06-07 December, 2012, Danang, Vietnam

i

Table of Contents

The 2012 International Conference on BioSciences and BioElectronics Table of contents _____________________________________________________________ i Message from the Chairs ______________________________________________________ iii Technical Program Committee _________________________________________________ v

Keynote Abstracts Automated CCTV Surveillance Research at the University of Reading ________________ vi Bioprospecting microbes to innovate new biotechnological uses for control of plant pests and diseases _________________________________________________________________ vii SESSION A: BIOELECTRONICS

A Survey on Advanced Video-Based Healthcare Monitoring Systems ____________________ 1 Le Thi My Hanh, Hoang Le Uyen Thuc, Tuan V. Pham.

Pulse Oximetry System based on FPGA ___________________________________________ 9 Kien Nguyen Trung, Ngoc Nguyen Van.

Towards Efficient Implementation of Neural Networks with Reduced Precision Floating Point Parameters ___________________________________________________________________ 14 Huynh Viet Thang, Nguyen Hai Trieu Anh.

Human-Computer Interaction using ultrasound hand tracking via USB interface ____________ 19 Sau V. Tran, Tuan K. Tran, Tu T. Hoang, Trung X. Pham.

A Real-Time Scheduling Scheme for Distributed Control System ________________________ 24 Trong Cac Nguyen, Xuan Hung Nguyen, Van Khang Nguyen.

Fall Detection Based On Hidden Markov Model _____________________________________ 32 Viet Q. Truong, Hieu V. Nguyen, Tuan V. Pham.

Measurement of Biological Concentration Using Magnetic Agents_______________________ 37 Cao Xuan Huu, Tuan V. Pham, Dang Duc Long, Nguyen Thi Minh Xuan, Pham Thi Kim Thao.

A Novel Approach to Protect Intellectual Property Core of FPGA-Based Partially Reconfigurable Systems ________________________________________________________ 42 Tran Thanh, Tran Hoang Vu, Pham Ngoc Nam, Nguyen Van Cuong.

ii

A DAG - SVM based Method for Face Recognition using PCA. ________________________ 46 Tran Thi Minh Hanh.

A Speaker Recognition System using Combination Method between Vector Quantization and Gaussian Mixture Model _______________________________________________________ 51 Ngo Quoc Hung, Tuan V. Pham.

Design of Television Remote Controller based on Vietnamese Speech Recognition _________ 55 Nguyen Tu Ha, Tuan V. Pham.

ECG Signal Transmission Using Wireless Technology in Patient Health-care and Monitoring System ______________________________________________________________________ 58 Duong Trong Luong, Nguyen Duc Thuan, Nguyen Hung.

Combination of analog and digital solutions for Wireless ECG Monitor __________________ 63 Khue Tra, Phung T.Kim Lai, Khiem Duy Nguyen, Tuan V. Pham.

SESSION B: BIOMETRICS

Biodegradation of Phenol by Native Bacteria Isolated from Dioxin Contaminated Soils ______ 68 Bui Ba Han, Nguyen Thi Lan, Dang Duc Long.

Biopolymer Film from Chitosan for Shelf-life Extension of Fruits _______________________ 74 NGUYEN, Thi Xuan Lam; NGUYEN, Thi Minh Xuan; DANG, Duc Long

Effect of CO2 Utilization on the Growth of Chlorella Vulgaris for Food Technology _________ 79 Nguyen Hoang Minh, Nguyen Thi Thanh Xuan, Dang Kim Hoang, Nguyen Dinh Phu

Effect of Carbon Sources on Proliferation of Zedoary (Curcuma Zedoaria Roscoe) Cell Suspension Cultures ___________________________________________________________ 83 Vo Chau Tuan, Tran Quang Dan

Use of BMWPVIET and ASPT Indices as Bioindicators for Testing Water Quality of Rivers in Danang city __________________________________________________________________ 88 Nguyen Van Khanh, Vo Van Minh, Kieu Thi Kinh, Tran Duy Vinh, Phan Thi Hien

Whole Cell Immobilisation of Bacillus Subtilis on Cellulose Carriers and Waste Water Treatment Application __________________________________________________________ 93 TRAN, Thi Xo, NGUYEN, Thi Diem Quynh.

Lactic Acid Fermentation from Jackfruit Seed Flour __________________________________ 96 Trương Thi Minh Hanh, Ho Thi Hao.

Gold Nanoparticles based Localized Surface Plasmon Resonance in Combination with MicroFluidic System for Biomolecule Determination. _________________________________ 100 Nguyen Ba Trung, Le Tu Hai, Yuzuru Takamura.

iii

Preface Dear readers and colleagues,

It is a great honour to introduce the proceedings of the BioSciences and BioElectronics Conference (ICBSBE2012). At the conference, many scholars and experts from industrial organizations and government agencies have participated in various sessions to share research ideas, potential collaboration opportunities and state-of-the-art achievements in the field of emerging BioSciences and BioElectronics and BioMetrics technologies throughout the world. In the coming decades, these technologies will certainly play a decisive role in the sustainable development of Vietnam and many other countries.

We acknowledge that BioSciences and BioElectronics are critical to sustainable development and poverty reduction efforts. It affects all aspects of development such as social, economic, and environment including livelihoods, access to water, agricultural productivity, health, population levels, education, and gender-related issues. None of the Millennium Development Goals can be met without major improvement in the quality and quantity of such BioSciences and BioElectronics services in developing countries. Therefore, this Conference is an excellent opportunity for researchers, industry professionals, academic and administrative organizations to exchange the latest innovations and developments in technologies between United Kingdom and Vietnam. Accordingly, conference outcomes will contribute to the environmental improvement as what we desire so that Vietnam can develop sustainably.

The 2012 ICBSBE is co-organized by Danang University of Technology – The University of Danang, British Council Vietnam, and the University of Reading. The goal of the conference is two-fold: First, to exchange the latest innovations and developments in the fields of BioSciences, BioElectronics, BioMetrics and Applications in related technologies; Secondly, to gather academics, researchers, industry professionals, and higher educational-related organizations from Vietnam and foreign countries, especially from United Kingdom. ICBSBE2012 will include keynote speeches, paper presentations, tutorials, panel discussions and field trips.

In recent years, we have set up some Teaching Research teams (TRTs) to conduct a series of research projects in the field of technology and engineering with our international partners. We have also focused on the collaborative programs such as student and staff exchange programs, advanced programs, co-organisation of international symposiums, especially joint research and co-publication with university partners.

The staff and students at Danang University of Technology - The University of Danang are very pleased to work as the host university for the two thousand and twelve

iv

(2012) BioSciences and BioElectronics Conference ICBSBE2012. This is a meaningful event for the University of Danang and Danang City, a potential and dynamic city in socio-economic development as well as in environmental protection. I am proud to say that Danang will become a smarter city to improve living and working environtments. It is evident that our staff and students and people in Danang will be of the beneficiaries from the outcomes of this important conference.

Thank you again for choosing Danang University of Technology, The University of

Danang as the location for the conference. Thank you for coming and attending the conference and I hope you enjoy the time in

Danang. I wish all of you good health and the conference the best success. Conference Chair Prof. Dr. Le Kim Hung

v

Technical Program Committee Honorary chairs:

Tran Van Nam (President, The University of Danang) Le Kim Hung (Rector, Danang University of Technology) Steven Mithen (Pro-Vice-Chancellor, University of Reading)

General chairs: • Vo Trung Hung (Director of Science Technology, The University of Danang) • Hoang Hai (Vice-Director of International Collaboration, The University of Danang) • Robin Rickard (Director, British Council Vietnam) • Vu Kieu Dung (Assistant Director, British Council Vietnam) • Truong Hoai Chinh (Vice-Rector, Danang University of Technology) • Le Thi Kim Oanh (Vice-Rector, Danang University of Technology) • Nguyen Dinh Lam (R&D, Post-graduate & Int. Collaboration, Danang University of

Technology)

Technical program • David Hogg (University of Leeds) • Jenq-Neng Hwang (University of Washington) • James Ferryman (University of Reading) • Rob Jackson (University of Reading) • Tang Tan Chien (The University of Danang) • Nguyen Van Tuan (Danang University of Technology) • Vo Van Minh (Danang University of Education) • Dang Duc Long (Danang University of Technology) • Huynh Huu Hung (Danang University of Technology) • Pham Van Tuan (Danang University of Technology)

Publicity • Duong Mong Ha (University of Danang) • Nguyen Thu Giang (British Council Vietnam) • Bui Thanh Nga (British Council Vietnam)

Finance • Lam Thi Hong Tam (Danang University of Technology) • Vu Ngoc Ha (Danang University of Technology)

Publication • Ngo Thai Bich Van (Danang University of Technology) • Huynh Viet Thang (Danang University of Technology) • Huynh Tan Tien (Danang University of Technology)

Local arrange • Le Minh Duc (Danang University of Technology) • Nguyen Le Hung (Danang University of Technology) • Tran Thi Huong (Danang University of Technology) • Nguyen Thi Minh Xuan (Danang University of Technology)

Contact: • Pham Van Tuan, Email: pvtuan@dut.udn.vn

vi

Keynote Talk 1 Automated CCTV Surveillance Research at the University of Reading

Abstract:

The Computational Vision Group (CVG), within the School of Systems Engineering at the University of Reading, are very active in security research, with a focus on automated surveillance including CCTV analysis for safety, security, and threat assessment. This talk will present an overview of security research activities with particular attention paid to recent and ongoing FP7 security projects SUBITO (detection of unattended baggage and the identification and tracking of the owner), Co-Friend (cognitive understanding of airport aprons), and EFFISEC (more efficient integrated security checkpoints), in which CVG play a significant role. The talk will also address the overarching themes of performance evaluation of detection systems, in particular details of the PETS (Performance Evaluation of Tracking and Surveillance) workshop series which has most recently addressed multi camera crowd image analysis.

Speaker’s Biography:

Dr. Ferryman leads both the Computational Vision Group (CVG) and the wider Computing Research Group in the School of Systems Engineering, University of Reading. Dr. Ferryman's personal research interests include automated CCTV surveillance of public spaces (including improving the effectiveness of CCTV), behavioural analysis, cognitive systems, multimodal interaction, robotics and autonomous systems, novel imaging modalities, and performance evaluation. Dr.Ferryman has received extensive funding from Research Councils UK, the EU, and Industry. Dr. Ferryman was Co-investigator on the EPSRC network ViTAB (Video based Threat Assessment and Biometrics) and was Principal Investigator on the EPSRC REASON project (EP/C533402) examining robust methods for monitoring and understanding people in public spaces. Dr. Ferryman has been Principal Investigator on a number of EU projects including the PASR project ISCAPS (013800) on integrated surveillance of crowded areas for public security, the EU Framework 6 Aeronautics project SAFEE, which addressed aircraft onboard threat detection, and Framework 7 Security/Cognitive Systems projects Co-Friend (airport apron monitoring) and EFFISEC (efficient integrated security checkpoints.) Dr. Ferryman is a member of the British Machine Vision Association (BMVA) and the Security Information Technology Consortium (SITC).

vii

Keynote Talk 2

Bioprospecting microbes to innovate new biotechnological uses for control of plant pests and diseases

Abstract:

With an increasing number of restrictions on the use of chemicals to control pathogens and pests of plants, there is an urgent need to identify alternative means of control. One approach that is receiving a resurgence of interest is the use of biological agents, which may include whole organisms, viruses, or components of them. In our laboratory, we employ bioprospecting to identify novel biological agents that may be utilised for pest and disease management. Our main targets are bacteria and bacteriophages, with a view to controlling fungal and oomycete pathogens, insect pests such as aphids and thrips, and bacterial pathogens. We also employ genetic screens to identify the systems employed by bacteria to survive in different niches and under different climactic conditions. As well as providing key insights to bacterial function, it can also help to reveal novel bioactive molecules such as toxins and surfactants. Taken together, these approaches can innovate the discovery of new biological controls for plant pests and diseases.

Speaker’s Biography:

Robert W. Jackson is a Lecturer in Microbiology within the School of Biological Sciences, University of Reading. He is also Admissions tutor for the BSc (Hons) Microbiology degree and Director of Enterprise for the School.

Having always had an interest in science from a young age, his appetite for research started during a 6 month Erasmus placement at the University of the Algarve in Portugal working with Prof. Jose Leitao – after a one month period of dossing on the beach and drinking too much wine from the local garrafao, he finally got down to work examining tissue culture methods to propagate local plant species. A further 6 month spell working with Dr David Royle on Septoria disease on winter wheat at Long Ashton Research Station was the project that sparked an interest in plant pathology and led to searches for a PhD in the discipline. His PhD (1994-1997) research was done in the laboratory of Alan Vivian at the University of the West of England examining the role of plasmids in Pseudomonas syringae pathogenicity on plants. A major breakthrough led to 3 years BBSRC postdoc work with Alan and John Mansfield (Imperial College). In 2001, he moved to the lab of Paul Rainey in the Department of Plant Sciences, University of Oxford, where a whole new world of [non-molecular] ecology and evolution did its best to befuddle him. After this highly inspiring and insightful spell, Rob was awarded a British Society for Plant Pathology fellowship to do a 2 month research project at the University of Auckland in 2004. The following 5 months were spent back at UWE working for Dawn Arnold and in late 2004, he joined the lab of Richard Cooper in the University of Bath studying bacterial extracellular polysaccharides. In 2006, he finally achieved his ambition

viii

by landing a lectureship at the University of Reading.

In his personal time, he enjoys cooking, wine, reading, gardening, scuba diving, hiking and dreaming that Liverpool might once again lift the league cup!

1

A Survey on Advanced Video-Based Healthcare Monitoring Systems

Le Thi My Hanh Department of Information Technology

Danang University of Technology Danang, VIETNAM ltmhanh@dut.udn.vn

Hoang Le Uyen Thuc, Pham Van Tuan Department of Electronic and Communication Engineering

Danang University of Technology Danang, VIETNAM

hluthuc, pvtuan@dut.udn.vn

Abstract—Video-based healthcare monitoring technique has become an important field in computer vision researches and applications. Based on the intelligent video analysis, numerous algorithms are proposed and implemented for healthcare applications such as monitoring the daily activities of elderly, detecting a fall, measuring the gait features to detect the early symptom of some illness, and evaluating the recovery progress of patients. This paper surveys recent video-based healthcare monitoring systems, which are quite infant, challenging but promising. Three main subjects are covered in this survey: first, we describe a typical video-based healthcare monitoring system; second, we select recent interested papers to summarize and discuss in terms of their proposed algorithms, applications, achievements and limitations; finally, we indicate the key technical problems and future research directions of the field.

Keywords- video-based, healthcare, fall detection, gait analysis, activity recognition, rehabilitation.

I. INTRODUCTION The World Health Organization (WHO) estimates that

there is a need of 4.3 million additional trained health workers worldwide to address basic health requirements [1]. Especially, the scarcity of trained health workers has reached serious crisis level in 57 countries. From among 57 above countries, 8 countries including Angola, Benin, Cameroon, Ethiopia, Zambia, Haiti, Sudan and Vietnam with acute shortages of human resources for health (HRH) are selected as pathfinder countries [2]. The shortage of health workforce causes serious damage to the human public health worldwide. Simultaneous combination of HRH development and relieving their workload is really a globally urgent issue, in particular to 8 above countries.

This motivates a promising research branch nowadays – designing, developing and applying the automatic health care monitoring systems. Among a series of such researches all over the world, video-based techniques are emerged as the widespread proliferation thanks to their remarkable advantages such as the easy installation, the unobtrusive operation and the convenient maintenance. This problem is great challenging and sophisticated due to very large variations of the context-dependent motion appearance in different viewing angles, different illumination and background, different clothes, different action speed, camera movement, and occlusion in

terms of human-human occlusion, human-body part occlusion, human-object occlusion, etc.

According to our careful observation, there is no survey on video-based healthcare monitoring systems. Our survey, which aims to provide the comprehensive view of the field, concentrates on the applications of intelligent video analysis to automatic healthcare monitoring systems, including monitoring the daily activities of elderly, detecting a fall, measuring the gait parameters, and evaluating the recovery progress of patients. Most of our selected papers are published after 2000 in order to ensure the up-to-date knowledge in the survey.

Video-based healthcare monitoring systems concerns to automatic recognition of human actions/activities from video sequences. More specifically, the systems are to “analyze” and “understand” patients’ actions and/or activities, so it can facilitate health workers to diagnose, treat and care patients, resulting in improving the reliability of diagnosis, decreasing the working load for the medical personnel, shortening the hospital stay for patients, and improving the quality of life for patients, as well.

Common functional modules of a typical video-based healthcare monitoring system include three main processing steps as shown in Fig. 1.

Figure 1. Overview of the general video-based healthcare monitoring system

In the first step, feature representation, the moving humans in video input are first extracted using human object detection and segmentation algorithms. The humans are then

2

continuously tracked using object tracking algorithms. The tracking results in every frame will be transformed into a reduced representation set of features (also called feature vector) simultaneously. In order to achieve the appropriate performance, the extracted features have to deal with spatial-temporal scaling variations associated with human actions, as well as encapsulate the unique characteristics of an action performed by different persons. A good feature descriptor is expected to be able to extract the most suitable information from the input video in order to perform the predefined tasks with sufficient accuracy without using the full size input. Thus, the first step is equivalent to condensing each input video frame into a specified multi-dimension feature vector.

After feature vectors are appropriately obtained, we can now implement the step 2 which is human action/activity recognition. Activities can be simple actions such as walking, waving; complex actions such as ballet dancing, doing exercise; the interactions between persons-persons such as hand shaking, hugging; or the interactions between humans - objects such as preparing a dinner, punching a punch-bag. Recognition of human activities is a challenging and sophisticated task due to great diversity in the way an activity is performed by different people or even by the same people with varying viewing perspectives or time duration.

The goal of the last step – applications – is to analyze the classified activities so that their semantic meaning can be understood.

In this paper, we do a review for all the 3 stages of the system. The rest of the paper is organized as follows. In Section 2, we describe feature representation methods. Section 3 mentions the recognition algorithms. The applications are introduced in Section 4. The discussion and summary are described in Section 5, followed by the conclusion in Section 6.

II. FEATURE REPRESENTATION Here we define the feature extraction as consisting of three

main processes: (1) data capturing, (2) object segmentation and (3) feature description.

A. Data Capturing

In the video-based healthcare monitoring system, the input data can be gathered from video cameras as a sequence of image frames. Video signal can be captured using one camera [3, 4] or multiple cameras. Normally, 2D analysis requires only one camera, but the performance can be affected by many factors such as the occlusion in terms of body - body part, human – object, human – human, the viewing angle change as well as the depth ambiguity. 3D analysis can overcome these limitations; however it usually requires more than one camera in order to measure the depth information of the subject to reconstruct the 3D positions of interested human points [5, 6]. In order to extract the depth information, inexpensive depth camera such as Microsoft Kinect is the new interest of many researchers [7, 8]. In some cases, video combined with data signals provided from optical markers attached to specific body points or other kinds of wearable sensors are used but this is out of scope of our paper.

B. Object Segmentation

This step in most of video-based systems is separating the interested objects from the rest of the image frame called the background. Based on the mobility of cameras, the object segmentation can be categorized as two types of segmentation: the static camera segmentation and moving camera segmentation. For the healthcare applications, the most popular method is static camera segmentation where the cameras are located in the specific positions with fixed angles.

The extraction of moving humans from the input video stream can be achieved by human object detection and segmentation algorithms based on temporal difference of two successive frames [9, 10], or background subtracting from each image [11, 12].

Background subtraction is known as a powerful object segmentation method and quite suitable for indoor static camera environment in recent years thanks to the development of sophisticated dynamic background estimation and updating techniques. The basic scheme of background subtraction is to subtract the image from a reference image that is the estimated background. The method by Stauffer and Grimson [11] has today become the most popular background subtraction method. Each pixel is represented by a mixture of Gaussians and then updated with new Gaussians during run-time. This update was done recursively, so as to model the slow changes in the scene such as illumination change and noises.

In addition, some recent researches implement the shadow detection and removal, random noise filtering as well as morphology operations to smooth the boundary and fill the small holes to create well-defined silhouette images, such as [9, 10, 13]. An example of original frame and corresponding segmented human silhouette is shown in Fig. 2.

Figure 2. An example of extracted human silhouette [13]

C. Feature Description

The step is to associate the detected humans in the current frame with those in the previous frames, providing the temporal trajectories of characteristics of the segmented objects such as shape, color and motions in the form of features. As mentioned before, the carefully chosen feature descriptor plays an important role in the entire video-based system.

There are many feature description methods proposed in literature, which we classify into two main types called numeric [14-18] and Boolean (binary) features [19, 20]. The numeric features are most popular and they are represented as continuous-valued numbers. Boolean features take either

3

number 0 or 1 to express the binary geometric relations between certain points of a human pose. Numeric features include two main categories which are shape-based methods and flow-based methods. Shape-based approaches attempt to extract silhouettes of moving people such as [14-16]. Flow-based methods estimate the optical flow field between adjacent frames such as [17-18]. One inherent disadvantage of shape-based features is that they can not capture the internal motion of the object within the silhouette region. Moreover, the state-of-the-art background subtraction techniques still cannot segment precise silhouettes, especially in dynamic environments. In contrast to shape-based methods, it is not necessary to implement background subtraction in flow-based techniques. However, the features are still dependent on the camera view and therefore are not suitable for action recognition from monocular videos.

Some of the effective feature descriptors can be listed as follow:

Bobick and Davis represent an action by a motion history image (MHI) [14]. In an MHI, pixel intensity is a function of the recency of motion in a sequence at that point, i.e., more recently moving pixel is brighter than the past one.

Diaf and Benlamri represent an action by a single image which is called motion intensity image (MII) [15]. They get the MII by aligning the human silhouette of each background-subtracted binary image to a reference point and then form a single intensity image of these silhouettes by taking into account the difference between each subsequent silhouette. By this way, the temporal occlusion and the imperfect extraction are effectively removed.

In [16], Kim et al. adopt the accumulated motion energy image (AMI) using image differences. For the robustness of the features, the AMI is then resized to a NxN sub-image by intensity averaging and a rank matrix is generated by ordering the sample values in the sub-image. Figure 3 shows an example of two quite different sub-images due to different clothes and backpack and two quite different sub-images but their corresponding rank matrices are identical.

The method performed by Ahmad and Lee is for human action recognition from any arbitrary view image sequence that uses combination of the Cartesian component of optical flow velocity and human body silhouette feature vector information [18]. The action region in an image frame is represented by Q-dimensional optical flow feature vector combined with R-dimensional silhouette feature vector.

Muller et al. introduce binary features which is a set of relations to test whether the position of a specified body point is in-front-of/behind or right/left a specified plane, the existence of a bent/stretched pose of a body part, the touched/untouched scenario of two specified body points [19]. For example, to characterize whether the leg is bent or stretched, the angle between thigh and lower leg is used. The corresponding “leg angle”

feature is 1 (bent pose) if that angle is less than 1200 and 0 (stretch pose) if that angle is over than 1200.

Figure 3. Example of two different images with identical features [16]

In general, the numeric features have been shown to achieve relatively good results in human action recognition. However, they are based on 2D information extracted from image sequences; therefore, are sensitive to occlusions and viewing dependent.

Binary features, which are mainly derived from 3D point coordinates, can better handle the occlusion problems. Moreover, the binary features are efficient to describe human poses based on low dimensional feature vectors. Unfortunately, due to the use of only binary numbers, 0s or 1s, to describe the geometric relation in a pose, binary features cannot be so discriminative in describing the sophisticated human body motion. For example, the Boolean feature describing the hand’s pose in a typical jogging action is not efficient because both hands are repeating a series of forward-backward movement, while always keeping in front of body plane, i.e., the Boolean feature values are consistently equal to 1 in every frame without revealing the actual motion of both hands.

III. ACTIVITY RECOGNITION

After obtaining a feature descriptor, the recognition problem is performed as a higher level task. It may be simply considered as a classification problem of temporal feature data, i.e., statistically identifies the sequence of extracted features into one of the categories of interested activities. We classify the recognition algorithms into two main categories of methods, namely, static and temporal classifications. Temporal classification can be further classified into template matching and state-space schemes.

A. Static classification

Static classification methods are not interested in the temporal information of the image frames.

K-nearest neighbor (K-NN) is a popular closest pattern-based method due to its simplicity and hence, is able to classify real-time at a very low cost. The training phase of the

4

algorithm is only labeling and then storing the training samples. In the testing phase, an unknown pattern is assigned to the group which has most samples among the K training samples closest to that testing point. Kumari and Mitra propose an action recognition method based on the average discrete Fourier transform (DFT) feature extracted from the small image blocks then apply K-NN algorithm for recognition [21].

Besides KNN, support vector machine (SVM) provides a powerful approach to the recognition problems. The main idea of the standard SVM algorithm is as follows: suppose there are two 2D data sets separated by some hyperplanes. In the training phase, it is necessary to choose the optimal hyperplane so that the margin is maximum in order to minimize the error rate. Once we have trained an SVM, in the testing phase, we simply determine on which side of the decision hyperplane a testing point lies and then assign the corresponding class label. The standard SVM algorithm is formulated to solve only linear two-class categorization problems. Therefore, extending standard SVM algorithm to multi-class and nonlinear has great significance. Schuldt et al. [22] use local space-time features in a video and apply SVM to recognize human actions. Laptev et al. [23] apply non-linear SVM with a multi-channel Gaussian kernel for recognition of various natural human actions, including AnswerPhone, GetOutCar, HandShake, HugPerson, Kiss, SitDown, SitUp and StandUp by building spatial-temporal bag-of-features.

B. Temporal classification

Unlike static classification method, temporal classification methods pay attention to the temporal information of the data. Temporal classification can be categorized as template matching and state-space schemes.

1) Template matching

Template matching methods represent a human action as a set of template feature sequences of executed action. When a video is coming, the sequence of feature vectors extracted from that unknown video will be compared with the template patterns to determine the similarity. The highest similarity (or the smallest distance) is chosen as the criterion for classification of the actions.

Humans may perform an action in different styles and/or different rates, and the similarity must be measured considering such variations. The dynamic time warping (DTW) has been widely used for matching two human movement patterns. DTW deals with differences between sequences by operations such as deletion-insertion, compression-expansion, and substitution of subsequences [24]. By defining a metric of how much the sequences differ before and after these operations, DTW classifies the sequences. Even if the time scale between a test pattern and a reference pattern may be inconsistent, DTW can still successfully establish the matching as long as time ordering constraints hold. Figure 4 shows an example of matching between two sequences with different execution rates [25]. Sempena et al. [26] use depth camera Microsoft Kinect to recover 3D human joints, and then build feature vector using joint orientation so that the orientation is invariant to human body size. For action recognition, the sequence of feature

vectors extracted from video input is compared to the list of defined feature vectors using DTW method. Upper part related actions such as clap, punch, smash and wave can be recognized quite well.

Figure 4. Example of two sequences with different execute rates. images with identical features [25]

The most significant advantage of template matching is simple implementation. The computational cost is proportional to the number of sequences in database. However, the template matching is sensitive to noise and spatial-temporal scale variances.

2) State-space methods

State-space methods represent a human action as a model composed of a set of states. The model is statistically trained so that it corresponds to sequence of feature vectors belonging to its action class. Generally, one model is constructed for each action. For each model, the probability of the model generating an observed sequence of feature vectors is calculated to measure the likelihood between the action model and the input image sequence. The maximum likelihood is selected as the criterion for recognition of the actions. In an action model, a state can be repeated generated itself with the arbitrary number of times, so it is superior to template matching in temporal data recognition problems; and therefore, are able to be widely applied in the human action recognition.

One very popular model used in the state-space methods is hidden Markov model (HMM) [27]. HMM structure includes a hidden Markov chain and a finite set of output probability distribution. More specifically, an HMM is completely characterized by a set of three matrices λ = A, B, π, where A = transition matrix = aij, with aij being the transition probability from state qi to qj, (i, j)ϵ [1: N); B = observation matrix = bj(k), with bj(k) being the probability of observed output (discrete) symbol vk at state qj, k ϵ [1: M); π = πi, with πi being the initial state probability.

In order to recognize the actions based on HMM, the use of HMM covers two stages: training and classification. In the training stage, each HMM is statistically trained for each action, using the feature vectors extracted from training data belonging to its action class. The number of states of an HMM must be specified, and the corresponding state transition and observed symbol probabilities are optimized in order that the generated symbols can correspond to the observed vector features. In the classification stage, for each model, the probability of the model generating an observed sequence of feature vectors is calculated to measure the likelihood between the action model and the input image sequence. If the calculated probability of a particular model is highest, it is able to decide that the action corresponding to that model occurs in the given input. In Figure 5, a human is assumed to be in one

5

state at each time, and each state generates an observation which is a feature vector. Each human image in the figure represents a pose with the highest observation probability bj(k) for each state qj.

Figure 5. Example of a HMM to model the waving action

Yamato et al. are the first authors applying HMM into human action recognition [28]. Every set of time-sequential images is transformed into a mesh feature vector sequence, and converted into a symbol sequence by vector quantization. They then apply HMM to recognize six tennis strokes and achieve the acceptable recognition rates. Ahmad and Lee present each action by a set of hidden Markov models [29]; each model is for each action at any viewing direction. The training data set contains eight specified views. Experimental results of different actions from any viewing direction are correctly classified, which indicates the robustness of this view-independent method.

IV. APPLICATIONS The goal of the last step – applications – is to analyze the

classified activities so that their semantic meaning can be understood. Activity understanding requires expert knowledge to characterize the uniqueness accurately and to build the scenario suitable to each specified application. This makes the activity recognition techniques to be more valuable and widely used in our daily lives in forms of numerous numbers of diversified applications including healthcare environment. In this section, we are interested in four healthcare applications: daily life activity monitoring, gait analysis, fall detection, and rehabilitation applications.

A. Daily Life Activity Monitoring

Daily life activity monitoring mainly focuses on learning and recognizing the daily life activities of seniors at home. The proposed systems are to provide seniors an opportunity to live safely, independently and comfortably. In order to accomplish this, most of proposed systems are to continuously capture the movements of individual senior/ multiple seniors at home, automatically recognize their activities and detect the gradual changes in baseline activities such as mobility functional disabilities, mental problems, as well as the urgent warning signs of abnormal activities such as falling down or stroke. Some of these applications can be listed as follows:

Respiration behavior can be critical in diagnosing patient illness or recognizing distress during sleep. Many diseases—such as obstructive sleep-apnea

syndrome, cardiovascular disease, and stroke—induce abnormal respiration. Automated respiration monitoring is performed by Lee et al. [30]. They use near-IR images, captured by a near-IR camera to measure the sleeper's respiration based on the periodic rising and falling motion of their chest or abdomen.

J. Gao et al. measure feeding difficulties in nursing home residents with severe dementia, by automatically measure number of hand movements to the mouth using motion feature vectors and HMM to identify the start and ending of individual dining events [31].

Huynh et al. present a method for detecting and tracking of face, mouth, hands and medication bottles in the context of medication intake monitored by a camera [32]. This aims to monitor medicine intake behavior of elderly at home to avoid the inappropriate use.

In order to recognize the activities as the higher semantic level, the activity duration, the position of human, the interaction between people and person-object are the essential elements received lots of interests.

For the activity duration, Luhr et al. use the explicit state duration HMM (ESD-HMM), in which a duration variable is introduced into the standard HMM [33]. In [34], similar to [33], Duong et al. also add the duration information to the standard HMM to result in hidden semi-Markov model (HSMM); however, they model the state duration by using the generic exponential family. In [35], Duong et al. introduce the switching hidden semi-Markov model (S-HSMM) which exploits the benefit of both the inherent hierarchical organization of the activities and their typical duration.

For the human position, in [33, 34], the door, the stove, the fridge, the sink, the cupboard, and the table areas are used to define the activities in kitchen room. For example, the meal preparation and consumption include twelve steps: take-food-from-fridge bring-food-to-stove wash-vegetable come-back-to-stove-for-cooking take-plates/cup-from-cupboard return-to-stove-for-food bring-food-to-table take-drink-from-fridge have-meal-at-table clean-stove wash-dishes-at-sink leave-the-kitchen. In [36], the kitchen is quantized into 28 square cells of 1 m2 each and the position of the human is captured by four cameras mounted at the ceiling corners, and the tracking system returns the list of cells visited by person.

For the interaction, Liu et al. propose an interaction-embedded hidden Markov model (IE-HMM) framework for detecting and classifying individual human behaviors and group interactions in a nursing home environment [37]. The framework comprises a switch control (SC) module, an individual duration HMM (IDHMM) module, and an interaction-coupled duration HMM (ICDHMM) module. The SC module is to extract the atomic behavior units as an individual behavior unit (comprising a single participant) or an

6

interaction behavior unit (comprising two or more participants) based on the distances between the various participants in each scene, and monitoring the duration for which these distances are maintained. The individual behavior units are passed to the IDHMM module to be classified as the corresponding human behavior. Similarly, the interaction behavior units are passed to the ICDHMM module, where the corresponding interaction mode is classified. The IE-HMM is applied to analyze the human actions and interactions in a nursing home environment and get the overall recognition performance of 100% when applied for individual human actions and 95% when applied for group interactions.

B. Gait Analysis

Human gait analysis can be simply described as the study of human locomotion by measuring body movements, body mechanics, and the activities of muscles. Most gait analysis researches focus on the walking gait although people can move by running, hopping, skipping, or even crawling. Intuitively, walking gait is just a simple action people do to move whole body from one place to the other. However, walking is actually a very complex act and it requires many human organs, such as brain, ears, eyes, and muscles, to coordinate closely to make it right.

Lots of researches conclude that loss of the ability to walk correctly can be a result from a significant health problem, due to the fact that pain, injury, paralysis or tissue damage can alter normal gait [38]. In addition, the mental problems also can cause the gait problems such as the gait slowing is a predictor of incident dementia [39]. Thus, the main goal of gait analysis for healthcare is to detect the gait abnormalities, early symptoms of some diseases, which may affect the gait.

Liao et al. presents a video-based system for analyzing four posture features of walking human including body line, neck line, center of gravity (COG) and gait width based on the extracted silhouettes from front view and side view [40]. Similarly, Leu et al also use two cameras and the feedback control structure at the segmentation level to extract the gait features such as torso angle, left and right thigh angles, and left and right shank angles [41]. With the aim of removing the inconvenience of many methods that require the captured images of the walking human from either front or side view, Li et al. propose a system with much less restriction on walking direction [42]. The system successfully extracts gait features such as COG and pace length from images obtained from two cameras with orthogonal view.

C. Fall Detection

Fall in the elderly is the major health risk as it caused serious injuries and it is known that fall is the leading cause of injury deaths among seniors. Foroughi et al. conduct some methods to real-time detect the fall [43-46]. For example, fall is detected based on human shape variation. Extracted features which are combination of best-fit approximated ellipse around the human body, projection histograms of the segmented

silhouette and temporal changes of head pose, are finally fed to a multi-class SVM [43] or MLP Neural Network [44] for precise classification of motions and determination of a fall event. The other features are based on the combination of integrated time motion images (ITMI) and Eigen space technique [45, 46].

D. Rehabilitation Applications

Traditional rehabilitation systems often require patients lots of clinical visits for the physical therapy exercises and the scheduled evaluation until full recovering of mobility function for daily activities. Such clinical visits can be decreased by innovative rehabilitation systems, which are home-centered and self-health care with the help of video-based activity recognition techniques. Moreover, by continuously monitoring the daily activities and gaits, the early symptoms of some diseases can be timely detected so that the diagnosis and the intervention are more useful. Some of these applications can be listed as follows:

Stroke is a major cause of disability and health care expenditure around the world. Ghali et al. design a system to provide real time feedback to stroke patients performing the everyday kitchen activities necessary for independent living e.g. making a cup of coffee [47]. They envisage a situation in which a stroke patient stands at a standard kitchen and makes a cup of coffee in the usual way. The position and movement of the patient’s hands and the objects he/she manipulates are captured by overhead cameras and monitored using histogram-based recognition methods. The key events (e.g. picking up a cup) are recognized and interpreted in the context of a model of the coffee-making task.

In order to objectively evaluate the improvement of motor functions of the elders at home, as well as reduces the burdens on fitness instructors, Ryuichi et al. propose the “multimedia fitness exercise progress notes” system [48]. They use a popular video camera to capture the movements of the elders doing exercises, then send videos to an analysis center: on a screen, video files can replay to extract still images, so that many kinds of measurements, such as lap time, distances and angles can be performed using a mouse.

Goffredo et al. propose the Gauss-Laguerre transform-based (GLT-based) motion estimation method in order to analyze the sit-to-stand (STS) motion from monocular videos [49]. STS movement mainly involves hip and knee flexion-extension, and ankle plantarflexion-dorsiflexion is analyzed by utilizing a 2D human body model that considers the projections of body segments on the sagittal plane.

V. KEY PROBLEMS AND POTENTIAL RESEARCH DIRECTIONS

Although the recent video-based healthcare monitoring approaches have achieved encouraging performance, there are apparent problems that make it extremely difficult for real-

7

world applications. The performance of the system is adversely influenced by several factors as follow:

Viewpoint issue may be the main challenge for human activity recognition. In the real healthcare monitoring system, the activity sequences are usually observed from arbitrary camera viewpoints; and therefore, the applications require the view-independent methods. This means that the performance of system needs to be invariant from different camera viewpoints. However, most recent algorithms are based on the constrained viewpoints, such as person has to be in front-view (i.e., face a camera) or side-view. Some effective ways to solve this problem have been proposed, such as using multiple cameras to capture different view sequences then combining them as training data.

Most of moving human detection algorithms are based on background subtraction, which requires a reliable background model. In practice, some humans often walk in a complex and dynamic cluttered background, and/or varied light condition (e.g., day, night).

The natural human appearance can be changed due to many factors such as the walking surface conditions (e.g., hard/soft, level/stairs, etc.), clothing (e.g., long dress, short skirt, coat, etc.), footgear (e.g., stockings, sandals, slippers), object carrying (e.g., backpack, briefcase).

Promising directions for research in future are outlined below as potential solutions to those challenging problems:

The view-invariant method from monocular videos, although getting some preliminary results, is still an open and challenging research direction.

The human detection is still a problematic for real-time video-based systems. A large number of key problems still need to be solved, such as how to decrease the computational requirement, how to deal with the practical cluttered background, how to distinguish the static background with the motionless human, how to track multiple people, how to handle the occlusion in terms of body-body part, human-human, human-objects, etc.

The change of human action appearance leads researchers to a promising research branch: how to describe the movement that is less sensitive to appearance but still capture the most useful and unique characteristics of each action.

In addition, the patient behavior understanding has become a dynamic and challenging research branch. It attracts more and more researchers all over the world. Behavior understanding requires the additional contextual information such as W5+ (who, where, what, when, why, how) [50]. The same action may have several different meanings depending on the context in which it is performed; therefore it can be recognized as different behaviors. “Place” context can provide the location information to be used to detect abnormal behaviors. For example, lying on the bed or a sofa is taking relax or sleeping, but at the irrelevant places such as floor in

bathroom or kitchen, it can be a falling or a stroke. “Time” context is also a popular contextual description for behavior understanding. For example, a person usually watching TV after 2 am can be regarded as insomnia. Another example is that a person will be detected as picking up stuffs if he squats and stands up soon. But if he squats for a while, he can have a motion difficulty due to osteoarthritis or senility. Moreover, the number of repetitions of an action is also a good hint. For example, eating too many times or too little a day can be an early symptom of depression. The interaction between people or between person and objects is also a good context to identify the action meaning. For example, if a person is punching a punch-bag, he might be doing exercise. But if he is punching himself, it will be a mental disorder.

VI. CONCLUSIONS There have been a series of research projects on video-

based healthcare monitoring system all over the world in recent years. The aim of the survey is to review the related works on video-based healthcare monitoring techniques, including all modules of a typical system. Four video-based scenarios which are monitoring of daily life behaviors, analysis of gait, detecting of fall and evaluating of therapy exercises of patients were discussed. Although video-based healthcare monitoring approaches have achieved preliminary achievements, at this stage, there are still lots of technical problems need to be solved for real-world applications.

REFERENCES [1] J. J. Divino, “New Financing Tool to Help Reserve Global Health

WorkforceShortage,”from http://www.who.int/workforcealliance/media/news/2009/rrtlaunchpr/en/, 2009.

[2] Global Health Workforce Alliance, WHO, “GHWA Collaboration with eight Pathfinder Countries,” Status Report, Geneva, Switzerland, Nov. 2008.

[3] Xin Zhanh and Guoliang, “Dual Gait Generative Models for Human Motion Estimation From a Single Camera,” IEEE Trans. On System, Man and Cybernetics – Part B: Cybernetics, vol. 40(4), 2010, pp. 1034-1049.

[4] Shian-Ru Ke, Jenq-Neng Hwang, Kung-Ming Lan, Shen-Zheng Wang, “View-invariant 3D Human Body Pose Reconstruction using a Monocular Video Camera,” 5th ACM/IEEE Int. Conf. on Distributed Smart Cameras, 2011, pp. 1-6.

[5] Yu-Ren Li, Shaou-Gang Miaou, Charles Kater Hung, and Julius Tube Sese, “A Gait Analysis System Using two Cameras with Orthogonal View,” Int. Conf. on Multimedia technology, 2011, pp. 2841-2844.

[6] Gunawan Ariyanto and Mark S. Nixon, “Model-based 3D Gait Biometrics,” Int. Joint Conf. on Biometrics, 2011, pp. 1-7.

[7] Erik E. Stone and Marjorie Skubic, “Passive In-home Measurement of Stride-to-Stride Gait Variability Comparing Vision and Kinect Sensing,” 33rd Annual Int. Conf. of the IEEE EMBS, 2011, pp. 6491-6494.

[8] Erik E. Stone and Marjorie Skubic, “Passive, In-home Gait Measurement Using an Inexpensive Depth Camera: Initial Results,” 6th Int. Conf. on Pervasive Computing Techologies for Healthcare and Workshops, 2012, pp. 183-186.

[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati, “Detecting moving objects, ghosts, and shadows in video streams,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 10, pp. 1337-1342, Oct. 2003.

[10] J. S. Hu, T. M. Su, and S. C. Jeng, “Robust background subtraction with shadow and highlight removal for indoor surveillance,” 2006 IEEE/RSJ Int. Conf. Intell. Robots and Syst., pp. 4545 – 4550, Oct. 2006.

8

[11] C. Stauffer and W. E. L. Grimson, “Adaptive Background Mixture Models for Real-time Tracking,” Proc. IEEE ICCV 1999, pp. 246-252, Jun. 1999.

[12] A. Elgammal, D. Harwood, and L. S. Davis, “Non-parametric Model for Background Subtraction,” Proc. ECCV 2000, pp. 751-767, Jun. 2000.

[13] Shian-Ru Ke, Jenq-Neng Hwang, Kung-Ming Lan, Shen-Zheng Wang, “View-invariant 3D Human Body Pose Reconstruction using a Monocular Video Camera,” 5th ACM/IEEE Int. Conf. on Distributed Smart Cameras, 2011, pp. 1-6.

[14] Aaron F. Bobick and James W. Davis, “The Recognition of Human Movement Using Temporal Templates,” IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 23(3), pp. 257-267, Mar. 2001.

[15] Abdunnaser Diaf and Rachid Benlamri, “An Effective View-based Motion Representation for Human Motion Recognition,” Proc. Of Int. Symposium on Modeling and Implementation of Complex Systems, May 2010.

[16] Wonjun Kim, Jaeho Lee, Minjin Kim, Daeyoung Oh, and Changick Kim, “Human action recognition using ordinal measure of accumulated motion,” EURASIP journal on Advances in Signal Processing, vol. 2010, Article ID 219109, 11 pages, Jun. 2010.

[17] S. Danafar and N. Gheissari, “Action recognition for surveillance application using optic flow and SVM,” Proc. ACCV, Tokyo, Japan, Nov. 2007.

[18] Mohiuddin Ahmad and Seong-Whan Lee, “HMM-based Human Action Recognition Using Multiview Image Sequences,” Proc. IEEE ICPR, 2006.

[19] Meinard Muller, Tido Roder, Michael Clausen, “Efficient content-based retrieval of motion capture data,” Association for Computing Machinery Transactions on Graphics, vol. 24, Issue 3, pp. 677-685, Jul. 2005.

[20] Tido Roder, “Similarity, Retrieval, and Classification of Motion Capture Data,” PhD Dissertation, Mathematic – Natural Sciences Department, University of Bonn, pp. 50-53, Sep. 2006.

[21] S. Kumari, S. and S.K. Mitra, “Human Action Recognition Using DFT,” Third National Conference on Pattern Recognition, Image Processing and Graphics, pp.239-242, Dec. 2011.

[22] C. Schuldt, I. Laptev, B. Caputo, “Recognizing human actions: a local SVM approach,” 17th Int. Conf. on Pattern Recognition, vol.3, pp. 32- 36, Aug. 2004.

[23] I. Laptev, M. Marszalek, C. Schmid, B. Rozenfeld, “Learning realistic human actions from movies,” IEEE Conf. on Computer Vision and Pattern Recognition, pp.1-8, Jun. 2008.

[24] J. K. Aggarwal and Q. Cai, “Human Motion Analysis: A Review,” Computer Vision and Image Understanding, 73(3), pp.428-440, 1999.

[25] J. K. Aggarwal and M. S. Ryoo, “Human Activity Analysis: A Review,” Journal ACM Computing Surveys, vol. 43 (3), Apr. 2011.

[26] Samsu Sempena, Nur Ulfa Maulidevi, Peb Ruswono Aryan, “Human Action Recognition Using Dynamic Time Warping,” IEEE ICEEI, pp. 1-5, Jul. 2011.

[27] Lawrence R. Rabiner, “A tutorial on hidden Markov models and selected applications in speech recognition,” IEEE, vol. 77(2), pp. 257-286, 1989.

[28] J. Yamato, J. Ohya, and K.Ishii, “Recognizing human action in time-sequential images using hidden Markov model,” Proc. IEEE CVPR, pp. 379-385, Jun. 1992.

[29] Mohiuddin Ahmad and Seong-Whan Lee, “HMM-based Human Action Recognition Using Multiview Image Sequences,” Proc. IEEE ICPR, 2006.

[30] Jiann-Shu Lee, Yung-Ming Kuo, and Pau-Choo Chung, “A Visual Context-Awareness- Based Sleeping-Respiration Measurement System,” IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 2, pp. 255-265, 2010.

[31] Jiang Gao, Alexander G. Hauptmann, Ashok Bharucha, and Howard D. Wactlar, “Dining Activity Analysis Using a Hidden Markov Model,” Proc. 17th Int. Conf. on Pattern Recognition, UK, Aug. 2004.

[32] Huynh H. H., et al., “Real-time Detection, Tracking and Recognition of Medication Intake,” World Academy of Science, Engineering and Technology, vol. 60, pp. 280-287, 2009.

[33] S. Luhr, S. Venkatesh, G. W. West, and H. H. Bui, “Explicit State Duration HMM for Abnormality Detection in Sequences of Human Activity,” Proc. 8th Pacific Rim Itl’l Conf. Artificial Intelligence, pp. 983-984, Aug. 2004.

[34] Thi V. Duong, Hung H. Bui, Dinh Q. Phung, and Svetha Venkatesh, “Human Behavior Recognition with Generic Exponential Family Duration Modeling in the Hidden Semi-Markov Model,” Proc. IEEE ICPR, pp. 1-6, 2006

[35] Thi V. Duong, Hung H. Bui, Dinh Q. Phung, and Svetha Venkatesh, “Activity Recognition and Abnormality Detection with the Switching Hidden Semi-Markov Model,” Proc. IEEE CVPR, vol. 1, pp. 838-845, Jun. 2005.

[36] Tsung-Yen Liao, Shaou-Gang Miaou, and Yu-Ren Li, “A Vision-based Walking Posture Analysis System without Markers,” 2nd Int. Conf. on Signal Processing Systems, pp. 254-258, 2010,

[37] Chin-De Liu, Pau-Choo Chung, Yi-Nung Chung, and Monique Thonnat, “Understanding of Human Behaviors from Videos in Nursing Care Monitoring Systems,” Journal of High Speed Network, vol. 16(1), pp. 91-103, 2007.

[38] http://www.patient.co.uk/doctor/Abnormal-Gait.htm [39] L. M. Waite, D. A. Grayson, O. Piguet, H. Creasey, H. P. Bennett, G. A.

Broe, “Gait Slowing as a Predictor of Incident Dementia: 6-year longitudinal data from the Sydney Older Persons Study,” Journal of the Neurological Sciences, pp. 89-93, 2005.

[40] Tsung-Yen Liao, Shaou-Gang Miaou, and Yu-Ren Li, “A Vision-based Walking Posture Analysis System without Markers,” 2nd Int. Conf. on Signal Processing Systems, 2010, pp. 254-258.

[41] A. Leu, D. Ristic-Durrant and A. Graser, “A Robust Markerless Vision-based Human Gait Analysis System,” 6th IEEE Int. Symposium on Applied Computational Intelligence and Informatics, 2011, pp. 415-420.

[42] Yu-Ren Li, Shaou-Gang Miaou, Charles Kater Hung, and Julius Tube Sese, “A Gait Analysis System Using two Cameras with Orthogonal View,” Int. Conf. on Multimedia technology, 2011, pp. 2841-2844.

[43] Homa Foroughi, Alizera Rezvanian, and Amirhossien Paziraee, “Robust Fall Detection using Human Shape and Multi-class Support Vector Machine,” IEEE 6th Indian Conf. on Computer Vision, Graphics and Image Processing, pp. 413-420, 2008.

[44] Homa Foroughi, Baharak Shakeri Aski, and Hamidreza Pourreza, “Intelligent Video Surveillance for Monitoring Fall Detection of Elderly in Home Environments,” 11th IEEE Int. Conf. on Computer and Information Technology, pp. 219-224, 2008.

[45] Homa Foroughi, Hadi Sadoghi Yazdi, Hamidreza Pourreza, and Malihe Jadivi, “An Engenspace-Based Approach for Human Fall Detection Using Integrated Time Motion Image and Multi-class Support Vector Machine,” 4th IEEE Int. Conf. on Intelligent Computer Communication and Processing, pp. 83-90, August 2008.

[46] Homa Foroughi, Aabed Naseri, Alizera Saberi, and Hadi Sadoghi Yazdi, “An Eigenspace-Based Approach for Human Fall Detection Using Integrated Time Motion Image and Neural Network,” IEEE Int. Conf. on Signal Processing, pp. 1499-1503, 2008.

[47] Ahmed Ghali, Andrew S. Cunningham, and Tony P. Pridmore, “Object and Event Recognition for Stroke Rehabilitation,” Visual Communication and Image processing Conf., 2003.

[48] Ayase Ryuichi, Hagashi Terumasa, Takayama Satory, Sagawa Setsuko, and Ashida Nobuyuki, “A Method for Supporting At-home Fitness Exercise Guidance and At-home Nursing Care for the Elders, Video-based Simple Measurement System,” IEEE 10th Int. Conf. on e-Health Networking, Applications and Service, pp. 182-186, 2008.

[49] Michela Goffredo, Maurizio Schmid, Silvia Conforto, Marco Carli, Alessandro Neri, and Tommaso D’Alessio, “Markless Human Motion Analysis in Gauss-Laguerre Transform Domain: An Application to Sit-To-Stand in Young and Elderly People,” IEEE Trans. on Information Technology in Biomedicine, vol. 13, no. 2, pp. 207-216, Mar. 2009.

[50] Maja Pantic, Alex Pentland, Anton Nijholt and Thomas S. Huang, “Human Computing and Machine Understanding of Human Behavior: A Survey,” Artificial Intelligence for Human Computing, 2007.

9

Pulse Oximetry System based on FPGA

Kien Nguyen Trung, Ngoc Nguyen Van Department of Electronic and Telecommunication

Danang University of Technology Danang,Vietnam

ntkien@dut.udn.vn, ngvng@gmail.com

Abstract—With Soft Processor, embedded DSP blocks, configurable logic blocks and high speed IOs, Field Programmable Gate Array (FPGA) are potential for a lot of high performance system on chip (SoC) applications. The paper proposes a completed method using System Generator to developing a low cost portable Pulse Oximeter System in FPGA for monitoring patient or old people where many commercial systems are working on. In FPGA hardware, many implemented preprocessing algorithms such as Led switching controller, noise filter and DC tracking, data computing run simultaneously and so that it is potential to give us very high performance, accuracy, real-time and low cost Pulse Oximeter System. Moreover, this Pulse Oximeter IP core can integrate into other IPs such as Power Monitor IPs, Security IPs to build the completed Smart House system.

Keywords- Pulse Oximeter, Blood Oxigen Saturation, FPGA, System Generator, PicoBlaze, SpO2.

I. INTRODUCTION One of biggest objectives in the development of Danang

city to year 2016 is to make the wifi network available everywhere in the city. This wifi network not only plays an important role in studying and accessing information from internet, but also constructs basic facility for many projects in information technology which applies into the citizen life such as remotely monitoring the traffic, automatically supplying tourist information, etc… To align with the network development and to take its advantages, this research develops a portable medical device to monitor online the old people’s oxygenation status. This device, with wireless module inside, can connect to the wifi network to transmit information to doctors then they can monitor the health of their patients over the internet. It can also connect to the mobile phone through bluetooth standard that helps the host software in the mobile phone monitor the customer’s health status and do the alarm action such as sending urgent messages or calling to relatives when the accident happens.

Pulse Oximetry is a rapid noninvasive measurement of arterial oxygen saturation and cardiac pulse using simple light emitter diodes (LED) and sensor application. The measurement is done directly without pre-adjustment or calibration. There are two types of pulse oximetry: the transmission type and the reflection type. Now, most of commercial products are in transmission type while reflection pulse oximetry is still in research and evaluation. The name transmission pulse oximetry means light sources are transmitted through a body part to a light receiver in opposite side to determine blood oxygen

saturation and the heart beat. With the development of modern technology, higher integrated circuit, better and smaller LED and optoelectronic sensor, the early large, heavy and expensive pulse oximeter with price around ₤5500 in the pass becomes much cheaper and more powerful personal portable devices nowadays as illustrated in figure 1.

Figure 1. Examples of portable pulse oximetry.

In the following sections, we first review the principle of the transmission pulse oximetry to understand the mechanism of calculating the blood oxygen saturation and cardiac pulse from sensor with LED, infrared LED (IR) and an optoelectronic sensor only. Then third section is our proposal about a pulse oximetry system based on FPGA technology. We will try to explain why FPGA was selected and how to build the system into FPGA. The last section discusses experimental results, conclusion and improvement solutions.

II. PRINCPLE OF REFLECTANCE PULSE OXIMETER The typical pulse oximetry configuration on a finger shown

in the figure 2 has two different wavelength light sources: red LED with 660 nm wavelength and IR LED with 910 nm wavelength, with a photo-detector in opposite positions. When only one light source is active, the incident light is the summary of transmitted light, absorbed light, scattered light and reflection light on the finger. Fortunately, commercial probes have very good mechanical structure which rejects almost scattered light and reflection light so that incident light is simplified to the summarize of transmitted light and absorbed light. The principle of pulse oximetry bases on the absorbance of lights when they are emitted through a part of body. Each wavelength light source has different light absorbance on different thickness of skin, color, tissue, bone, blood, and other material of the body part.

10

Figure 2. Typical finger pulse oximeter with two wavelengths light source is emitted, diffusely scattered through the finger and detected on the opposite

side by a photodetector. [4]

According to the Beer-Lambert law, the absorbance (A) of the wavelength light source with a molar absorptivity (ϵ) is directly proportional to both the concentration (c) and pathlength (l) of the absorbing material: A = ϵcl. Actually, there are four difference components: oxyhemoglobin in the blood (concentration co, molar absorptivity ϵo, and effective pathlength lo), “reduced” deoxyhemoglobin in the blood (concentration cr, molar absorptivity ϵr, and effective pathlength lr), specific variable absorbances that are not from the arterial blood (concentration cx, molar absorptivity ϵx, and effective pathlength lx), and all other non-specific sources of optical attenuation, combined as Ay, which can include light scattering, geometric factors, and characteristics of the emitter and detector elements. The total absorbance at the two wavelengths can then be written:

During a cardiac pulse, the absorbances that are not from

the arterial blood and other non-specific sources of optical attenuation have constant values. If we assume two blood pathlength changes are equivalent, the ratio R of time rate change of the absorbance at wavelength 1 to that at wavelength 2 can be reduced to below equation.

where I1, I2 are light intensity of received light from photo detector and I0 is light intensity of emitted light. In this case Io can be removed because it does not change in time rate.

Because the oxygen saturation is calculated by equation S = co /(co+cr) so that we can rewrite it in term of the ratio of R in following equation

,

by which the saturation was defined as shown in figure 3 (on the right). The absorptivity (ϵ) of the red LED and of the infrared LED was illustrated in Figure 3 (on the left).

Figure 3. Relationship of red R/infrared (IR) numeric ratio value to arterial oxygen saturation (SaO2). [1]

III. LOW COST PULSE OXIMETER BASED ON FPGA This section proposes a design of a low cost pulse oximeter

based on FPGA. Normally, the low cost FPGA family does not have any analog devices such as op-amp, Analog Digital Converters (ADCs), Digital to Analog Converters (DACs), discrete electric elements such as capacitors, resistors, inductors. Therefore, system has to use external devices as presented in figure 4. All external circuits are used to support preprocessing signal, amplifier, noise filter, ADC, DAC and the LED gain controller circuit. With rich logic cell resource, FPGA is suitable for implementing digital filters, finite state machine controllers, arithmetic computations and control algorithms. Moreover, a soft micro controller named PicoBlaze of Xilinx FPGA is proposed to writing the embedded program to control the operation of system. Although PicoBlaze is only an 8 bit RISC micro controller, it has many advantages included free, easy to assembler, powerful performance, and small logic size. PicoBlaze Micro-controller embedded in FPGA provides the optimal balance between FPGA and Micro-controller solution for the project that requires both higher performance and controlled state machine application.

In this proposed pulse oximetry system, the DC tracking using IIR filter and the automatic LED gain controller need to response immediately to the input signal, hence they should be implemented by FPGA logic element for parallel computing. DC signal after the IIR filter then is subtracted by the input signal to get the AC signal for computing the oxygen saturation value and heart beat. This work can run with algorithm code in PicoBlaze micro-controller but in this case, this micro-controller is assumed to interface all time with the communication modules such as Bluetooth or Wifi through common control protocol UART or SPI.

Figure 4. System design of a finger pulse oximetry bases on FPGA

11

IV. EXPERIMENT Initial experimental results showed that the maximum

amplitude of signal from pulse oximetry probe is small around 40 - 60 mV and contains a lot of noises. Therefore, integrated circuit Op-amp AD704 is considered to be used in the pre-amplifier and noise filter because it is a pico ampere input current quad bipolar Op-amp specialized for ECG/EKG instruments. Some great advantages of the Op-amp are active low pass filter inside with 0.5 µVpp 0.1Hz to 10Hz noises and very low offset voltage 15 µV. This helps the system removes low noise at the beginning stage but it protects the DC value from drifting offset for accuracy oxygen saturation computing in later phase.

Figure 5. a) A simple 3.55Hz lowpass filter and pre-amplifier. b) The 50 Hz notch filter

A simple circuit low pass filter in figure 5a) was used to remove non-information signal from photo-detector in pulse oximetry probe. There are always existence of high frequency noises from fluorescent lights or radio waves. This low pass filter was designed to have cut-off frequency f = 1/(2πRC) = 3.55 Hz and its gain is H = 100K/5K = 20. In the next stage of the preprocessing phase, circuit notch filter is very important to remove noises with frequency 50 Hz which comes from electrical grid. This type of noise exists in any stage of the circuit because of the electromagnetic interference from AC sources. Therefore, notch filter was considered to be put into the end of the pre-processing stage of the system. In figure 5b), the cut off frequency is calculated in equation fc = 1/(2πRC) = 50 Hz, of which R = R1 = R2 = R3 = R4 = 33 kΩ and C = C1 = C2 = C3 = C4 = 100nF.

Figure 6. a) Model of DC tracking filter. b) DC tracking filter designed in System Generator.

An DC filter tracking is shown in figure 6a). In fact, it is an simple IIR filter of which the value of K found by experiment has a value of about 1/f sampling rate [6]. Figure 6b) shows us the IIR filter design in development tool in Simulink, the Xilinx System Generator. This tool has all necessary block set for implement DSP system in FPGA quickly. Moreover, it has some special tools such as Resource Estimation, Timing and Power Analysis, and Matlab EDA Simulator that allows us to evaluate the product in the system design stage. It is known that Matlab EDA Simulator is a very useful tool to co-simulate between Simulink and real FPGA hardware (as seen in figure 7).

Figure 7. Co-simulation with Matlab EDA Simulator Link.

V. RESULTS AND DISCUSSIONS In the first stage of the research, the pre-processing based

on external electronic device was implemented and tested carefully. Simulation of feature DC extracting of IIR filter is illustrated in figure 8 shows that designing digital filters in FPGA using System Generation gives us not only a very fast tool but also a low cost and low power consumption method. The filter uses a few hardware resources and draws only 81 mW power supply for operating. System Generator works in the visualization environment Simulink where many powerful tools of Matlab can be used to speed up the system design and verification system.

Figure 8. IIR simulation in System Generator: Channel 1 is a input sample signal and the Channel 2 is the DC level of input signal

Figure 9a) shows us that the simple low pass filter works well in removing most of high frequency noises and pre-

(a)

(b)

(a) (b)

12

amplifies original signal with the gain 20. However, signal after the low pass filter is still contain the 50 Hz interference noise and this noise was removed completely after the notch filter as illustrated in figure 9b). Power consumption of two filters which is about 20 mA is quite good for the mobile pulse oximetry system for monitoring health status of old people. In near future, this pre-processing stage will be used to collect many data from different people whose skin color, skin thickness, and finger size are different. These data will be used to calculate range of the gain in the programmable amplifier to adapt signal to the ADC circuit. It makes sure that the variation of input signal is still in the range of the analog input to the ADC. This information decides the resolution and the frequency sampling of the ADC.

In second stage, as seen in the figure 4, DC value after the IIR has two parts DC value of Red LED and DC value of the IR LED. The difference between two values will be used to control the LED Gain Controller automatically to receive the same DC value (or the light intensity) in the receiver. In FPGA, designing a finite state machine for this work is easy and it will

work independent with PicoBlaze. Accordingly, it frees the micro-controller for focusing only on communication tasks. The more convenient is System Generator also supports block PicoBlazed MicroController that means we don’t need any more FPGA development tools to develop the completed SpO2 system into FPGA.

In the last stage of the project, analyzing the hardware resource usage and power consumption of the system is the most important work that makes the system can be commercialized. The product should be more accuracy than 95%, compared to the common commercial product. The first selected communication using in this system is Bluetooth because it is popular in most of mobile phones or laptops. That helps building software in Android mobiles to transmit old people’s oxygen status to doctor or relatives through GSM/GPRS. It is potentially a realistic solution because the system can also be converted to Very Large Scale Integrated (VLSI) circuit to be the Pulse Oximetry on Chip with very low cost in high volume.

Figure 9. a) Noise in signal from the pulse oximetry photodetector (aboved figure) and the effect of pre-amplifer and noise filter on signal. b) 50 Hz noise signal was rejected by the 50 Hz notch filter.

VI. CONCLUSIONS This paper proposed a completed method to design and

implement a Pulse Oximetry system on chip based on FPGA technology. In this method, Simulink and System Generator blocks are used in all stages of the design from design system

to implementing in to the hardware. Moreover, Matlab EDA Simulator which can co-simulate between Matlab and FPGA board, helps the designing faster, easier and more visualization. First results of this method show us pulse oximetry based on FPGA is potential compact, low cost and low power consumption that DSP equation can be implemented and

(a) (b)

(IR LED)

(IR LED)

13

evaluated quickly. In the future, the obtained results can be improved to a commercialization product of which the higher compatibility will be achieved when working with Wifi and Bluetooth available in all mobile phones, computers, and laptops.

ACKNOWLEDGMENT The authors would like to thank the Center of Excellence of

Danang University of Technology in Vietnam for their support of this research project.

REFERENCES [1] Wukitsch MW, Petterson MT, Tobler DR, Pologe JA. “Pulse oximetry:

analysis of theory, technology, and practice”, J Clin Monit, Vol 4, pp. 290-301, 1988.

[2] Yitzhak Mendelson, “Pulse Oximetry: Theory and Applications for Noninvasive Monitoring”, Clinical Chemistry, Vol. 38, No. 9, pp. 1601-1607, April 1992

[3] K. Ashoka Reddy, “Novel Methods for Performance Enhancement of Pulse Oximeters”, Phd Thesis, Indian Institute of Technology Madras, April 2008

[4] Ed. Joshep D.Brozino, “The Biomedical Engineering Handbook”, Second Edition, CRC Press LLC, 2000

[5] Serhiy Matvieyenko,”Pulse Oximeter”, Application Note, Cypress, 11/09/2005.

[6] Vincent Chance, Steve Underwood, “A Single-Chip Pulsoximeter Design Using the MSP430”, Texas Instrument, November 2005

[7] Vishal Markandey, “Pulse Oximeter Implementation on the TMS320C5515 DSP Medical Development Kit (MDK)”, Texas Instrument, November 2010

[8] PicoBlaze 8-bit Embedded Microcontroller User Guide for Spartan-3, Virtex-II, and Virtex-II Pro FPGAs, Xilinx, November, 2005

[9] Xilinx, System Generator, http://www.xilinx.com/ise/optional_prod/ system_generator.htm.

[10] MathWorks, Simulink, http://www.mathworks.com/products/simulink/. [11] “A Hands-on Guide to Effective Embedded System Design”, Xilinx

white paper, 2007 [12] “System Generator for DSP User Guide”, Xilinx white paper, 2008

14

Towards Efficient Implementation of Neural Networks with Reduced Precision Floating Point Parameters

Huynh Viet Thang, Nguyen Hai Trieu Anh Danang University of Technology (DUT), Danang, Vietnam

Emails: thanghv@dut.udn.vn; nhtanh@dut.udn.vn

Abstract— Multilayer perceptron artificial neural networks have widely been implemented on reconfigurable hardware (FPGAs) to perform a variety of applications including classification and pattern recognition. These networks are often trained with double-precision floating-point representations for input data and parameters using software like MATLAB before they are implemented on reconfigurable hardware with the optimal parameters represented in reduced-precision formats. This paper investigates the effect of reduced precision floating-point formats, used for the representation of the optimal parameters, on the recognition rate of the neural network at hand. By gradually reducing the precision of the floating-point weighting coefficients, we observed that a reduced-precision floating-point format of 4 mantissa bits is able to provide the same recognition rate as provided by the highly accurate double-precision floating-point format. Our work allows for an efficient investigation of tradeoffs in operand word-length, recognition rate and hardware resources of floating-point neural network implementations on reconfigurable hardware.

Keywords-neural network; reconfigurable hardware; floating-point; bit width allocation; reduced precision; MPFR; design space exploration.

I. INTRODUCTION Artificial neural networks have widely been used in

biometrics for the identification of humans. Typical applications of neural networks in biometrics include fingerprint and/or face classification/recognition. As of today, efficient implementations of neural networks on hardware are very desirable.

Over the last two decades, neural networks have been implemented on field programmable gate arrays (FPGAs) with fixed-point arithmetic to perform a variety of machine learning applications [1-4]. The main reason for using fixed-point arithmetic for hardware implementations of neural networks in existing work is that fixed-point numbers require less hardware resources to implement on FPGAs and provide faster learning time than floating-point numbers. With respect to numerical accuracy, floating-point arithmetic has, however, been proven as the best number format as it can guarantee a more accurate computational result. Additionally, designing with floating-point arithmetic is much more convenient for designers as overflow and underflow are automatically handled.

Despite the superior numerical benefit of floating-point arithmetic, the use of standard floating-point number formats (i.e., single precision or double precision) in implementing neural networks has not attracted many designers because of

two main reasons: i) standard floating-point formats are not area-efficient as fixed-point formats, and ii) neural networks seem to be tolerant of precision loss, i.e., a neural network implementation with 16-bit fixed-point weights could provide the same recognition performance as provided by another neural network implementation with 53-bit (double precision) floating-point weights. In the following, we briefly give a summary of related work.

A. Related work Dating back in 1991, Holt and Baker [1] conducted

extensive simulations over a wide range of data sets to compare the accuracy of limited precision fixed-point neural network implementations with the accuracy of floating-point ones. A 16-bit fixed-point format was used in their simulations. The experiments showed that the limited precision fixed-point simulation is able to perform as well as the floating-point simulation, suggesting that the 16-bit fixed-point format can potentially be the typically chosen number format for the weights of neural network implementations.

Nichols et al. [2] examine the feasibility of using floating-point arithmetic for FPGA based neural network implementations. A classic logic-XOR problem is used to benchmark the learning ability of neural networks. A 16-bit fixed-point implementation is compared with a 32-bit (single precision) floating-point implementation. It is shown in [2] that the 16-bit fixed-point implementation uses less area than and outperforms the single precision floating-point implementation in terms of training speed, suggesting that single precision floating-point format is not feasible for FPGA based neural network implementation.

In [3], a three-layer fully connected feed-forward neural network used for character recognition problem has been developed and implemented on FPGA. Each neuron uses 12-bit fixed-point weights. Experimental results reveal that the character recognition system with 12-bit fixed-point weights provides a comparable recognition rate with another character recognition system using full precision floating-point weights.

Focusing on a similar direction, Savic et al. [4] has investigated the impact of number representation on implementing multilayer perceptron neural networks on FPGAs. Again, a neural network for solving the classic logic-XOR problem is implemented. Both floating-point and fixed-point formats are studied and the effect of precision of representation and FPGA area requirements are considered. It is shown in [4] that a multilayer perceptron network with back-propagation learning uses less clock cycles and consumes less

15

Figure 1. IEEE-754 floating-point number formats.

Figure 2. A simple neural network with one single neuron.

w1

w2 w3

b

hardware resources when compiled in a fixed-point format, compared with a larger and slower functioning compilation in a floating-point format with similar data representation bit width or a similar precision and range.

Even though many authors have made comparison between using fixed-point format and using floating-point format for implementing neural networks, it seems that the use of floating point number formats for neural networks implementations and machine learning applications has not fully been studied, as another way of employing floating-point number format is not completely explored.

B. Scientific contributions Recently, some research groups, including ourselves, have

found custom precision floating-point number format very potential for achieving efficient implementation of floating point algorithms on reconfigurable hardware. More specifically, we are interested in using floating point formats with customizable bit width configurations, i.e., using the most suitable bit width for the implementation of a given algorithm, retaining the result’s required accuracy while minimizing a cost function including hardware resources, execution time and power consumption. Custom precision floating-point applications considered include dot products [5-6], support vector machines [7] and Bayesian network classifiers [8].

Motivated by recent optimistic results in using custom precision numbers, we aim to examine the feasibility of floating point arithmetic for implementing multilayer perceptron neural networks on reconfigurable hardware. Neural networks operate in two phases: back propagation for obtaining the optimal weights and forward computation for classification (i.e., we will consider a character recognition problem in this paper). In the back propagation phase, neural networks are often trained with very high numerical precision, i.e., double-precision floating-point representations for input data and weights, before they are implemented on reconfigurable hardware with the optimal parameters represented in reduced precision formats in the forward computation phase. In this paper, we only focus on the forward computation phase.

This paper investigates the effect of reduced precision floating-point formats, used for the representation of the optimal weights, on the recognition rate of the neural network at hand. By gradually reducing the precision of the floating point weighting coefficients, the optimal precision guaranteeing a required recognition rate can be identified.

To the best of our knowledge, we are the first to investigate the impact of reduced precision floating-point arithmetic on the recognition rate of neural networks. Investigation of the impact of reduced-precision floating-point formats on the back propagation phase and hardware implementation of multilayer perceptron neural networks with reduced precision floating-point parameters will be reserved for future work.

The rest of this paper is organized as follows. Section II briefly gives some background on floating-point arithmetic and multilayer perceptron neural networks followed by a motivating example showing the potential of reduced precision formats in neural network implementation in Section III. Next, Section IV presents our design of a multilayer neural network

for the character recognition problem and experimental results, which investigate the impact of reduced precision formats onto recognition rate. Optimal floating-point bit width used for efficient hardware implementation of neural networks is also identified. Finally, Section V concludes the paper.

II. BACKGROUND In this section, we briefly present the background for

further studies of custom precision floating-point neural networks in the next sections.

A. Floating-Point Arithmetic and MPFR Floating-point arithmetic is the standard approach for

approximating real number arithmetic in modern computers. The floating-point number format as specified by the IEEE-754 standard [9] consists of three fields: a sign bit (s), a biased exponent (e) and a mantissa (f). The precision p is equal the mantissa bit width plus one bit for the implicit leading one. Figure 1 shows the fields’ positions as defined for the IEEE-754 single (p = 24) and double precision (p = 53) formats.

Hardware cost is typically dominated by the size of the mantissa field. It is known that reducing the precision (mantissa bit width) of floating-point operands directly translates into increased parallelism, and if exploited correctly, to increased performance [5, 10]. Evaluation of custom precision floating-point computations can be carried out quite effectively in Matlab environment via the GNU MPFR Library [11] (version 3.0.0), in which MPFR C functions are compiled into MEX files and called from Matlab.

B. Multilayer Perceptron Neural Networks Neural networks are among state-of-the-art algorithms that

mimic how the human brain works. Neural networks were widely used throughout the 1980's and 1990's, but their popularity dropped in the late 90's due to the fact that they are computationally expensive algorithms. As of today, since computers are fast enough to run large-scale and compute-intensive applications, neural networks have a major resurgence.

16

Figure 3. Multilayer perceptron neural network

Figure 4. Effect of reduced-precision on error of sigmoid function

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240

0.05

0.1

0.15

0.2

0.25

Precision

Abs

olut

e er

ror

Absolute error versus precision of the function 1/(1+exp(-z))

The simplest possible neural network consists of a single neuron, shown in Figure 2. This neuron receives three inputs

321 ,, xxx and produces one output

)()()( 3

1,

i iiT

bW bxwfbxWfxh , where )(zf is

the activation function, b is the bias, TwwwW ],,[ 321 is a column vector representing the weights of the neuron,

Txxxx ],,[ 321 is the input column vector, and

bxWz T . Different activation function can be used depending on the inputs/outputs of the network and the application requirement as well. In this paper, we use the sigmoid activation function defined by Equ. (1).

)exp(1

1)(z

zf

. (1)

Multilayer perceptron neural networks (Figure 3) consist of many layers; each layer in turn has multiple simple neurons that are distributed and operating in parallel. The outputs of the preceding layer become the inputs of the next layer. In a fully connected neural network, the output of each neuron in the preceding layer is connected to inputs of all neurons in the next layer. In Figure 3, the leftmost layer is the input layer and the rightmost layer is the output layer.

In the forward computation phase, the computations performed at each layer include evaluating z and applying the activation function )(zf . While the activation function )(zf can be a linear or non-linear transform, the evaluation of z is a linear transform. The evaluation of z is basically a matrix-vector multiplication. This matrix-vector multiplication is based on the dot-product computations conducted for every neurons, in which the quantity zi of each neuron is the dot-product of the input vector and the weight vector.

III. A MOTIVATING EXAMPLE In multilayer perceptron neural network implementations, it

has been shown that precision loss (due to using reduced precision number formats) does not significantly degrade the recognition rate [1, 3, 4]. In the forward computation phase, two numerical operations are performed at the hidden and output layers: the matrix-vector multiplication of the weight matrix and the input vector to the respective layer, and the sigmoid activation function. As the rounding error of the

matrix-vector multiplication scales with the precision p following an exponential function p2 (see [6] for more detail about relation between rounding error of floating-point dot-products with working precision; and the dot product is the basic operator for matrix-vector multiplications), reducing one mantissa bit approximately increases the error of the matrix-vector multiplication by a factor of 2. This relation suggests that the less degradation of recognition rate of the neural network due to using reduced precision floating-point format in the forward computation phase could mainly be due to the numerical behavior of the sigmoid activation function. In the following, we present an example to verify this assumption.

We set up a simple example in Matlab to investigate the effect of reduced precision onto the resulting rounding error of the sigmoid function. We compute the absolute error of the reduced precision sigmoid function ))exp(1/(1 z when gradually reducing the working precision in unit step from single precision 24p to the minimum precision 2p . MPFR is employed for performing custom precision computations. The exponent width of reduced precision numbers is fixed at 11 bits as it is a constraint of MPFR. The double precision computed version of the sigmoid function is chosen as the reference value. The chosen input range is

]10,10[ and 1000 random numbers are used to obtain each data point corresponding to each precision in the experiment.

The experimental result is presented in Figure 4. As can be seen, there is no accuracy loss due to reduced precision formats for high precisions from 24p to 10p . From 9p , the accuracy loss starts increasing and rapidly becomes even worse with small precisions.

It is evident from Figure 4 that using a reduced precision floating-point format with 10 mantissa bits can provide the same numerical accuracy as using a double precision format for the sigmoid activation function. As the activation function mainly defines the output of the network, this motivating example suggests that the use of reduced precision floating-point formats is very potential for achieving efficient hardware implementation of neural networks. An open question is how a floating-point neural network implementation behaves when

17

Figure 5. Samples of training set (left) and testing set (right)

Figure 6. Impact of reduced precision floating-point formats on the F1 score of the character recognition problem using neural network.

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Precision

F1 s

core

F1 score versus precision for the character recognition problem

Average F1 scoreMin. F1 scoreRespective F1 score in double precision (p=53)

we reduce the precision of the weight coefficients, which will be studied in more detail in the next section.

IV. MULTILAYER PERCEPTRON NEURAL NETWORK WITH REDUCED PRECISION FLOATING-POINT PARAMETERS

In this section, we study the impact of reduced precision format onto the performance of a multilayer perceptron network used for character recognition problem.

A. Neural networks for character recognition problem There is a large number of character recognition techniques reported. Most of them use feature extraction technique, which extracts some features from individual characters, to distinguish the characters. The extracted features form a feature vector which is compared with the pre-stored feature vectors to measure the similarity. After feature extraction, many different classifiers can be used to recognize characters, such as neural networks, support vector machines and hidden Markov model.

In this paper, we directly use the pixel’s intensities as the features to a single hidden layer neural network. We build and train a 3-layer perceptron neural network with 576 neurons in the input layer, 88 neurons in the hidden layer and 35 neurons in the output layer. The back propagation technique is used for training the network in this work.

B. Database and Measurement metrics To evaluation the effect of using custom precision floating point number to the classification performance of neural networks, we use a database consisting of 6651 characters of 24 by 24 dimensions for designing and testing the neural network. The database is built by our colleagues from Danang University of Technology. The database is then divided into a training set of 80% samples and a testing set of 20% samples. In the training phase we use hold-out cross validation for tuning the parameters of the neural network; thus, only 60% of samples were actually used for training, the remaining 20% is for validating purpose. Samples for training and testing sets are presented in Figure 5.

In this paper, we use the F1 score for evaluating the test accuracy of the neural network based recognition system. The F1 score considers both recognition precision (RP), i.e., the number of correct results divided by the number of all returned results (note that in this paper the precision p denotes

the precision of a floating-point number format); and recall (RC), i.e., the number of correct results divided by the number of results that should be returned. The F1 score can be seen as a harmonic mean of RP and RC and is defined as follows:

RCRPRCRPF1

*.2 . (2)

C. Results Figure 6 presents the F1 scores versus precision obtained

by simulations with custom precision floating-point arithmetic in Matlab via MPFR library. The average and minimum (worst case) F1 scores are both reported. The F1 score corresponding to using double precision format is used for comparison. When full precision floating numbers are used, our experimental results show that the average and minimum F1 scores on the test set are 99.37% and 92.06 %, respectively.

For both average and worst cases, the recognition rate of the neural network degrades significantly when the used precision is less than or equal to 3 bits.

With a reduced precision format with 4 mantissa bits, the neural network for character recognition problem can perform as well as the highly accurate double precision format version.

V. CONCLUSION In this work, we have investigated the impact of reduced

precision floating-point numbers on the recognition rate of multilayer neural networks. By gradually reducing the precision of the floating point weights, we observed that a floating point format of 4 mantissa-bits is able to provide the same recognition rate as provided by the highly accurate double-precision format.

Our work allows for an efficient investigation of tradeoffs in the operand word-length, recognition rate and hardware resources of floating-point neural network implementations on

18

reconfigurable hardware. In fact, with a very compact floating-point format with 4 mantissa bits for character recognition problem, it could also be possible to try implementing neural network with reduced precision floating-point number format on low cost, low power embedded systems.

The approach presented in this paper can be repeated with a very little effort in other recognition and classification problems using neural network as well, for example, face recognition or fingerprint classification.

Future work will i) investigate the impact of reduced precision floating-point parameters on the learning speed of the back propagation technique and ii) focus on the implementations of artificial neural networks on hardware.

ACKNOWLEDGMENT We would like to thank Dr. Tuan V. Pham for his valuable

advice and for allowing us to use the database of characters for training and testing the neural network.

REFERENCES [1] J. L. Holt and T. E. Baker, "Back propagation simulations using limited

precision calculations," in Neural Networks, 1991., IJCNN-91-Seattle International Joint Conference on, vol. ii. IEEE, Jul. 1991..

[2] K. R. Nichols, M. A. Moussa, and S. M. Areibi, "Feasibility of Floating-Point arithmetic in FPGA based artificial neural networks," in In CAINE, 2002, pp. 8-13.

[3] M. Hoffman, P. Bauer, B. Hemrnelman, and A. Hasan, "Hardware synthesis of artificial neural networks using field programmable gate arrays and fixed-point numbers," in Region 5 Conference, 2006 IEEE.

[4] A. W. Savich, M. Moussa, and S. Areibi, "The impact of arithmetic representation on implementing MLP-BP on FPGAs: A study," Neural Networks, IEEE Transactions on, vol. 18, no. 1, pp. 240-252, Jan. 2007.

[5] M. Mücke, B. Lesser, and W. N. Gansterer, “Peak performance model for a custom precision floating-point dot product on FPGAs,” in 3rd Workshop on UnConventional High Performance Computing, 2010.

[6] T. V. Huynh, M. Mücke, "Error analysis and precision estimation for floating-point dot-products using affine arithmetic," in The 2011 International Conference on Advanced Technology for Communications (ATC2011). IEEE, Aug. 2011.

[7] B. Lesser, M. Mücke, and W. N. Gansterer, "Effects of reduced precision on floating-point SVM classification accuracy," in International Conference on Computational Science (ICCS 2011). Elsevier, Jun. 2011.

[8] S. Tschiatschek, P. Reinprecht, M. Mücke, and F. Pernkopf, "Bayesian network classifiers with reduced precision parameters," in European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2012.

[9] “IEEE standard for floating-point arithmetic,” IEEE Std 754-2008, pp. 1 –58, 29 2008.

[10] T. V. Huynh, M. Mücke, and W. N. Gansterer, "Native double-precision LINPACK implementation on a hybrid reconfigurable CPU," in 18th Reconfigurable Architectures Workshop (RAW 2011). IEEE, May 2011.

[11] The GNU MPFR Library: http://www.mpfr.org/ (last accessed: Nov 09, 2012).

19

HUMAN-COMPUTER INTERACTION USING ULTRASOUND HAND TRACKING VIA USB INTERFACE

Sau V. Tran, Tuan K. Tran, Tu T. Hoang Electronics and Telecommunication Department

Danang University of Technology The University of Danang

tranvan678,boydtvtbk,thanhtu08dt3@gmail.com

Trung X. Pham Electronics and Telecommunication Department

Danang University of Technology The University of Danang

trung_ete@dut.edu.vn

Abstract—This paper presents a hand tracking method by adopting ultra-sound sensors. Besides, the paper demonstrates a new device that can be connected to computer via USB interface in order to recognize complex one-hand gestures. The device uses the Doppler effect to characterize complex movements of one-hand through the spectrum of an ultra-sound signal. The parameter of time-delay, amplitude and spectrum at the receiver changes according to the movement of the hand such as speed, direction and location. The outstanding character of this method andgadget is useful and user-friendly because it could potentially control and drive any application on computer; furthermore, any untrained person could use it easily. (Abstract)

Keywords-Ultra-sound; Ultra-sound sensor; hand tracking; hand gestures; human-computer interaction(key words)

I. INTRODUCTION The act of gesturing is an integral part of human

communication because it may be used to express a variety of feelings and thoughts such as fury, despair, pleasure, etc…. In fact, gestures can be the most natural way for humans to communicate. In the digital era, there are more and more communication technology achievements in recognizing hand gestures as a major mode of interaction between the users and the system. Such gesture-based interfaces would operate effectively when they understand these gestures accurately. However, this is a difficult task and remains an area of active research. In order to simplify the task, gesture-recognizing interfaces in our research typically use a variety of simple assumptions.

In order to create interaction between users and the devices of the DiamondTouch, Microsoft Surface and iPhone, people have to touch their surface. This means that they require the users to hold a device and make the simplest inferences that might be deduced from the acceleration of a hand-held device. To other mechanisms, they have to make more generic inferences to classify into [1] mouse or pen-based input, [2] methods that use data-gloves, and [3] video-based techniques. Mouse/pen-based methods get highly accurate at identifying gestures; however, these methods require the user to be in physical contact with a mouse or pen (in fact, the DiamondTouch, Surface and iPhone may all arguably be claimed to be instances of pen-based methods with a special glove). In addition, such required actions could be just applied

to a few applications as well as not suitable to the disable people.

In this paper, we propose a dissimilar method based on the Doppler Effect for the recognition of one-hand gestures. Our device is structured by a single ultrasonic transmitter and a receiver. The transmitter emits an ultrasonic tone that is used to reflect hand movement. This tone will continue to be processed through a Doppler frequency shift phase that is dependent on the current velocity of the hand. Then, it forms the general characteristic of the hand velocity in multiple directions as function varying according to time. At the end, the signals are captured by the receiver to recognize the specific gesture.

Another strength of this device is its cost effectiveness, it costs under $5.33. Indeed, the ultrasound device based on gesture recognizer is really cheap because its processing signal is simple. Besides, experiments show the exact recognition of the number of vocabulary of one-handed gestures in the office environment of classification schemes and the ultrasound device is also a factor to reduce cost.

II. THE DEVICE STRUCTURE AND OPERATION PRINCIPLE The ultrasound Doppler based device used for gesture

recognition is an extension of the device suggested by Kalgaonkar and Raj [8]. The device uses the Doppler Effect to characterize complex movements of objects through the spectrum of an ultrasound signal and time among signal transmit and receiver. The reflected signal is captured by receiver sensor in order to characterize the motion of object.

A. The device structure

Figure 1: Ultrasound device.

20

Figure 1 shows a picture of our Doppler Effect-based gesture recognition device. It consists of a MSP430F5510 microcontroller and a LED indicator. The microcontroller is the main component putted at the central processing circuit to connect with the computer via USB port - 12mHz external crystal. With the aim of simplifying the design and reducing the size of the circuit, a Bootstrap loader button is added in order to download firmware from computer via USB instead of the JTAG interface. Another advantage of MSP430F5510 microcontroller is an integrated 3.3V low drop-out regulator with sufficient output to power entire MSP430 and system circuitry from 5V of VBUS.

The 40kHz-signal which is generated by the algorithm, will be taken to ultrasound transmitter through a transistor power amplifier. The power amplifier is used to control the range of the device. Long-range sensor can be used by people with disabilities efficiently.

The characteristic of the hand movement are presented by the signals and frequency shifts of the 40 kHz center receiver. This bandwidth of the signal receiver which is less than 40 kHz would be not recognized. The received signals are digitized by sampling. When the receiver is turned highly, the principle of band pass sample may be applied; then, the received signals are sampled at conversion rate of 200ksps.

All gestures to be recognized will be performed in front of the device. The sensitive range of the receiver sensor depends on the power of the transmitter and the gain of receiver circuit. A noise-remover algorithm will be added to avoid capturing random movements in the field of the sensor.

B. Principle of operation

The ultrasound device operates basing on the principle of the Doppler effect and the distance from the hand to the device.

The phenomenon capturing movement in the shift of frequency of ultrasound in response to a moving object is called the Doppler Effect. This frequency shift is proportional to source regularly and to the velocity of the moving object. In our approach, the original transmitter and listener are stationary, thus in absence of any motion, there is no change of frequency. Hand movement would be reflected by waves causing a shift in frequency. The reflected signals are then measured by the receiver sensor (fr) and can be described by the following equation which is used for Doppler radar as well as for estimating frequency changes in reflection of light by a mobile mirror.

).(vcvc

tfrf

(1)

Where,

fr: perceived frequency at receiver.

ft : original frequency from transmitter.

c: speed of sound in air.

v: velocity of target (hand).

The distance is calculated on the basis of propagation time by the following formulae:

23460010_ 6

timepropdis

(2)

Where,

dis : distance from object to device (cm).

prop_time: propagation time (us).

III. HAND MOVEMENT RECOGNITION ALGORITHM The signal which reflects on this device would be

processed to recognize the gestures of hand. Before a command is sent to computer from microcontroller, we build a program execution consists of two main components: an application program running on the computer and a processing data running on the MSP430 microcontroller.

A. Application program

Figure 3 shows the tasks of this overall block receiving data from the computer's USB port, then processing data and finally sending dummy commands to control applications.

USB Driver: Exchange data between computer and the device. Here we use the driver developed by Texas Instruments called "MSP430 USB Debugging Interface".

User Interface: The program interface and algorithm are designed by using C# and Microsoft Visual Studio SDK that aim to reduce time to develop an application.

Figure 2: Transmitted and received

Figure 3: Computer application diagram

21

Command Mapping: Translate data that is received from the device to correspond key code.

Get Active Windows: Get handle of current active window.

Send Command: Send a dummy command to control application which is active window; for example, mouse click event, mouse move event.

MSP430 sends a key code to computer by using USB communication. According to operation of the computer, USB drivers will receive and transfer data to our control application. Here, the key code will be mapped onto corresponding command. Rich set of command is really simple such as: next - back, up - down, enter, back, exit, page up, page down... Then, this command will be sent to the current active window that we focus on.

For example, when users are suffering the internet and viewing a nice picture on Chrome browser, if they wave their hand from right-to-left, the program will send a command "back" or if they wave their hand from left-to-right, the program will send a command "next" to the browser to view next image without using mouse or keyboard.

B. Data processing program USB Interface: Exchanging data between microcontroller

and computer by using USB communication is very difficult to implement. To save time for developing applications, we use the USB framework which was provided by Texas Instruments. This is a powerful tool and fully compatible with the MSP430.

Ultrasound Generator: MSP430 drives the transmitter to generate a 10-cycle burst of 40 kHz square-wave signal which derives from the crystal oscillator with a time interval of 3ms.

Distance Measurement: By measuring travel time between the first pulse at transmitter and the first received pulse at receiver, we can determine the distance between the object and the sensor.

Fast Fourier Transform: Transform data from the time domain to the frequency domain.

Command Recognition: Result of FFT algorithm will be analysed and mapped onto corresponding command such as next, back, up, down, mouse move.

After a period of 3ms, MSP430 sends a 10-cycle burst of 40 kHz square-wave signal sequence to the transmitter amplifier circuit. Ultrasound waves will spread over space with propagation velocity 346m/s.

When encountering obstacle, the ultrasound waves will react back to the receiver sensor. Here, MSP430 will measure transmission time to determine distance from device to obstacle. Then, microcontroller will get 128 samples from analog-to-digital module for FFT module. Fast Fourier Transform algorithm will transform all data from time domain to frequency domain. Command recognition module will compare the received signal spectrum with original spectrum to determine the movement of obstacle by using the principle of the Doppler effect. Finally, a key code will be sent to computer to control current active application.

The following is flowchart the algorithm of program

Two characteristic of distances and the frequency spectrum are used to recognize hand gesture. These gestures are performed within the range of the device. The orientation of the fingers and palm has no bearing on recognition or the meaning of the gesture. Figure 6 shows the sequence distance from the hand to the device during time a gesture command is executed.

Command enter-back: the undulating form according to hand moving up and down.

Command left-right: the hand is horizontal movement so that the distance is virtually unchanged.

Figure 5: Transmitter and processing

Figure 4: Data processing program

22

Command page up-page down: there is a strong variation of distance.

Figure 7 shows the spectrum of the received signal. This spectrum has two peaks:

The first peak represents the negative artifacts caused by signal frequency (<20kHz).

The second peak represents the contrast and the signal spectrum when the arm does not move.

IV. EXPERIMENTS AND RESULT Each command was tested 60 times, in four different

scenarios with condition of noise, temperature and humidity to evaluate the stable of the device.

Office: This is the most common environment with 50dB intensity of noise, temperature of 250C and air humidity of 70%.

Outdoor: 55dB intensity of noise, temperature range: 250C ÷ 350C and air humidity of 80%.

Class: 60dB intensity of noise, temperature range: 270C and air humidity of 70%.

Coffee shop: 75dB intensity of noise, temperature of 300C and air humidity of 70%.

Table 1 shows accuracy of algorithm, especially in low-noise environment. Some other environments with high intensity of noise such as coffee shop, the accuracy may be decreased but it still meets the requirement to control applications.

Command Office Outdoor Class Cafe Shop

Up 98.3% 98.3% 96.7% 96.7% Down 98.3% 98.3% 96.7% 96.7% Left 96.7% 96% 93.3% 93.3% Right 95.0% 95.0% 93.3% 93.3% Page up 96.7% 96.7% 95.0% 93.3% Page down 96.7% 95.0% 93.3% 93.3% Enter 95.0% 91.7% 91.7% 90.0% Back 93.3% 90.0% 88.3% 83.9%

Enter and Back commands are often confused with each other and with the Left, Right commands (Figure 9). The confusion could exist because the users are not familiar with usage, and the other part because of high-noise environment.

V. DISCUSSION AND FUTURE WORK This paper shows that an ultrasound can identify the good

gesture of the objects. We can control the computer without having contact with the device.

The high recognition of command performance averaged at 96.25% in the office environment. We believe that command recognition performance will greatly increased if the next version having the increasing number of sensors combining with statistical methods such as HMM, neural network, fuzzy control methods and confused phenomenon between enter and back commands is removed. At the same time, it will help typical gestures to clearly reflect on receiving device. This makes control easier.

Figure 7: Fourier spectrum

(c) Page down waveform.

(a) Enter – Exit waveform. (b) Left – Right waveform.

(d) Page up waveform.

Figure 6: the type of waveform

TABLE 1: AVERAGE % OF CORRECTLY RECOGNIZED GESTURES ACROSS THREE USERS IN FOUR PLACES.

Figure 9: Error rate and distribution for different commands.

23

REFERENCE [1] H. He, and J. Liu, “The design of ultrasonic distance measurement system based on S3C2410,” Proceedings of the IEEE International Conference on Intelligent Computation Technology and Automation, Oct. 2008, pp. 44-47. [2] Y. Jang, S. Shin, J. W. Lee, and S. Kim, “A preliminary study for portable walking distance measurement system using ultrasonic sensors,” Proceedings of the 29th Annual IEEE International Conference of the EMBS, France, Aug. 2007, pp. 5290-5293. [3] Sidhant Gupta, Dan Morris, Shwetak N Patel, Desney Tan, “SoundWave: Using the Doppler Effect to Sense Gestures”, In Proceedings of ACM CHI 2012, May 2012. [4] Murugavel Raju, “Ultrasonic Distance Measurement With the MSP430”, Application Report, October 2001. [5] S. S. Huang, C. F. Huang, K. N. Huang and M. S. Young, “A high accuracy ultrasonic distance measurement system using binary frequency shift-keyed signal and phase detection”, Review of scientific instruments, October 2002. [6] Texas Instruments, “Programmer’s Guide: MSP430 USB API Stack for CDC/PHDC/ HID/MSC”, in January 2012. [7] Silicon Labs, “Using Microcontrollers in Digital Signal Processing Applications”, Rev. 0.2 8/08. [8] Kaustubh Kalgaonkar, Rongquiang Hu, Bhiksha Raj,” Ultrasonic Doppler Sensor for Voice Activity Detection”, Mitsubishi Electric Research Laboratories, Inc., 2008 [9] Silicon Labs, “Using Microcontrollers in Digital Signal Processing Applications”, Rev. 0.2 8/08. [10] http://msdn.microsoft.com/en-us/library/k50ex0x 9.aspx [11] Kaustubh Kalgaonkar, Rongquiang Hu, Bhiksha Raj,” Ultrasonic Doppler Sensor for Voice Activity Detection”, Mitsubishi Electric Research Laboratories, Inc., 2008 [12] Kaustubh Kalgaonkar, Bhiksha Raj,” ONE-HANDED GESTURE RECOGNITION USING ULTRASONIC DOPPLER SONAR” 2009 IEEE.

24

A Real-Time Scheduling Scheme for Distributed Control System

Trong Cac Nguyen, Xuan Hung Nguyen, Van Khang Nguyen

Hanoi University of Science and Technology e-mail: cacdhsd@gmail.com

Abstract – Real-time scheduling is an important mechanism for Distributed Control Systems (DCS). The real-time scheduling is, in several networks, mainly based on static priorities associated to the messages (for example the network CAN). This static priority scheme has intrinsic limitations (which have been studied in [1]) and so, in order to overcome these limitations, we propose in this paper a real-time scheduling strategies based on three hybrid priority schemes.

Keywords - CAN bus; hybrid priority; real-time scheduling

I. INTRODUCTION DCS are gaining increased popularity because they offer

several advantages such as modular and flexible system design (e.g. distributed processing and interoperability), simple and fast implementation (e.g., small volume of wiring and powerful configuration tools), and powerful system diagnosis and maintenance utilities (e.g., alarm handling and supervisory packets). DCS contain a large number of interconnected devices that exchange data through communication networks; examples include industrial automation, building automation, office and home automation, intelligent vehicle systems, and advanced aircraft and spacecraft. The specific application imposes different degrees of timing requirements to the DCS implementation [2]. Therefore, the successful implementation of DCS depends on many variables and on a good integration of different bodies of knowledge. A distributed process control application includes three remote tasks (the sensor task, the controller task and the actuator task) which are on different sites and which then require the periodic exchange of two message flows through the network: the sensor flow which concerns the transfer of the output samples, from the sensor of the process to the controller which computes the control law; the controller flow which concerns the transfer of the control samples from the controller to the actuator of the process. The scheduling through the network of the messages of these two flows is an essential mechanism which strongly influences the settling time and the stability [3] of a process control application.

This paper is precisely concerned with this problem by considering the network CAN (Controller Area Network) [4] and the MAC layer which implements the scheduling of the frames. The scheduling is done by means of priorities which are represented in the IDentifier (ID) field of the frames. Different types of priorities (static priority, hybrid priority) can be considered. Here this paper focuses on hybrid priority schemes.

Related Work

Branicky et al. [5] studied network issues such as bandwidth, quantization, survivability, and reliability. They applied the RMS algorithm to schedule a set of NCSs. Wu et al. [6] proposed a distributed dynamic message scheduling method based on deadline of message in order to satisfy timeliness of messages and improve the system’s flexibility based on CAN. There are various resource allocation and scheduling techniques now. Simple queuing methods sample the output of the plant periodically and place the resulting data record into a first-in–first-out queue. However, the sensor sampling period must be larger than the average transmission time interval; otherwise, the queue overflows. Another one is try-one-discard, which does not implement a queue. Rather, it discards the data whenever transmission is not possible, for example, network is unavailable [7]. Martí et al. [8] showed that the codesign of adaptive controllers and feedback scheduling policies allows for the optimization of the overall QoC. They described an approach for adaptive controllers for the NCS to overcome some of the previous restrictions by online adapting the control decisions according to the dynamics of both the application and executing through message scheduling. Li and Chow [9] proposed sampling rate scheduling to solve the problem of signal fidelity and conserve the available data transmission. Al-Hammouri et al. [10] proposed an asynchronous, scalable, dynamic, and flexible bandwidth allocation scheme for NCS, formulating the bandwidth allocation as a convex optimization problem in a fully distributed manner. Hong et al. [11] scheduled a life science high-throughput platform using timed transition Petri nets and heuristic searching. Martí and Velasco [12] reviewed basic control models for flexible scheduling in real time and built a new model which allowed irregular sampling while still having better schedulability and robustness. Pereira et al. [13] dealt with scheduling real-time transmission in a wireless industrial network. They developed a new collision-free wireless medium access protocol (MAC) with static-priority scheduling. Grenier and Navet [14] highlighted a class of online scheduling policies targeted at scheduling frames at the MAC level.

The goal of this paper is mainly to evaluate the Quality of Service (QoS ) provided by three hybrid priority schemes to a process control application, to compare these QoS and to explain their differences. This study is done by using the simulator TrueTime [15] which allows to represent NCS (both network and control aspects).

25

II. CONTEXT OF THE STUDY

A. The process control application considered This application is represented on the Fig. 1: the process

control application has a transfer function 1000( 1)( ) s sG s ; the

controller is a proportional derivative controller which considers the output derivation (K is the proportional gain, K=1.8; Td is the derivative time constant, Td = 0.032s). We have an overshoot of 5% and the rise time tr is about 40ms.

Figure 1. Model

The closed loop transfer function F(s) of this application is:

2

( ) 1000( )1 ( )(1 ) (1 1000 ) 1000d d

KG s KF sKG s T s s KT s K

(1)

B. General view of the implementation through the network CAN

1) The network operates both The network operates both is presented in Fig. 2.

Figure 2. Implementation through the network CAN

Between computer 1 (Cl) in association with the numerical information provided by the AD conversion (this computer includes a task that we call the sensor task and which generates the sensor flow addressed to the controller, we note fsc this flow) and computer 2 (C2) where we have the reference and the controller (in C2 we have a task called controller task which generates the controller flow addressed to the actuator, we note fca this flow); fsc goes from Cl to C2 and fca goes from C2 to the computer 3 (C3).

And between C2 and C3 which provides numerical information to the DA conversion in front of the Zero Order Hold (ZOH). The task generating the flow fsc is time-triggered whereas the task generating the flow fca is event-triggered. Generally a network is not dedicated to only one application but shared between different

applications. In order to make a general study of the behavior of the process control application, when it is implemented through a network, we have to consider, in particular, the influence of the flows of the other applications. It is why we have, in Fig. 2, what we call the external flow, denoted as fex, which globally represents an abstraction of the flows of all other applications. We also consider that fex is periodic.

2) The choice of the sampling period (h) of the process control application is a basic action.

The sampling period has, from the control point of view, an upper bound. But from the network point of view, a value that is too small gives load that is too great. So, the choice results in a compromise. The relation 10 4

r rt th , which has been given in [16, page 66] is generally used. We consider here the bound

4rt . As tr 40ms, we have h = 10ms.

3) In a general way, the information transmission rate requested by the applications to a network is the pertinent parameter to compare the efficiency of the message scheduling.

We call here this parameter the User Request Factor (URF). Concerning the network CAN, the scheduling is done by the MAC layer and concerns the frame scheduling. By calling:

Dca, Dsc, Dex the duration of the fca frame, the fsc frame and the fex frame, respectively;

h the sampling period of the process control application (the period of fsc and consequently of fca) and Tex the period of the external flow. We have:

ca sc ex

ex

D D DURF

h h T (2)

In the context of this work, we will consider the following numerical values:

Bit rate in the physical layer of CAN: 125 Kbits/s;

Length of 10 bytes for the fsc frames and fca frames (thus a duration of Dsc = Dca = 640s);

Length of 15 bytes for the fex frames (thus Dex=960s).

The component ca scD Dh h of the URF, which concerns the

process control application and which represents the network capacity used by this application, has the value 12.8%. The use by the external frame of the network capacity will depend on its period Tex. It is this parameter that we will vary during our study in order to analyze the robustness of the scheduling of the process control application frames.

The frame scheduling in the MAC layer of CAN [4] is done by comparing the field ID bit by bit (we start from the MSB). In CAN the bit 0 is a dominant bit and the bit 1 is a recessive bit. The lower the numerical value of the field ID, the higher the priority. We consider here the standard length of 11 bits for the field ID.

G(s)

Process Output

y u

Controller Input reference

K

1+sTd

r _ +

Process to

control

Output

y

N E T W O R K

C A N

Controller

Input reference

r u

C2 fca

fsc

fex

y

C1

D A

Z O H

C3

Sensor

A D

h

26

C. General idea on hybrid priorities 1) The field ID and the scheduling execution.

The identifier field of a frame is divided into two levels, see in Fig. 3.

Figure 3. The structure of ID field (MSB: Most Significant Bit; LSB: Least

Significant Bit)

In Fig. 3: the first level represents the priority of a flow (it is a static priority specified off-line); the second level represents the priority of the transmission urgency (the urgency can be either constant or variable). The idea of structuring the ID is present in the Mixed Traffic Scheduler [17] which combines EDF (dynamic field) and FP (static field). In [3] the authors propose encoding the weighted absolute value of the error in the dynamic field (this idea is also presented in [18]) to resolve the collisions with the static field. A constant transmission urgency is characterized by a static priority (one m bit combination) specified off-line. A variable transmission urgency is characterized by a dynamic priority (which can take, generally speaking, m-bit combination among a subset of the m-bit combinations). The frames of the flows fsc and fca of a process control application have variable needs (strong urgency in a transient behavior after an input reference change, weak urgency in a permanent behavior). That is why, in this study, we consider that the dynamic priority of the frames of the flows fsc and fca of a process control application can take any m-bit combination of the set of the m-bit combinations. The scheduling is executed by, first, comparing the second level (needs predominance), and, secondly, if the needs are identical, by comparing the first level (flow predominance).

Remark: for the first level of the field ID we will consider here: Priority fex > Priority fca > Priority fsc.

2) Cohabitation of flows with constant needs and flows of process control applications (variable needs)

We have the objective of good performances for the process control applications in transient behavior. This means the urgent needs of these flows must be satisfied very quickly. For that, we impose a maximum with constant needs for the priority of these needs (concept of Pr_th for the constant needs). In this way, a strong transmission urgency of a process control application flow (dynamic priority with a very high value i.e. higher than Pr_th) will be scheduled first.

Remark: the external flow fex will have in this study constant needs (characterized by Pr_th).

D. Three hybrid priority schemes We have defined three schemes. The first is what we call

the strict hybrid priority (hp) scheme (computation of the dynamic priority directly from a function of the control signal u; re-evaluation after each sampling instant). The second is the hp scheme extended with a static time strategy (sts) for the re-evaluation of the dynamic priority (re-evaluation not always

after each sampling time). This scheme is noted hp+sts. The third is a scheme which does not compute the dynamic priority directly from the control signal u (definition of a timed dynamic priority reference profile and trip in this profile by means of an on-line temporal supervision based on a function of the control signal u). The dynamic priority is re-evaluated after each sampling instant. This third scheme, which implements a dynamic time strategy for the trip in the timed dynamic reference profile, is noted hp+dts. We now detail these three schemes.

1) hp scheme The needs are translated into a dynamic priority by

considering an increasing function of u (call it ( )f u characterized by a saturation for a value of u less than the maximum of u (noted maxu ). We do not want the dynamic priority to take its highest value only when u is maximum but already for values before the maximum, in order to react quickly as soon as the needs begin to become urgent. So we decide (it is an arbitrary choice) to take 2

3 maxu as the value of u where the dynamic priority reaches its highest value Pmax.

Several functions ( )f u have been studied, for this work we consider the function ( )f u represented in Fig. 4. This function is defined by:

2

max 3 max23 max

2max 3 max

,0

,

uP u u

f u uP u u

(3)

Figure 4. The considered nonlinear function

The computation of the dynamic priority is done by the controller each time it receives a frame that the sensor sends after each sampling instant (dynamic priority re-evaluated after each sampling instant). Then, after the reception of a frame from the sensor, the controller sends a frame with the value of the new dynamic priority. This frame reaches all the sites and as the sensor site knows the first level of the ID of fca (it is a constraint for our implementation), it will learn the dynamic priority that it will put in the next frame that it will send (the dynamic priority is then used by the two flows of a process control application) .

Taking into account the task implementation (sensor task is time-triggered, controller task is event-triggered), it is the sensor task which transmits the first frame at the start of the

second level first level

MSB LSB

m bits (n-m) bits

27

application. For this first frame, the sensor site has no information about the dynamic priority and thus we consider that it uses the maximum priority. This way, the first fsc reaches the controller site as quickly as possible.

2) (hp+sts) scheme A criticism of the hp scheme is that we can have oscillatory

behavior of the dynamic priority values (resulting from a damped sinusoidal transient behavior of u). We can have, for example, this scenario for the dynamic priority values at three successive re-evaluation instants [19]: the highest value at the first re-evaluation instant, then an intermediary value at the second, and again the highest value at the third re-evaluation instant. Such an oscillatory behavior shows that the control of a situation requiring a big value of the dynamic priority is inadequate in terms of the maintenance of this big value, since after leaving this value for an intermediary one, at the second re-evaluation instant, we come back to this big value at the third re-evaluation instant. The observation of this phenomenon suggests increasing the duration of the dynamic priority with a big value in order to improve transient behavior.

The (hp+sts) scheme is then the following. Contrary to the scheme hp, where the dynamic priority is re-evaluated in the controller site, after each reception of a fsc frame, the instant of the re-evaluation is no longer so closely related to the sampling instants. Here the duration of the time interval between two successive re-evaluations depends on the value of the dynamic priority at the beginning of the time interval. This duration must be relevant, in particular, from the point of view of the transfer function of the process control application and more precisely, of its transient behavior (defined before its implementation through the network). We considered the following algorithm:

If the dynamic priority has a value between the highest priority (Pmax) and half the highest priority ( 1

2 Pmax), we keep this value for 4 sampling intervals and we re-evaluate the dynamic priority afterwards; this duration is equal to the rise time tr which represents a good characteristic of the transient behavior.

If the dynamic priority has a value inferior to half the highest priority, we re-evaluate it after each sampling instant as in the previous algorithm.

3) (hp+dts) scheme Main ideas: We define, at first, a reference profile of

dynamic priorities for apprehending with efficiency one transient behavior (i.e. an input change or a disturbance). It consists in a continuous decreasing time function from a priority Pmax (start of the transient behavior) to a priority Pmin (end of the transient behavior and then the beginning of the permanent behavior), which gives the values of the dynamic priorities at all the sampling times (these values are decreasing).

However the only consideration of the reference profile is not enough to handle the actual behavior. In the actual behavior, we have to take into account for the influence of the network and also the possibility of successive input changes and/or disturbances which lengthen the transient behavior with

respect to the one considered in the reference profile. Then, actually, the temporal evolution of the dynamic priorities, cannot be always decreasing, i.e., at a sampling instant, we can, by considering the reference profile curve, move back to values higher than the value of the previous sampling instant.

So, in order to take into account for an actual behavior, we add a component, called on-line temporal supervision based on a function g(u) which will allow to do, with respect to the reference profile, the temporal repositioning of the values of the dynamic priorities. We consider the reference profile represented in Fig. 5.

The function P(t) is defined by:

2

max max min maxmax

( ) ,0tP t P P P t tt

(4)

Pmin is the priority used in the permanent behavior. The dynamic priority decreases slowly at the beginning of the transient behavior (we need several successive sampling instants with high priority in order to be reactive) and more quickly towards the end. The reference profile expresses that the priority, related to the sampling instants, tk, is lower than the priority related to the previous sampling time, tk-l.

Concerning the time tmax: as our objective is to tend towards a transient behavior guided by the transient behavior of the process control application without the network, we take tmax equal to the settling time at 5% of the process control application without the network.

Figure 5. Reference profile

On line temporal supervision: We have defined several functions g(u) which allow, at the sampling instant tk, to move back in the reference profile with respect to the previous sampling instants tk-l. These functions g(u) give the time values which must be subtracted to the value tk-l+h to come back more or less towards the beginning of the reference profile (then using, at the instant tk, a priority higher than at the instant tk-1). Note that the maximum of this time value can be tmax. Here we use the function g(u) represented in Fig. 6 and defined by:

2

max 3 max23 max

2max 3 max

, 0( )

,

ut u u

g u ut u u

(5)

Then max( ) 0,g u t

28

Figure 6. Example of g(u)

III. STUDY OF THE THREE SCHEMES BASED ON HYBRID PRIORITIES

A. Study conditions We consider the process control application which was

presented in the subsections II-A and II-B. The input is a position step which starts at time 0 and we study the transient behavior until it reaches permanent behavior. The QoS parameters, which need to be taken into consideration, are the mean delay D of the control loop and its standard deviation . The QoC parameter is the settling time at 5% (ts) which is obtained directly from the tool TrueTime. In order to evaluate the QoS parameters, we use, on the one hand, the message exchange temporal diagrams which are also provided by TrueTime, and the value of ts. From the message exchange temporal diagrams, we can get the delay in the control loop (delay of the message of the flow fsc+delay of the message of the flow fca+Dsc+Dca) for each sampling period (call Di this delay for the sampling period i). Counting the number n of sampling periods in the settling time ts, we deduce the value

of D and by these formulas: 1

nii

D

nD and 21

nii

D D

n

In order to make a quantitative analysis, we cause a variation in the network load (URF) by varying the period Tex of the external flow: we consider an external flow, the frequency of which (noted 1/Tex) is a multiple of the sampling frequency (1/h). The different URFs being considered are given in Table I.

TABLE I. DIFFERENT URFS

URF (%) Multiple of 1/h Tex (ms) 99.2 9 1.1111 89.6 8 1.25 80 7 1.4286

70.4 6 1.6667 60.8 5 2.0 51.2 4 2.5 41.6 3 3.3333 32 2 5.0

22.4 1 10.0

The following important points must still be emphasized:

The flows fsc (which are generated at the sampling times) and fex are synchronous (starting at the same time) and as we consider the cases where the frequency of fex is a multiple of the sampling frequency, then

their medium access attempts coincide at every sampling time.

Up to the value 70.4% of the URF (value of 1.6667ms for Tex), we can see that during Tex, one frame of each flow can access the medium: 0.96ms+0.64ms=1.6ms<1.6667ms (the third flow can begin to be transferred and then cannot be interrupted).

A last point must be still noted: at the beginning of a transient behavior, as the control signal is at a maximum, the dynamic priority of the flows of the process control application is Pmax.

B. hp scheme Concerning the process control application, we give D and

in Table II and ts in Table III. The values depend on the network load URF, and on the Pr_th.

TABLE II. HP SCHEME (QOS): D AND (MS)

Pr_th

URF 0.9Pmax 0.5Pmax 0.25Pmax

(%) D D D 99.2 5.333 1.680 3.743 2.262 1.804 1.380 89.6 3.264 1.286 2.240 1.228 1.629 0.846 80 2.48 0.887 1.987 0.828 1.5418 0.592

70.4 1.891 0.462 1.716 0.478 1.472 0.384 51.2 1.891 0.462 1.716 0.478 1.472 0.384 22.4 1.891 0.462 1.716 0.478 1.472 0.384

TABLE III. HP SCHEME (QOC): TS (MS)

URF Pr_th (%) 0.9Pmax 0.5Pmax 0.25Pmax 99.2 359 228 105 89.6 148 110 103 80 111 108 101

70.4 107 105 99 51.2 107 105 99 22.4 107 105 99

Concerning the values of D , we observe the following

main points:

1) For each value of Pr_th: For URF70.4%, we note that we have the same values

of D and whatever the value of URF is. This is a consequence of the fact that the two frames of fsc and fca, during each sampling period, can be sent during the period of fex, which is not the case with URF>70.4% where D and increase with the value of URF. We explain the difference by means of two exchange temporal diagrams provided by TrueTime (Fig. 7 and 8). On the Fig. 7, we see that the frames fsc or fca can be delayed, during a sampling period, for the duration of one frame of fex (0.96ms). On the Fig. 8, we see that the two frames of fsc and fca can be delayed and the delays for the frame of fca can be more than the duration of one frame of fex. Note then, when URF>70.4% and for increasing values of URF, D increases because the network load increases (thus more chances to delay the frames of fsc and fca).

29

Figure 7. hp scheme (URF=70.4%, Pr_th=0.9Pmax): time exchanges

Figure 8. hp scheme (URF=89.6%, Pr_th=0.9Pmax): time exchanges

2) For increasing values of Pr_th: D also increases because the dynamic priorities of the

frames of fsc and fca have fewer chances of being higher than the threshold.

3) Concerning the values of , we have the following comments:

For each value of URF, the variation of , when Pr_th increases, presents a maximum. The explanation is given by means of the Fig. 9, 10 and 11 (which represent the dynamic priority variation for Pr_th=0.25Pmax, Pr_th=0.5Pmax and Pr_th=0.9Pmax). These figures allow us to evaluate the number of times where, during the ts, the frames of fca have a higher or lower priority than the threshold (a higher priority means a lower delay; a lower priority means a bigger delay). Then we can see that we have for Pr_th=0.5Pmax, the maximum value of . For Pr_th=0.25Pmax (Pr_th=0.9Pmax), the number of times where the dynamic priorities are higher (lower) than the threshold is much greater than the number of times where the dynamic priorities are lower (higher) than the threshold. Thus, we have values of smaller than with Pr_th=0.5Pmax (in the case of Pr_th=0.25Pmax with a small value of D ; in the case of Pr_th=0.9Pmax with a higher value of D ). Obviously, for each value of Pr_th, increases with URF (the reason is still the increase of the network load). Important remark: for Pr_th0.15Pmax i.e. low threshold (we have not represented the

results for reasons of limited space), we have the minimal value for D (1.28ms i.e. a frame of fsc (0.64ms) and then a frame of fca (0.64ms) that always use the medium before the frames of fex because the dynamic priority is always higher than Pr_th during the settling time). Then, of course, =0.

Figure 9. hp scheme, URF=99.2%, Pr_th=0.25Pmax

Figure 10. hp scheme, URF=99.2%, Pr_th=0.5Pmax

Figure 11. hp scheme, URF=99.2%, Pr_th=0.9Pmax

C. hp+sts scheme For the hp scheme, we give D and in table IV and ts in

table V. The values are obviously function of URF and Pr_th.

TABLE IV. (HP+STS) SCHEME (QOS): D AND (MS)

Pr_th

URF 0.9Pmax 0.5Pmax 0.25Pmax

(%) D D D 99.2 2.589 2.138 2.589 2.138 1.28 0.0 89.6 1.856 1.152 1.856 1.152 1.28 0.0 80 1.664 0.768 1.664 0.768 1.28 0.0

70.4 1.28 0.0 1.28 0.0 1.28 0.0 51.2 1.28 0.0 1.28 0.0 1.28 0.0 22.4 1.28 0.0 1.28 0.0 1.28 0.0

30

TABLE V. (HP+STS) SCHEME (QOC): TS (MS)

URF Pr_th (%) 0.9Pmax 0.5Pmax 0.25Pmax 99.2 103 103 50 89.6 100 100 50 80 98 98 50

70.4 50 50 50 51.2 50 50 50 22.4 50 50 50

We can see important differences with the hp scheme:

1) For URF70.4% D is now always constant, whatever the Pr_th is (this is for

two reasons: the first one is because of the consequence of the property URF70.4%) indicated in the subsection III-A; the second is the fact that now, at the beginning of the transient behavior, the dynamic priority is used by the flows fsc and fca for a duration, at least, equal to 4h. Obviously as D is constant, = 0.

2) For Pr_th=0.25Pmax We have D which is constant for all URF values (this

means that, on all the network load conditions, the dynamic priority is higher than the threshold). The explanation is given by the exchange temporal diagram in Fig. 12.

Figure 12. (hp+sts) scheme (URF=99.2%, Pr_th=0.25Pmax): time exchanges

3) For Pr_th>0.25Pmax We have the same values of D and whatever the value of

Pr_th. The explanation is given by the exchange temporal diagrams of the Fig. 13 and 15 where we consider URF=99.2%. These diagrams are identical.

4) For URF>70.4% We note an increase of D and with URF (the explanation

is given by the Fig. 14 and 15); the delay of the frame fca in the Fig. 15 is higher than in Fig. 14.

With respect to the hp scheme, all the improvements (which give best settling time for the process control application) result from the fact that the dynamic priority Pmax is used a longer time. In Fig. 16, we have an example of the evolution of the dynamic priority (we have Pmax during 8h), compare the Fig. 16 with the Fig. 10.

Figure 13. (hp+sts) scheme (URF=99.2%, Pr_th=0.5Pmax): time exchanges

Figure 14. (hp+sts) scheme (URF=80%, Pr_th=0.9Pmax): time exchanges

Figure 15. (hp+sts) scheme (URF=99.2%, Pr_th=0.9Pmax): time exchanges

Figure 16. (hp+sts) scheme (URF=99.2%, Pr_th=0.5Pmax): time exchanges

31

D. hp+dts scheme We give, as for the previous schemes D and in Table VI

and ts in Table VII. We can see now that we always have the minimum constant value D (duration of the fsc frame (0.64 ms)+duration of the fca frame (0.64 ms)), then =0, and the best settling time (50 ms). This is a consequence of the fact that the dynamic priority is continuously controlled (by the control signal u) and that it is higher than the threshold for a time longer than the ts (Fig. 17).

TABLE VI. (HP+DTS) SCHEME (QOS): D AND (MS)

Pr_th

URF 0.9Pmax 0.5Pmax 0.25Pmax

(%) D D D 99.2 1.28 0.0 1.28 0.0 1.28 0.0 89.6 1.28 0.0 1.28 0.0 1.28 0.0 80 1.28 0.0 1.28 0.0 1.28 0.0

70.4 1.28 0.0 1.28 0.0 1.28 0.0 51.2 1.28 0.0 1.28 0.0 1.28 0.0 22.4 1.28 0.0 1.28 0.0 1.28 0.0

TABLE VII. (HP+DTS) SCHEME (QOC): TS (MS)

URF Pr_th (%) 0.9Pmax 0.5Pmax 0.25Pmax 99.2 50 50 50 89.6 50 50 50 80 50 50 50

70.4 50 50 50 51.2 50 50 50 22.4 50 50 50

Figure 17. (hp+dts) scheme, URF=99.2%, Pr_th=0.9Pmax

IV. CONCLUSION This study has presented the interest of a hybrid priority

strategy for the real-time scheduling on the network CAN where we have two distributed applications with different needs in terms of transmission urgency in their messages flows (variable needs for the process control application). In particular, an important characteristic in an DCS context is the capacity to implement distributed process control applications with good performances whatever the network load is. We have precisely shown that real-time scheduling strategies, based on hybrid priority schemes, allow the implementation of a distributed process control application even if the network load is important. We have considered three hybrid priority

schemes and we have demonstrated the particular interest of a scheme, call (hp+dts), with a double aspect: dynamic priority based on a temporal supervision function of the control signal of the process control application and a reference profile.

REFERENCES [1] G. Juanole, G. Mouney, and C. Calmettes, "On different priority

schemes for the message scheduling in Networked Control Systems: Definition and Analysis of a Hybrid Priority Scheme for the Message Scheduling", in Proc. 16th Mediterranean Conference on Control and Automation, France, Jun. 2008, pp. 1106-1111.

[2] J. Yépez, P. Martí, and J. M. Fuertes, “Control loop scheduling paradigm in distributed control systems,” in Proc. IEEE IECON’03, Roanoke, VA, Nov. 2003, pp. 1441–1446.

[3] G. Walsh and H. Ye, "Scheduling of Networked Control Systems", Control Systems Magazine, IEEE, Vol. 21, No. 1, Feb. 2001, pp. 57-65.

[4] Bosch GmbH, CAN specification 2.0 (A), www.semiconductors.bosch.de/pdf/can2spec. pdf, 1991.

[5] M. S. Branicky, V. Liberatore, and S. M. Phillips, “Networked control system co-simulation for co-design,” in Proc. Amer. Control Conf., 2003, pp. 3341–3346.

[6] W. Qinmu, L. Yesong, and Y. Qin, “A scheduling method based on deadline for CAN-based networked control systems,” in Proc. IEEE Int. Conf. Mechatronics Autom., Jun. 2006, pp. 345–350.

[7] G. C. Walsh, Y. Hong, and L. G. Bushnell, “Stability analysis of networked control systems,” IEEE Trans. Control Syst. Technol., Vol. 10, No. 3, May 2002, pp. 438–446.

[8] P. Martí, J. Yez, M. Velasco, R. Villà, and J. M. Fuertes, “Managing quality-of-control in network-based control systems by controller and message scheduling co-design,” IEEE Trans. Ind. Electron., Vol. 51, No. 6, Dec. 2004, pp. 1159–1167.

[9] Z. Li and M.-Y. Chow, “Sampling rate scheduling and digital filter codesign of networked supervisory control system,” in Proc. ISIE, 2007, pp. 2893–2898.

[10] A. T. Al-Hammouri, M. S. Branicky, V. Liberatore, and S. M. Phillips, “Decentralized and dynamic bandwidth allocation in networked control systems,” in Proc. 20th Int. Parallel Distrib. Process. Symp., Apr. 2006, 8 pp.

[11] T. Hong, M.-Y. Chow, P. Haaland, D.Wilson, and R.Walker, “Scheduling a life science high-throughput platform under starvation constraints using timed transition Petri nets and heuristic search,” in Proc. IEEE Int. Symp. Ind. Electron., Jun. 2007, pp. 1893–1898.

[12] P. Martí and M. Velasco, “Toward flexible scheduling of real-time control tasks: Reviewing basic control models,” in Proc. 10th Int. Conf. Hybrid Syst.: Comput. Control, Vol. 4416, LNCS, 2007, pp. 710–713.

[13] N. Pereira, B. Andersson, and E. Tovar, “WiDom: A dominance protocol for wireless medium access,” IEEE Trans. Ind. Informat., Vol. 3, No. 2, May 2007, pp. 120–130.

[14] M. Grenier and N. Navet, “Fine-tuning MAC-level protocols for optimized real-time QoS,” IEEE Trans. Ind. Informat., Vol. 4, No. 1, Feb. 2008, pp. 6–15.

[15] M. Ohlin, D. Henriksson, and A. Cervin, TrueTime 1.5-Reference Manual, Lund Institute of Technology, Sweden, Jan. 2007, pp. 1-107.

[16] K. J. Astrom and B. Wittenmark, Computer-Controlled Systems-Theory and design, Third Edition-Prentice Hall, 1997.

[17] K. Zuberi and K. Shin, "Scheduling messages on Controller Area Network for real-time CIM applications", IEEE Transactions On Robotics And Automation, Vol. 13, No. 2, 1997, pp. 310-314.

[18] P. Marti, J. Yepez, M. Velasco, R. Villa, and J. Fuertes, "Managing Quality-of-Control in network-based control systems by controller and message scheduling co-design", Industrial Electronics, IEEE Transactions on, Vol. 51, No. 6, Dec. 2004, pp. 1159-1167.

[19] H. Nguyen Xuan, “Réseaux de communication et applications de contrôle-commande,” Ph.D. dissertation, Univ de Toulouse, INSA, LAAS, Toulouse, France, Dec 2011.

32

Fall Detection Based On Hidden Markov Model

Viet Q. Truong, Hieu V. Nguyen, Tuan V. Pham Electronic & Telecomm. Engineering. Dept. Danang University of Technology,

Danang, VietNam Email: truongquocviet.genie@gmail.com, nvhieu@dut.udn.vn, pvtuan@dut.udn.vn

Abstract—Risk of falling accident in the elderly is very high. If this accident is not detected in time, it will become much more serious and the victim can’t be cured. In this paper, a novel fall detector has been built based on Hidden Markov Model (HMM). A feature set is extracted from an ellipse model which is constructed from a focused object. Besides, the fall detector using Artificial Neural Network (ANN) [5] has been rebuilt to compare with the proposed detector. The experimental results obtained from a self-built database show that: (i) The proposed detector achieves very high accuracy for clean data set which is 95.96%, (ii) For noise data sets, the accuracy is 72.41% for Test2 and 76.92% for Test3. These results are similar to the results of the detector using ANN.

Keywords: fall detection, human behavior understanding, recognition, neural network, hidden markov model.

I. INTRODUCTION There is nearly 40-60% of falling accidents causing

dangerous injuries, especially the elderly [2]. Moreover, it’s more dangerous if the elderly live in their own houses where it is difficult to detect a fall. As reported in [3], there are approximately 28-34% elderly in the community experience at least one fall every year. Therefore, automatic falling detection is really necessary to detect a fall in time. It has potential applicants in medical care as well as health-care or behavior recognition domain [4]. Many research on detecting falling accidents, which use sensors, cameras or computer technologies have been studied so far [6][7].

One of highly appropriated approaches to detect a fall is based on intelligent video analysis. According to Anderson’s report, motion history image and ellipse fitted human body are efficiently used to detect falls [8]. In image processing methods, the cameras’ positions are quite important because these positions affect the detection accuracy [9]. Furthermore, real time is always concerned in surveillance system as proposed in [10] using Template Matching (TM) algorithm. In order to improve recognition performance, more sophisticated methods have been proposed based on Artificial Neural Network [1][5] or Hidden Markov Model.

In this paper, a fall detector using HMM will be designed and tested on self-built database. Figure 1 shows the proposed detector include four steps which system processes in sequence: object segmentation, feature extraction, index selection and recognition. First step, background subtraction is used to extract object’s silhouette and Mathematical Morphology method is used to clear this binary image. Next, five features extracted from binary image sequence are Current Angle, CMotion, CTheta, Eccentricity and CCentroid. Then, these features will be clustered based on Euclidean distance to every codeworks in the codebook has been built previously. Finally, HMM algorithms are used to detect a fall.

II. OBJECT SEGMENTATION

A. Background subtraction First, object segmentation block is responsible for getting

silhouette images using background subtraction method [10].

From [12], a certain pixel Pi from current frame Ii is marked as foreground if

|Pi – Bi | > τ

where Bi is same position pixel of background frame Bi and τ is a “predefined” threshold. And, the updated background Bi+1 is calculated by:

Bi+1 = αIi + (1 − α)Bi

where α is chosen 0.05 as in [10].

B. Pre-processing Next, the pre-processing block removes redundant pixels

from images which result from previous block as well as noise. Mathematical morphology method used for smoothing silhouette images.

III. FEATURE EXTRACTION

A. Ellipse model Ellipse model is usually used for tracking object because it

is quite easy to draw an ellipse around object [13]. To build an ellipse, proposed system determines 3 parameters as follow:

1) Centroid of ellipse For each binary frame, the centroid coordinate of ellipse

O(Ox, Oy) is determined: Abscissa (Ox) and ordinate (Oy) are average of the all x coordinates and the all y coordinates of the white pixels [9].

2) Current angle This parameter is also called vertical angle which is the

angle between horizontal line and major axis.

Figure 1. Proposed system

Feature extraction

Output info.

Object segmentation

Index selection

Calculate likelihood to recognize

Trained HMMs

Codebook

Input video

(1)

(2)

33

i j j).P(i,2yi j j).P(i,2xi j j)x.y.P(i,2

.arctan21θ (3)

where

i, j : position of pixel (i from 1 to width of frame, j from 1 to height of frame).

x = i – Ox and y = j – Oy (Ox, Oy: position of centroid).

P(i, j) : value of pixel (i, j).

The formula (3) derived tangent of double angle (2θ) is used so that vertical angle of object is more stable [9].

3) Major and minor axes of ellipse Major and minor axes are determined as [1].

B. Feature extraction 1) Current Angle

This feature is calculated as formula (3). Furthermore, system would well-control if current angle is limited.

In figure 2, the non-cross area is the limited domain of current angle. System modifies so that current angle belongs to non-cross area. For example, current angle is 2250 which system modifies to 450 by subtracting 1800. Figure 2 describes Theta sequences in different activities.

2) Coefficient of motion (Cmotion) The object’s motion rate is determined by considering 15

successive frames; then, system creates an image that saves the object’s motion history [14].

pixel White pixelGray

pixelGray motionC

(4)

Figure 3 shows a Motion History Image which is created from synthesizing a sequence of 15 successive frames. Figure 4 presents the changes of Cmotion in different activities.

3) Deviation of vertical angles (Ctheta) This parameter is standard deviation of a buffer of 15

successive frames (figure 5 describes the different Ctheta of object’s different activities) [1].

4) Eccentricity As reported [9], this parameter is really efficient when

detecting a direct fall (figure 6 describes the different eccentricity of object’s different activities).

5) Deviation of centroid (Ccentroid) Ccentroid is defined as standard deviation of 15 ordinates

of 15 successive centroids. Figure 7 show the different Ccentroid sequences of object’s different activities).

Figure 3. Motion history image

Figure 4. Cmotion in different condition.

Figure 2. Current angle in different condition.

Figure 5. Ctheta in different condition.

Figure 6. Eccentricity in different condition.

Figure 7. Ccentroid in different condition.

34

IV. RECOGNITION WITH HIDDEN MARKOV MODEL In this proposed system, recognition block based on HMM

is used to detect a fall. A popular HMM is specified by a set of state S = s1,s2,s3 ….,sN and a set of parameters Ө = π,A,B, where vector π contains the prior probabilities, A is the transition probability matrix and B is the density probabilities [11]. The following subsections present HMM training and testing of recognition block.

A. Training 1) Codebook generation

From three extracted features, it’s necessary to use vector quantization in order to reduce feature-vectors’ dimensions. This quantization generates a codebook which is used to change a 5-feature vector into certain index number. Therefore, there are 2 steps to create the codebook:

Extracting five features from frames of all training clips included fall and non-fall activities and putting all training feature vectors into the concatenation of 5-columns matrix.

Then, K-means clustering algorithm is used to generate a codebook which is used as index mapping table for all built HMM models. K-means clustering (MacQueen, 1967) is a method commonly used to automatically partition a data set into k groups which have k initial arbitrary centers of clusters. This algorithm iteratively refines centers of clusters to generate a codebook [16].

Figure 8 shows how to produce a codebook through K-means clustering algorithm.

2) HMM models In this work, the obtained symbol sequences are used to

train HMMs to learn the proper model for falls and other activities. Five states left to right HMM (Figure 9) is used to build fall model and non-fall model. Two HMMs are trained based on Baum-Welch algorithm. B. Testing

Testing progress (also recognition) in proposed system is illustrated as following flow chart

After training, proposed system uses two HMMs (fall model and Non-fall model) in recognition block to detect a fall by processing two main blocks: clustering and HMM decoding.

1) Clustering Before HMM decoding, recognition block determines index

by clustering block. At first, clustering block calculates Euclidean distances from all code-words to feature vector of current frame. Then, index of current frame is assigned as index of code-words which has minimum Euclidean distance to feature vector (figure 11 shows flow chart of clustering block) [17]. After that, recognition block checks if index-buffer is full of 20 values. If not, object’s action seems to be normal and clustering block continues getting index of next frame.

If index-buffer is full of 20 indexes, this buffer will be added new index at the end and removed first index. Then, it will be decoded by 2 HMMs (fall and non-fall models) in HMM decoding which is presented following section.

2) HMM decoding and detecting

Joining all feature vector

K-means clustering Codebook

Feature vector

sequence

Figure 8. Codebook generation.

From Feature Extraction block

Clustering

Fall detection

Y

HMM decoding

Figure 10. Flowchart algorithm for testing.

Feature vector from feature extraction block

Full index-buffer?

Index

N

Full label-buffer?

Label

No. Fall label>Th?

Normal action

N

N

Y

Y

Add new index at the end

Add new index at the end and

remove first index

Find minimum distance

Index Euclidean distance to every

codeword

Figure 11. Clustering block and index selection.

Feature vector

1 2 3 4 5

Figure 9. Five states left to right HMM.

35

After checking full index-buffer, 20-index buffers are used for decoding both two built models to find which fall model or non-fall model is more likelihood

As a result, the output of every 20-frame shift-buffer is “0” (if non-fall model is more likelihood) or “1” (if fall model is more likelihood). Another buffer (called the label-buffer) is responsible for storing 20 consecutive labels. When the total number of “1” in this buffer is greater than “predefined” threshold, the falling incident is detected.

V. RECOGNIZE WITH ARTIFICIAL NEURAL NETWORK According to [5], a two-layer feed forward ANN is

initialized. The input layer consists of either 5 inputs (ANN1) or 100 inputs (ANN2). And the results of the study in [5] indicate that ANN2 is more reliable than ANN1 because the inputs of ANN2 are sequence of feature which are extracted in a period of 20 consecutive frames. Therefore, in this study, ANN2 is rebuilt to compare with HMM. The optimization criterion was chosen to be the minimization of the mean square error (MSE) derived over the whole training set. The training set is first split up into a training subset and a validation subset. The training of the ANN only stops if the MSE derived from the validation subset could not be reduced within 5 consequent training epochs. The Scale Conjugate Gradient (SCG) algorithm is used for training the neural network with hidden layers is fixed at 50 units and validation subsets is 20% [5].

VI. EVALUATION

A. Database In this system, the DUT-HBU database used in our

previous work [20] has been reused. A data set containing 217 video clips as being depicted in Table I is constructed. There are 107 video clips in which falling actions are recorded, and the rest of 110 video clips are recorded non-falling ones. While each falling video contains two actions including walking and falling, each non-fall video can contain one of actions shown in Table I.

TABLE I. CLASSIFYING VIDEOS ACCORDING TO ACTIVITIES

To facilitate the comparison and evaluation, the algorithms are both trained with clean data (Scenario 1) and tested with all three testing data subset.

The Test1 subset: Contents and activities in the video clips for training are very similar to the ones for testing. In each clip, there is only one object with stable background. This set has 12 falling videos and 11 non-fall videos.

The Test2 subset: Includes videos that are similar to training videos but the brightness of the Test2 videos is not good or positions of the camera has changes. This set consists of 15 falling videos and 14 non-fall videos.

The Test3 subset: They have much changes compared to training videos. It appears naturally. For example, part of the object is obscured, background isn’t stable, or there are more than two motion objects in these video.

B. Evaluation criteria In this study, for comparing and evaluation, the following

statistical measures are used: Recall (RC), Precision (PR), Accuracy (Acc).

Recall or Sensitivity is the proportion of Real Positive cases that are correctly Predicted Positive.

Precision or Confidence denotes the proportion of predicted Positive cases that are correctly Real Positives. This statistic measures is calculated as follows:

The accuracy is the proportion of true results (both true positives and true negatives) in the population. This statistic measures is calculated as follows:

Where TP, TN, FN, FP are defined as follows:

True positives (TP): amount of fall actions which are correctly classified as fall.

False positives (FP): amount of non-fall actions which are wrongly considered to be fall.

False negatives (FN): amount of fall actions which are wrongly rejected and classified as non-fall actions.

True negative (TN): amount of non-fall actions which are correctly classified as non-fall.

C. Experimental results 1) Result of the proposed detector using HMM

Figure 12 shows the result of the proposed detector using HMM. For clean data set, the result is very high. Recall, precision and accuracy are 100%, 92.31% and 95.95% respectively. Statistical results getting decreased dramatically in the noise data. The accuracy is only 72.41% for Test2 and 76.92% for Test3.

Scenario1

Scenario2

Test1

Test2

Test3 ALL

Cross (Fc) 4 18 4 4 10 18Direct (Fd) 4 19 4 6 9 19Side (Fs) 7 17 4 5 7 16Cross (Ncb) 1 4 1 1 1 3Direct (Ndb) 3 5 1 1 1 3Side (Nsb) 1 3 1 2 2 5Cross (Ncc) 1 3 1 2 1 4Direct (Ndc) 2 4 1 1 1 3Side (Nsc) 1 4 1 1 1 3Cross (Ncl) 1 3 1 1 2 4Direct (Ndl) 3 5 1 1 0 2Side (Nsl) 1 4 1 1 2 4Cross (Ncs) 0 2 0 1 2 3Direct (Nds) 3 6 1 1 1 3Side (Nss) 1 4 1 1 1 3

0 12 0 0 11 1133 113 23 29 52 104

Sitting

DATABASEtrain test

Other (No)SUM

Fall

NonFall

Bend-ing

Creep-ing

Lying

(7)

(5) (6)

36

2) HMM in comparison with ANN for fall detection Figure 13 shows that: (i) For clean data in Test1, the results

of HMM and ANN are the same, (ii) For noise data, ANN has higher accuracy than HMM in Test2 but HMM is better in Test3. In general, the results of ANN and HMM are equivalent.

VII. CONCLUSION In this system, most of the recognition algorithms are very

high results for the video was made in good environmental conditions. This indicates that five extracted feature will provide the reliable information if the object segmentation works well. For the noise video in Test2 and Test3, Template Matching algorithm gives the very bad result, the result of ANN1 (5-features input) is also low [15]. ANN2 with inputs are features were extracted from 20 consecutive frames and HMM have better result for noise data because the variation of parameters over time is expressed. However, for the healthcare and warning system, these results are still quite low. These results are quite low due to the object segmentation algorithm is relatively simple, it can’t extract and tracking multiple objects at the same time. So, when the environmental conditions are not good or there are many objects in the video, the results will not be as expected. The results can be greatly

improved by using the better algorithm to extract the object and develop more sustainable attributes.

ACKNOWLEDGMENT The authors would like to thank the research group at DUT:

Hoan V. Tran, Phung K. Lai, Khue Tra, Duy H. Le.

REFERENCES [1] Y T. Ngo; Hieu V.Nguyen; Tuan V. Pham, “Study On Fall Detection

Based On Intelligent Video Analysis”, IEEE The International Conference on ATC 2012.

[2] J. Teno, D. Kiel, and V. Mor, “Multiple strumbles: a risk factor for falls in community-dwelling elderly,” J. America Geriatrics Society, vol. 38, no. 12, pp. 1321-1325, 1990.

[3] Chia-Wen Lin, Zhi-Hong Ling “Automatic Fall Incident Detection in Compressed Video for Intelligent Homecare”, Computer Communications and Networks, 2007. ICCCN 2007. Proceedings of 16th International Conference, pp. 1172-1177.

[4] Chieh-Ling Huang, E-LiangChen and Pau-Choo Chun “Fall Detection Using Modular Neural Networks With Back-Projected Optical Flow”, Biomedical Engineering: Applications, Basis and Communications, Vol. 19, No. 6 (2007) 415–424.

[5] Y T. Ngo, Phung K. Lai, Tuan V. Pham, “A single-camera fall detection using neural network”, United Kingdom Vietnam Advance Surveillance Systems workshop 2012.

[6] Young-Sook Lee; Wan-Young Chung, "Automated abnormal behavior detection for ubiquitous healthcare application in daytime and nighttime", Biomedical and Health Informatics (BHI), 2012 IEEE-EMBS International Conference on , vol., pp.204-207, 5-7 Jan. 2012.

[7] Arie Hans Nasution; and S. Emmanuel,” Intelligent Video Surveillance for Monitoring Elderly in Home Environments”, International Workshop on Multimedia Signal Processing, Greece, October 2007.

[8] D. Anderson;, “Recognizing Falls from Silhouettes”, Proceedings of the 28th IEEE EMBS Annual International Conference, 2006.

[9] J. Q. Lin, “The Behavior Analysis and Detection of Falling”, Master Thesis, Department of Computer Science and Information Engineering, National Central University, Taiwan, 2004.

[10] N. Chan, A. Patel, A. Chandrasekhar, H. Lee, “Fall Detection (ECE4007L05),” Georgia Institue of Technology, April 2009.

[11] Barbara Resch, Hidden Markov Models A Tutorial for the Course Computational Intelligence..

[12] M.M.Ivor, A. “ Background subtraction techniques ”. [13] Nait-Charif H, McKenna SJ, Activity summarization and fall detection

in a supportive home environment, in Proc of the 17th Int Conf on Pattern Recognition 4:323–326, 2004.

[14] Muhammad Jamil Khan and Hafiz Adnan Habib "Video Analytic for Fall Detection from Shape Features and Motion Gradients", Proceedings of the World Congress on Engineering and Computer Science 2009 Vol II, WCECS 2009, October 20-22, 2009, San Francisco, USA.

[15] Khue Tra, Viet Q. Truong, Tuan V. Pham, “Threshold-based versus Artificial Neural Network for fall detection”, Danang University Conference, 2012.

[16] Kiri Wagstaff, Claire Cardie, Department of Computer Science, Cornell University, Ithaca, NY 14853 USA, Constrained K-means Clustering with Background Knowledge,2001.

[17] Md. Zia Uddin, Nguyen Duc Thang, Jeong Tai Kim, and Tae-Seong Kim, Human Activity Recognition Using Body Joint-Angle Features and Hidden Markov Model,2011.

[18] Saleh Alaliyat, Department of Computer Science and Media Technology, Gjøvik University College, ‘Video - based Fall Detection in Elderly’s Houses’, 2008.

[19] N. Chan, A. Patel, A. Chandrasekhar, H. Lee, Georgia Institue of Technology, “Fall Detection (ECE4007L05)” April 2009.

[20] Hoan V. Tran; Phung K. Lai; Khue Tra; Viet Q. Truong; Duy H. Le;, “DUT-HBU database”, Danang University of Technology, 2012.

Figure 12. Result of Hidden Markov Model.

Figure 13. Accuracy of HMM and ANN.

37

Measurement of Biological Concentration Using Magnetic Agents

Cao Xuan Huu, Pham Van Tuan

Department of Electronic and Telecommunication Danang University of Technology, UD

Danang, Vietnam

Dang Duc Long, Nguyen Thi Minh Xuan, Pham Thi Kim Thao

Department of Chemistry Danang University of Technology, UD

Danang, Vietnam

Abstract—Detecting and measuring the concentration of biological agents are of interest in many biomedical applications. This paper presents a method for measuring of that concentration with the aid of magnetic nanoparticles. Here biological agents (proteins or DNAs) were attached to magnetic nanoparticles in specific binding ratios. The attached biological agents interacted with their specific targets in solutions and allowed those targets’ concentrations being measured based on the number of nanoparticles determined via the measurement of their induced magnetic field. A sensitive magneto-biosensor using giant magneto-resistance sensing element together with a new sensing technique has been introduced and employed as a novel method for a low magnetic field measurement. The technique demonstrates that ultralow biological concentration can be detected and measured in a timely fashion without the need of a special shielding medium.

Keywords-ultralow magnetic field, biosensor, sensing technique, GMR, magnetic nanoparticle

I. INTRODUCTION Detection and measurement of biological agents like

proteins or DNAs are of interest in most of biomedical applications. In many methods of protein detection platforms, the binding event of a protein to a specific recognition molecule must be detected with a signal transducer. In ELISAs, protein microarrays [1], and quantum dot [2] detection platforms, the readout is based on a fluorescent or colorimetric signal. However, many biological samples or reagents exhibit inherent autofluorescence or optical absorption which in turn becomes a major limiting factor of the measurement. Similarly, other detection methods using nanowires [3], microcantilevers [4], carbon nanotubes [5] and electrochemical biosensors [6] rely on charge-based interactions between the protein or tag of interest and the sensor, making each system unreliable in conditions of varying pH and ionic strength.

Over recent years, magnetic detection of biological agents and activities has gained much attention and be one of the top interests. The primary attraction of the biomagnetic method is its ability to assess non-invasively the state of health of various physiological systems [7]. The strength of biomagnetic signals is nearly zero (for general proteins or DNAs) or of many orders of magnitude smaller than that of the earth’s magnetic field (for other biological agents). Furthermore, magnetism of biological agents, if exist some, will not interfere with signal magnetic detecting transducer mechanism. Therefore, biomagnetism

detection using magnetic field becomes appropriate in detection and measurement of biological concentration [8]. Recently, magnetic sensing techniques exploit a broad range of ideas and phenomena from the fields of physics and material science. For detecting and measuring biological agents like as DNAs, proteins, etc, a diverse range of biosensor utilizing the MTJ, GMR, GMI materials has been put into focus [9]. Normally, those sensors work with the static field of the agents or measure statically the agents with a biasing technique. It is important to note that geomagnetic field noise (GMN) is of about 0.1 nT and has a 1/f-like frequency spectrum for a long spatial range (about kilometers). Thus, to get absolute value of biomagnetic signals, one must shield the system out of the effect of the noise. However, the absolute values need a clear and unchangeable off-set point, of which definitely unreachable. In addition, either the shielding is inconvenient as carrying out experiments in narrow shielding space or having a construction cost for a large enough shielded space of which few laboratories can effort. To overcome the difficulties, an advanced method has been explored in which the difference between the readings of two or more separated sensors has been taken into measurement. With this technique, it is possible to measure magnetic field changes smaller than the GMN with high reliability and without the use of magnetic shielding. Nonetheless, the method requires at least two sensors and gives an output just a single reading.

Recently, researchers working in biomedicine and on biosensor platforms have begun to examine the potential of magnetic nanoparticles (MNPs), on the scale of 5–150 nm, for biological labeling, drug targeting, hyperthermia, and as MRI contrast agents. For a recent review of MNPs employed in these and other biomedical applications the interested reader is directed to [10]. In the present study, the ultralow field detection assay is presented of which the application is focused onto biological agent detection with the aid of the magnetic nanoparticles. Specifically, in this study the biological agent to be detected and measured was proteins. The measuring technique expressed a promising capability of overcoming the week points mentioned above. There are a plethora of synthesis and coating strategies, and these are covered in more detail in a recent review article [11]. Therefore, here we will only touch on the basic concepts of MNPs for use in biomedicine and then introduce the method for measuring biological concentration via the measure of an ultralow magnetic field with the aid of MNPs.

38

II. BIOLOGICAL ATTACHMENT AND MEASUREMENT PRINCIPLE

A. Binding Progress of Proteins onto MNPs In the configuration illustrated in Figure 1, an immobilized

antibody adheres to the immobilization substrate and makes a connection with biomarker and biotinylated antibody in a ‘sandwich’ structure which in turn attached to avidin and nanoparticle by a specific reaction. For this scheme, antibodies on the substrate specifically trap the biomarker of interest from a sample. The substrate is then washed with a solution containing more biotinylated antibodies, which also bind specifically to the biomarker. These antibodies are tagged with biotin molecules, which bind to the protein avidin. When avidin-coated MNPs are passed over the substrate, they adhere to the biotinylated antibody. Finished sample is then put into sample holder and exposed to an external magnetic field created by two permanent magnets for measurement.

(a) (b) (c) (d) (e) Figure 1. Binding progress of a biological agent of interest to magnetic

nanoparticle.

The MNPs in attachment with biological agents are induced by the external magnetic field, for which the magnitude of the induced field will correlate with the number of bound nanoparticles [12]. Therefore, the purpose of the using MNPs is to magnetically mark the biological agents of interest and to create an output signal. The signal stays in a range around GMN regime. A sensor will measure the change in magnetic induced field of nanoparticles in such small regime using an advanced measuring technique presented in the next section.

B. Sensor Detecting Principle Conceptually, the easiest input circuit to consider for

detecting changes in magnetic fields is a pure magnetometer. However, magnetometers are extremely sensitive to all magnetic signals in the environment. This may be done without wondering when measured inside a well-shielded space [13]. A method which is frequently used is to employ another sensor to measure ambient fields then taking the difference between two sensors. However, if the magnetic signal of interest is weak, then environmental magnetic interference may prevent measurements. Eventually, a detecting method has been developed in which the measurements can be done in a common space, without shielding, with the help of MNPs.

In order to measure the change in magnetic field created by sample which is smaller than GMN, here we employ only one sensor and take the advantage of a scanning system with a controlled frequency. For this configuration, the sensor is located at fix position and the sample is moving back and forth

with certain amplitude and frequency which is controlled by a scanning mechanism. Figure 2 shows the principle of the measurement using one sensor. In this setup, the measuring signal after each cycle has been collected and then be taken the mean value after a number of cycles for the final record. The ultralow magnetic field of MNPs creates a change in total magnetic field around sensor head at a frequency of 2.0 Hz. This specific signal is taken into measurement without difficulty.

S

Bias field Moving sample

Scanning

Sensor head

Magnetizing field N

Figure 2. Configuration setup for the measurement of MNPs' induced field.

III. MATERIALS AND MEASUREMENTS

A. Synthesis of Iron Oxide Nanoparticles MNPs which are iron oxide nanoparticles in this study were

synthesized by coprecipitation methods from a mixture of FeCl2 and FeCl3 (1:2 molar ratio). Typically, 0.86 g FeCl2.4H2O and 2.35 g FeCl3.6H2O were mixed in 40 mL de-ionized (DI) water and heated to 80oC. 5 mL of NH4OH was added by syringe while vigorously stirring the reaction mixture and then heating continued for 30 min. After that, for the nanoparticles-coating SiO2, 1 g of silicon oxide anhydrous in 2 mL DI water was introduced, the temperature was increased to 95oC and stirring continued for 30 min. The nanoparticles was washed three times by DI water and kept at 4oC for long time use.

B. Preparation of Protein-coating MNPs Protein used in this research was enzyme protease Alcalase

purchased from Sigma Company.

Iron oxide nanoparticles with or without SiO2-coating were dissolved in protein-containing buffer (100 mg/mL, pH=7.5) to a final concentration of 10.8 mg/mL. This mixture was incubated overnight with lightly shaking at 4oC. After that, the excess protein was taken out. The iron oxide nanoparticles were washed three times by PBS (phosphate buffer saline) and used to determine the amount of proteins which were absorbed on nanoparticles by Bradford assay.

To protect linking between nanoparticle and protein from the effect of environment such as solution pH and ambient temperature, alginate coating was employed. Briefly, total above nanoparticle-absorbed protein complex was diffused into

39

1 ml sodium alginate solution (2 %). The mixture of nanoparticle-protein alginate was added dropwise into calcium chloride solution (0.2M) with continuous shaking at 4oC to form beads. These beads were washed three times with NaCl (0.9 %) solution and stored in NaCl (0.9 %) solution at 4oC for long time use. The biological specimens have been prepared in which the proteins were caught with biomarkers through biotinylated antibody and avidin, and fixed onto MNPs by a specific bonding mechanism [14]. Surface-functionalized MNPs are widely employed in various protein or DNA detection systems as immobilization platform with interest for purification/incubation processes [15].

C. Magnetic Measurements Sensor uses Giant Magnetoresistor (GMR) element to

detect an ultralow magnetic field caused by adhering MNPs in a magnetizing field. GMR sensor requires an uniform bias field created by an electromagnet couple for a linear detecting region ranging from 800 A/m to 2400 A/m. MNPs have been induced by a magnetizing field in vertical direction, which is produced by two permanent magnets as shown in Figure 2. The magnetizing field at sample has magnitude of 3.2 kOe, or 254.8 kA/m.

In order to protect unexpected contact with biological specimen, the sensor head and the specimen were separated by a cover glass of 120 µm thick. The distance from the sample to the upper sensor head was approximately 0.5 mm. The specimen was attached onto the slider of the scanning system, which was moving at 2.0 Hz with amplitude of 5 mm. The signal detecting was taken for 1 minute, which was of 120 cycles.

IV. RESULTS AND DISCUSSION Two types of MNPs, with and without SiO2, have been

fabricated. Magnetic measurements done by using Vibrating Sample Magnetometer (VSM), as shown in Figure 3 for their magnetization curves, identify that MNPs exhibit superparamagnetic properties.

Figure 4 shows one of transmission electron microscopy

(TEM) images for the obtained MNPs without SiO2 coating.

One can see that the particle size drops in range of approx. 8 – 12 nm. The inset in Figure 4 expresses the X-ray pattern of the sample of which the diffraction peak (311) happened at 2 36o (where with arrow mark) shows clear evident for a theoretical particle size of 8.2 nm when calculated by Scherrer’s formula [16].

Figure 5 presents the amount of proteins physically

absorbed on MNPs with or without SiO2. Diluted concentration range of bovine serum albumin (BSA) from 0-0.7 mg in 1 mL NaOH 1N plus 10.8 mg MNPs were used to make the standard line in Bradford assay (Figure 5a). The trendline and equations were generated by excel solver. After incubated with protein overnight, MNPs were washed and diffused in NaOH 1N. Dilution range of MNPs in NaOH 1N to 1:2, 1:4, 1:8 were prepared to determine concentration of proteins absorbed on MNPs (Figure 5b) based on the equations in Figure 5a.

Figure 5: The amount of proteins physically absorbed on MNPs with or without SiO2. Conc. stands for concentration; and Abs. for absorption.

Figure 4. TEM image of MNPs. The inset is the X-ray pattern for MNPs.

Figure 3. Magnetization curves measured at room temperature for two

MNP samples

40

Experimental results shown in Figure 5 revealed that while 10.8 mg nanoparticles without SiO2 coating absorbed 0.8 mg proteins, 10.8 mg nanoparticles surrounded by SiO2 shell absorbed 2 mg proteins. This means SiO2 shell increased affinity between MNPs and proteins. The hydrophilic properties of SiO2 may explain for this phenomenon. Indeed, Jin Chang et al. demonstrated that the core-shell MNPs-SiO2 penetrated into tumor cells more effectively than MNPs without SiO2 even the size of MNPs-SiO2 beads were more than MNPs’ size [17]. Similarity, composite of methylene blue and SiO2 increased enzyme (horseradish peroxidase) loading [18]. In our technique, the rate-limiting steps are the binding of targets to the substrate and antibody-linked nanoparticles to the target.

The beads containing MNPs and proteins still showed

superparamagnetic properties and catalyst activity of proteins, as seen in Figure 6. Alginate shell at concentration of 2% did not show any effect on activities of MNPs and proteins.

Due to the fix position sensor and the sample is moving relatively to it, the signal induced on sensor head is independent on sample scanning speed. Therefore, a higher scanning speed will give a shorter measuring time. However, the scanning speed of a mechanical system is limited by the inertia of the moving part. With a constant driving power supporting to the moving slider, higher scanning speed will bring shorter scanning amplitude. In this work, maximum speed is limited to about 5.0 Hz. For an optimum operation, scanning frequency was chosen by 2.0 Hz. One of the advantages of this measurement technique was its short measurement time. Suppose the data will be collected after one minute, 120 data points are taken into calculation for a mean value.

Initial results measured for the large amount of MNP samples (above 0.1 mM) express a good linearity of the concentration measurement. The MNPs’ induced field which was in range of 10-4 – 10-1 A/m can be detected and measured by only one GMR sensor using a special scanning technique. The results can be refined in the continuation of research when the optimization is considered. Experiment also confirms that the bio-concentration measurement with the aid of magnetic agents, MNPs, has not been affected by magnetically interference of biological molecules to be measured. Four samples have been prepared with the same initial MNP amount of 0.3 mM. Two first samples were original MNPs, one without SiO2-coating and the other with SiO2-coating. The last two samples were of the first samples with protein attachment. These four samples were put into measurement at the same settings and the obtained results were shown in Figure 7.

From Figure 7, it can be concluded that the proteins to be measured contribute zero biomagnetism effect on MNP’s induced magnetic field. The fluctuation in signal is due to the reduced signal of MNPs with SiO2 coating compared to that of MNPs without coating, as seen clearly in magnetization curves in Figure 3. This initial result show the error bar at each data point. It is the uncertainty determined by magnetic field measurements where somewhat of electromagnetic noise influenced on the output of the sensor. This implies that the uncertainty of measurement was in order of about 10-2 nM.

V. CONCLUSIONS This paper presents a biophysical method for measuring

protein concentration. Magnetic nanoparticles were of iron oxide prepared by chemical coprecipitation and then employed as active agents for the measurement. The concentration was determined via the amount of MNPs in attachment with proteins. In order to show the confirmation of the correctness of the technique, some initially obtained results were presented which showed that proteins attached to MNPs have no contribution to magnetism of total induced magnetic field of MNPs. The detecting and measuring technique expressed the advantage of a timely fashion and independent behavior of protein on the measuring signal. The results can be improved in the continuation of research where the optimization is being considered. Due to several advantages, the method of employing the MNPs as active agents for detecting and measuring biological concentration is becoming one of promising techniques for an application trend in biomedicine.

ACKNOWLEDGMENT This research was supported by Ministry of Science and

Technology - Vietnam for the Potential Project implementing in 2012. The authors are grateful to Professor Derac Son at Hannam University for suggestions on experimental idea and valuable discussions.

MNPMNP-SiO2

MNP-protein

MNP-SiO2-Protein4.6

4.7

4.8

4.9

5.0

5.1

5.2

5.3

Mag

netic

fiel

d (A

/m)

Sample

Measured data

Figure 7. Four samples with the same initial MNP amount: MNPs,

MNPs with SiO2 coating, MNPs attached with proteins, and MNPs with SiO2 coating attached with proteins.

Figure 6: The persistence of superparamagnetic properties of MNPs core in alginate coating

41

REFERENCES [1] Chan, S.M. et al. Protein microarrays for multiplex analysis of signal

transduction pathways. Nat. Med. 10, 1390–1396 (2004). [2] Shingyoji, M. et al. Quantum dots–based reverse phase protein

microarray. Talanta 67, 472–478 (2005). [3] Zheng, G. et al. Multiplexed electrical detection of cancer markers with

nanowire sensor arrays. Nat. Biotechnol. 23, 1294–1301 (2005). [4] Ji, H. et al. Microcantilever biosensors based on conformational change

of proteins. Analyst 133, 434–443 (2008). [5] Ghosh, S., Sood, A.K. & Kumar, N. Carbon nanotube flow sensors.

Science 299, 1042–1044 (2003). [6] Drummond, T.G., Hill, M.G. & Barton, J.K. Electrochemical DNA

sensors. Nat. Biotechnol. 21 1192–1199. [7] L. A. Bradshaw, A. Myers, W. O. Richards, W. Drake, J.P. Wikswo,

Vector projection of biomagnetic Fields, Medical & Biological Engineering & Computing, 43 (2005) 85.

[8] C.R. Tamanaha∗, S.P. Mulvaney, J.C. Rife, L.J.Whitman, Magnetic labeling, detection, and system integration, Biosensors and Bioelectronics 24 (2008) 1–13.

[9] J. Llandro, J. J. Palfreyman, A. Ionescu, C. H. W. Barnes, Magnetic biosensor technologies for medical applications: a review, Med Biol Eng Comput (2010) 48:977–998, DOI 10.1007/s11517-010-0649-3

[10] Berry CC (2009) Progress in functionalization of magnetic nanoparticles for applications in biomedicine. J Phys D 42: 224,003.

[11] Roca AG, Costo R, Rebolledo AF, Veintemillas-Verdaguer S, Tartaj P, Gonzalez-Carreno T, Morales MP, Serna CJ (2009) Progress in the preparation of magnetic nanoparticles for applications in biomedicine. J Phys D 42:224,002.

[12] Gaster, R. S. et al. (2009) “Matrix-insensitive protein assays push the limits of biosensors in medicine”, Nature Medicine, doi:10.1038/nm2032.

[13] KarstenSternickel, and Alex I Braginski, Biomagnetism using SQUIDs: status and perspectives, Supercond. Sci. Technol. 19 (2006) S160–S171.

[14] Ilia Fishbein and Robert J. Levy, The matrix neutralized, Nature - Analytical Chemistry, 461 (2009) 890.

[15] Arben Merkoci, Nanoparticles-based strategies for DNA, protein and cell sensors, Biosensors and Bioelectronics 26 (2010) 1164–1177.

[16] Patterson, A. L,, The Scherrer Formula for X-Ray Particle Size Determination, Physical Review, 56 (1939) 978-982.

[17] Jin Zhang et al. (2006), “Design nanoparticles as Drug carriers for cancer therapy”, Cancer genomics and proteomics review, 3, 147-158.

[18] Hui Yao, Nan Li, Shu Xu, Jin-Zhong Xu, Jun-Jie Zhu∗ , Hong-Yuan Chen (2005), “Electrochemical study of a new methylene blue/siliconoxide nanocomposition mediator and its application for stable biosensor of hydrogen peroxide”, Biosensors and Bioelectronics, 21, 372-377.

42

A Novel Approach to Protect Intellectual Property Core of FPGA-Based Partially Reconfigurable

Systems Tran Thanh1, Tran Hoang Vu1, Pham Ngoc Nam1, Nguyen Van Cuong2

1School of Electronics and Telecommunications, Ha Noi University of Science and Technology, Ha Noi, Viet Nam. Email:nb10069, nb10002, pnnam-fet@mail.hut.edu.vn

2 Faculty of Electronics and Telecommunications, Da Nang University of Technology, Da Nang, Viet Nam. Email: nvcuong2000@gmail.com

Abstract Reuse of designed hardware modules, also called Intellectual Property (IP) cores is an important part of the effort to design and implement complex systems. An IP core design can take months to develop, and be stolen in seconds. Therefore, IP protection of hardware designs is the most important requirement for many Field Programmable Gate Array (FPGA) IP vendors. There have been already many proposals to overcome this problem using symmetric encryption techniques with a cryptographic key to be stored in a non-volatile memory located on FPGA or in a battery-backed RAM as done in some of the current FPGAs. This paper proposes a novel approach for protecting IP cores on embedded systems based on FPGA by building an AES decryptor in the partially reconfiguration region. The AES decryptor can be released from the partial reconfigurable region when it is not in use and then be replaced with another useful application. Experimental results and analysis show that the proposed technique reduce resource overhead, increase the flexibility of the system and it is robust against attacks.

Keyword- reconfigurable, bitstream security, embedded system, partial reconfiguration.

I. INTRODUCTION The advance in semiconductor processing technology has

led to a rapid increase in the design of complex integrated circuit. Design methodology based on the reuse has prevailed in the IC design field. Reuse of Intellectual property (IP) can reduce product costs, shorten design cycles and decrease design risk. Nowadays, reusable IP cores have become the basic units of system design. IP cores are widely exchanged and are being sold as an independent design. Most of reusable IP cores can require a lot of time and effort to be implemented and verified by IP providers, but these IP cores can be easily copied or modified. Therefore, IP providers apply security solutions to against the stealing and tampering their products. In addition, users expect that purchased IP core is correct and legal. How to protect IP core reuse can effectively has become a serious problem and therefore an increasing number of attention has been paid.

A. Security risks Design outsourcing of FPGA-based systems to third-party

vendors and open source communities gives rise to network of trust relationships between parties and toolsets. This multi-party arrangement leads to increasing security risks [1].

Individual IP cores are typically verified, and then certified as secure. However, controlling the consistency of a complete design flow, from the system specification through IP core integration and deployment of the final product, becomes more challenging when a PR flow is considered. PR introduces a risk of replacing part of the design with malicious (Trojan-horse) functionality and thus a risk of compromising the entire system [2]. During self-reconfiguration, illegal connections (called covert channels) can potentially be set up, allowing for unwanted ad-hoc interactions between IP cores. This introduces the risk to IP ‘secrets’, e.g. encryption key extraction, algorithm theft etc. IP protection methods proposed to date can be classified into two main groups: low-cost security and high-end security.

Low cost methods target mainly massive scale (industrial) IP theft. This approach may not prevent the efforts of motivated peers to obtain FPGA configuration details and to publish results [3]. “Security-by-obscurity” is considered an adequate hindrance for resource-limited and time-limited attackers.

High-end protection includes methods aimed at providing security of the design IP against all but sophisticated attackers who have unlimited resources. IP protection provided by FPGA design software [4, 5] supports IP core licensing and Digital Rights Management (DRM), down to the design netlist-level. The software-level protection effort is augmented by configuration bitstream encryption, being a de-facto industry standard provided by FPGA-specific tools [6]. However, this approach is limited and does not support PR. When PR is used the attacker could, under certain conditions, intercept the plain-text content of the IP core by accessing it, using the ICAP for example.

B. AES algorithm The Rijndael algorithm was adopted as the Advanced

Encryption Standard (AES) in 2001. The AES algorithm is a symmetric block cipher that encrypts and decrypts electronic data. Symmetric-key means the key used for encryption is the same as the one used for decryption. Block cipher means the data is processed in blocks. Symmetric-key block-cipher encryption algorithms are widely used in many industries to provide data security because of its high security protection

43

and efficiency, ease of implementation, and fast data-processing speed.

AES was announced by the U.S. National Institute of Standards and Technology (NIST) [7]. AES has a fixed block size of 128 bits and a key size of 128, 192, or 256 bits. So far, the AES encryption algorithm is still considered safe, and in practice it means that data encrypted with the AES algorithm has not been broken. The key length of AES algorithm is enough to protect information classified as secret. For the more safety information the key length of 192 bits or 256 bits will be used.

The choice of key storage is the second most important design consideration. A key is stored in either volatile or non-volatile storage, depending on the chip vendor. Once power is off for volatile storage, the key is lost unless there is an external battery connected to the chip as a backup power supply. On the other hand, non-volatile key storage hands the designer greater flexibility.

C. Partial Reconfiguration Partial Reconfiguration (PR) technology, offered by some

FPGA vendors, provides full access to the FPGA configuration memory during system runtime, using for example an Internal Configuration Access Port (ICAP). This enables modification of the underlying hardware configuration during runtime, e.g. insertion of IP cores, without restarting the system.

Figure 1. The model of partial reconfigurable system

Partial Reconfiguration is the ability to dynamically modify blocks of logic by downloading partial bit files (IP cores) while the remaining logic continues to operate without interruption. In this system, Figure 1:

Reconfigurable Partition (RP) is design hierarchy instance marked by the user for reconfiguration.

Reconfigurable Module (RM) is a portion of the logical design that occupies the RP and each RP will have multiple RMs.

Static logic is all logic in the design that is not reconfigurable.

The same framework and embedded platform in [8], in this paper, we presented a new method by building a AES decoder

into partial reconfigurable region. This method is really useful for low-cost FPGAs.

The structure of the paper is as follows: Section II reviews the related work. Section III describes the proposed design flow to IP core protection. Section IV details the system implementation and evaluation. Section V concludes the paper.

II. RELATED WORK This section reviews current research in security of the

partial reconfigurable system and focuses on two main factors: system integrity and design privacy protection. System integrity protection is mainly used in systems at which high security levels are required. Hadzic [9] describes the threat of hardware viruses in FPGAs.

Kean [10] and Bossuet [11] highlighted vulnerability of volatile FPGAs to IP piracy and reverse engineering, and proposed bitstream encryption as a countermeasure. Drimer [12] more recently examined a wide range of attack mechanisms and countermeasures. A more general review of security challenges facing embedded systems can be found in [13]. Adi [14] proposed a system based on the use of public and secret-key cryptography. In [15], Yuan summarized current IP protection goals and proposed various solutions.

Most of the proposed approaches rely on the assumption that a secret can be stored safely in the FPGA fabric. The secret is usually an encryption key, and its storage is permanent, one time programmable, or volatile.

Ideally, IP-blocks should be tested, and verified by an external Trusted Authority. In reality, IP cores are typically obtained through in-house design reuse, from 3rd party IP vendors, or using automatic IP core generation tools. Thompson [16] challenged us to question the security of tools.

III. PROPOSED DESIGN FLOW We aim to implement the static system and the

reconfigurable modules in completely isolated design steps. Then, all modules can be developed without the existence of the final static system. Moreover, modules are encapsulated and might be ported among different designs without any additional synthesis or place&route step. This is possible by statically partitioning the routing resources of the FPGA into the resources used for the top level communication and into resources that are used for the implementation of the static system or the reconfigurable modules.

A. Implementing the Static System

For defining a reconfigurable partition within the static system and for constraining the interface wires that are allowed to cross the border for connecting a partial module, the user has to floorplan the system. This is supported by a comfortable GUI. For a selected reconfigurable region, our tools will generate constraints for the physical implementation of the static system that ensure that no logic and routing resources will be used within the selected reconfigurable partition.

44

The Xilinx vendor tools provide such prohibit constraints only for logic resources. We solved the remaining problem of constraining routing resources by generating blockers that occupy within a definable region all wire resources that will then not be used by the Xilinx router for implementing the static system. For each interface signal, we leave a hole in the blocker such that the corresponding dummy sink or source can be connected.

As shown in Figure 2, the static system includes an embedded processor, a reconfigurable controller with an integrated ICAP core, UART interface, Compaq Flash card connection, etc.

Figure 2. Block diagram of a FPGA system containing reconfigurable AES IP core controlling access to FPGA reconfiguration region via the ICAP port.

B. Implementing the Reconfigurable Modules The implementation of the reconfigurable modules follows

the same idea of constraining than the one that has been applied to the static system. But this time, we constrain a particular module into an encapsulated region having the size and containing the same resource layout than the reserved reconfigurable partition of the static system.

Figure 2 illustrates the partial configurable region, we have built the configurable partitions which consist of a partition to load the AES core to perform the decoding when IP core load to the reconfigurable partition. When updating is not required, AES core can be released and another application can be replaced here.

D. FPGA system multi-party environment To implement the proposed model, we consider a multi-

party environment of IP core design and protection. The parties and the implementation process are shown in Figure 3.

System Integrator (SysInt) designs the FPGA system and provides it to the user. SysInt has physical access to the product and can issue a product upgrade in the field. A typical system consists of custom elements and multiple third-party (IPVend) IP cores. IP cores can be distributed in various formats: HDL sources, netlists or FPGA device-specific partial bitstreams, depending on the level of trust between SysInt and IPVend.

IP Vendor (IPVend) provides reusable components (IP cores), and related data sheet, or may design an IP block to meet a provided subsystem specification. IPVend is typically not directly involved in the system level FPGA design process. IPVend wishes to protect its own design secrets.

Trusted Authority (TAut) is an authorization and/or certification center. TAut confirms the key generation process during initial system start-up and certifies the resulting key material. TAut is assumed to be trustworthy by all parties and is not involved in the system development process.

As System Integrators (SysInt) need IP core for his system, he sends requirements for an IP core to IP Vendor (IPVend). The IPVend sends a request of encryption key to encrypt his IP before sending it to SysInt. After receiving requests from IPVend, SysInt will send back an encryption key of your system, IPVend will send this key to Taut to verify whether this key is the the key of SystInt system or not. If it is true, Taut will send back a confirmation and IPVend will use this key to encrypt your IP core and sent to SysInt. Finally, SysInt integrated IP core into his system and saved in the CF card.

Figure 3. Flowchart for IP core installation

IV. IMPLEMENTATION AND EVALUATION

A. Prototype system To test the proposed method, we have built a prototype

system(Figure 4) which consists of a reconfigurable embedded platform based on Xilinx Virtex-6 XC6VLX240T-1FFG1156 FPGA ML605 board, a laptop and a CF card (Figure 4). The ML605 board and the laptop are connected via a TCP/IP connection.

Figure 4. Prototype system

45

On the Virtex-6 XC6VLX240T chip, we embedded a MicroBlaze soft-core microprocessor using Xilinx Embedded Development Kit (EDK) ver. 14.1 software. EDK toolset allows designers to easily create platforms based on either MicroBlaze. EDK offers a variety of peripherals (UARTs, counter, Ethernet, memory controller, general-purpose I/O and so on) and a one-connection solution based on the structure of the IBM CoreConnect bus [17].

In this prototype system, we build AES-256 Decryptor into a partial bitstream file saved in the CF card.

B. Resource overhead and performance of system Table 1 shows the resources used for the modules of

hardware system based on Virtex-6 ML605 Xilinx board.

TABLE I. HARDWARE UTILIZATION OF AES DECRYPTOR CORE.

AES-256 Decryptor LUTs Registers BRAM

% of Virtex-6 (LX240T) 4 0.5 5

The proposed test program measures the execution time

consumed by the consecutive decryption of data blocks until 105KB of the partial bitstream file are processed. The data to cipher is stored in Block RAM and configured on the fly.

As shown in Table 2, with the measured throughput from given partial bitstream file, it takes about few milliseconds to update the partial bitstreams for Virtex-6 LXT FPGA. This result is acceptable for a system update.

TABLE II. PERFORMANCE OUR SECURITY SYSTEMS..

Virtex-6 (LX240T) Time Throughput

AES-256 Decryption 2.5ms 350Mbps

We also consider the resource overhead when

implementing with the Xilinx Spartan-6 LX45 and Virtex-5 LX50T FPGA. Table 3 shows that the resource overhead is significant and the release of AES core when not being in use provides a substantial practical benefits.

TABLE III. HARDWARE UTILIZATION OF AES CORE OF SPARTAN-6 AND VIRTEX-5.

AES-256 Decryptor LUTs Registers BRAM

% of Virtex-5 (LX50T) 17.5 15 25

% of Spartan-6 (LX45) 8.2 4 7

V. CONCLUSIONS In this paper, with the proposed method, all the partial

bitstream files to ensure always be encrypted when stored outside. The flexibility of the system is that the AES core is implemented in the partial reconfigurable region. When the partial modules installed or updated completely, the reconfigurable partition of AES core may be released for another application to execute.

The proposed method for PR systems offers improved security and convenient reuse of external third-party IP cores, protect the IP Digital Rights Management as well as system security and integrity in embedded design environment with multi-party participants.

Future work will optimize the performance of the reconfiguration flow via Ethernet interface to remote updating.

REFERENCES [1] K. Thompson, “Reflections on trusting trust”, Communications of the

ACM, pp. 761-763. USA, 1984. [2] S. T. King, J. Tucek, A. Cozzie, C. Grier, W. Jiang, and Y. Zhou,

“Designing and implementing malicious hardware”, in Proc. First USENIX Workshop on Large-Scale Exploits and Emergent Threats (LEET), 2008.

[3] J. Note, E. Rannaud, “From the bitstream to the netlist”, In Proc. of the 16th international ACM/SIGDA symposium on FPGA, ACM, 2008, pp. 264-264.

[4] H. Walker, “Xilinx speeds HDL simulation with SecureIP and FAST Simulation Mode Models”, http://www.fpgajournal.com/articles_2008/20080610_xilinx.htm, 2008.

[5] M. Miller, “Synplicity introduces secure IP flow for FPGAs, signs ARM, Tensilica as partners “, http://www.edn.com/index.asp?layout=article&articleid=CA6551580 Xilinx encryption, 2008.

[6] A. Lesea, “IP Security in FPGAs”, unpublished, www.xilinx.com, 2007.

[7] FIPS-46: “Data encryption standard,” NBS, US Department of Commerce, January 1977.

[8] Tran Thanh, P. N. Nam, T. H. Vu, Ng. V. Cuong, “A Framework for Secure Remote Updating of Bitstream on Runtime Reconfigurable Embedded Platforms,” In Proceeding of the fourth International Conference on Communications and Electronics (ICCE 2012), pp. 471-476. Hue, Vietnam, 2012.

[9] Hadzic, I., Udani, S. & Smith, J. M., “FPGA viruses”, FPL '99: Proceedings of the 9th International Workshop on Field-Programmable Logic and Applications, Springer-Ver-lag, 1999, 291-300.

[10] Kean T., “Secure configuration of Field Programmable Gate Arrays”, FPL '01: Proceedings of the 11th International Conference on Field-Programmable Logic and Applications, Springer-Verlag, 2001, 142-151.

[11] Bossuet L., “Dynamically Configurable Security for SRAM FPGA Bitstreams”. In International Journal of Embedded Systems, IJES, Inderscience Publishers, Vol. 2, Nos. 1/2, pp 73-85, 2006

[12] Saar Drimer, “Volatile FPGA design security – a survey,” Journal of Engineering, Computer Laboratory, University of Cambridge, Version 0.96, April 17, 2008.

[13] Ravi, S., Raghunathan, A., Kocher, P. & Hattangady, S., “Security in embedded systems: Design challenges”, Trans. on Embedded Computing Sys., ACM Press, 2004.

[14] Adi, W., Ernst, R., Soudan, B. & Hanoun, A., “VLSI design exchange with intellectual property protection in FPGA environment using both secret and public-key cryptography”, Emerging VLSI Technologies and Architectures, 2006.

[15] Yuan L., Qu G., Ghout L. & Bouridane A., “VLSI Design IP protection: solutions, new challenges, and opportunities”, NASA/ESA Conference on Adaptive Hardware and Systems, IEEE Computer Society, 2006.

[16] Mark McLean, J. M., “FPGA-based single chip crypto-graphic solution”, Military Embedded Systems, 2007.

[17] Xilinx Inc., “MicroBlaze Processor Reference Guide,” UG081 (v9.0), http://www.xilinx.com/support/documentation/sw_manuals/mb_ref_guide.pdf

46

A DAG - SVM BASED METHOD FOR FACE RECOGNITION USING PCA Tran Thi Minh Hanh

Danang University of Technology, Danang, Vietnam

Email: ttmhanh@dut.udn.vn

Abstract— PCA (Principal Component Analysis) is a well-known feature extraction and data representation technique widely used in the areas of pattern recognition, computer vision and signal processing. In this paper, PCA is implemented for face feature extraction. These features are then used to train and test with support vector machine (SVM) classifiers, which is based on decision-tree pair-wise classification called Directed Acyclic Graph (DAG). A comparative study with Principal Component Analysis (PCA) and k Nearest Neighbor techniques is also presented. The result on two well-known databases JAFFE and ORL show that this method could achieve high classification rate. .Keywords-Face recognition, Principle Component Analysis, Support Vector Machine, Eigenfaces, Directed Acyclic Graph.

I. INTRODUCTION Face recognition has emerged as an active research

area in the field of computer vision and pattern recognition. It can be used in a wide range of applications such as identity authentication, access control, and surveillance [8]. Most of the face recognition system has two main processing steps. The first step is feature extraction, which is performed to provide effective information that is useful for distinguishing between faces of different persons and stable with respect to the geometrical and photometrical variations. In the second step, classification, the extracted feature vector of the input face is then matched against those of enrolled faces in the database.

Figure 1. Face recognition system with PCA and DAG-SVM

Many methods have been proposed for feature extraction, such as geometric feature-based [7], template matching based [8], and graph matching method [3]. In geometric feature-based methods, local features such as the eyes, nose, and mouth are first extracted and their locations and local statistics (geometric and/or appearance) are fed into a structural classifier. Facial feature detection and measurement techniques developed have not been reliable enough to adapt the need. In contrast, template matching based method generally is operated directly on an image-based representation (i.e. pixel intensity array). Because the

detection and measurement of face features are not required, this class of methods has been more practical and reliable as compare to geometric feature-based methods.

A well-know face feature extraction using template matching is that based on the eigenface representation Matthew Turk and Alex Pentland [5]. Nevertheless, PCA could not capture even the simplest invariance unless this information is explicitly provided in the training data. To deal with this problem, some researchers proposed other approaches: elastic bunch graph matching Wiskott et al. [3]; Bartlett et al. [6] proposed using independent component analysis (ICA) for face representation and reported that it performed better than PCA; Kernel PCA is suggested by Ming-HsuanYang [9] for face feature extraction and recognition. However, the performance costs of ICA and Kernel PCA are higher than PCA. In this paper, PCA is used as the feature extraction.

There have been a few of classification methods such as k-Nearest Neighbors (k-NN), Fuzzy C-Means (FCM) and Artificial Neural Network (ANN). Recently, most of researchers have focused on the classification method, Support Vector Machine (SVM). In 1995, Vapnik and Cortes [1] presented the foundations for SVM. Since then, it has become the prominent method to solve problems in pattern classification and regression. The basic idea behind SVM is, with a set of points belonging to two classes, finding the optimal hyperplane that separates the largest possible fraction of points of the same class on the same side, while maximizing the distance from either class to the hyperplane. For linearly non-separable data, SVM maps the input to a higher dimensional feature space where a linear hyperplane can be found. Although there is no warranty that a linear solution will always exist in the higher dimensional space, it is able to find effective solutions in practice.

To deal with the face classification, many researchers [2, 4] have applied SVM in their studies and stated that the experiment results are very positive. In this paper, an algorithm using SVMs for face recognition is presented (Fig.1). The SVMs based recognition is compared with the most popular eigenface approach, which use the k -Nearest Neighbor classification.

The remaining sections of the paper are organized as follows. Section 2 and 3 gives details of the first module (i.e, feature extraction) and the second module (i.e, face classification base on SVM), respectively. Section 4 shows implementation and experiments. Finally, the conclusion is mentioned in Section 5.

47

II. FACE FEATURE EXTRACTION

A. Normalization The first module is to normalize the input image. The goal

of the normalization module is to transform the facial image into a standard format that removes or attenuates variations that can affect recognition performance.

Figure 2. Input image (ORL database) before (a) and after (b)

normalization; global vector (c)

B. Global vector The face images then are represented as global vectors

xglobal = (x1, x2,…, xm*n)T , where m*n and xi is image

resolution and gray value of each pixel, respectively.

C. Feature extraction One of well-known methods for extracting facial feature is

PCA as mentioned above. It was first applied in face classification by Sirovich and Kirby and then Matthew Turk and Alex Pentland [5]. Now it has become the standard method in this field. The input is a dataset D = Nix i ,...,2,1;)( Step 1: Compute the mean of data

N

i

ixx1

)(

(1)

Step 2: Compute the covariance matrix of data

N

i

iTi xxxxN

C1

)()( )()(1

(2)

Step 3: Compute the eigenvectors vi and eigenvalues λi of matrix C Step 4: Order the eigenvectors according to their corresponding eigenvalues from high to low. Keep only K eigenvectors associated with K largest eigenvalues

Finally, each of the centered training images (i.e, xx i )( ) is projected into the eigenspace ,...,, 21 kvvvV . To project an image into the eigenspace, calculate the dot product of the image with each of the ordered eigenvectors.

)( )(1' xxVx i (3)

The new vector ,...,, '2'

1''

kxxxx of the projected image will contain as many values as eigenvectors.

III. CLASSIFICATION USING SUPPORT VECTOR MACHINE In this section, the basic theory of the SVM is described

first, and then present the multi-class strategy for SVMs to solve the multiclass recognition.

A. Two-class classification The goal of SVM classifiers is to find a hyperplane that

separates the largest fraction of a labeled data set. The most important requirement, which the classifiers must have, is that they have to maximize the margin (the distance between the hyperplane and the nearest data point of each class) (shown in Fig. 3). This linear classifier is termed the optimal separating hyperplane (OSH).

Figure3. A Typical SVM classifier

Consider the problem of separating the set of training vectors belonging to two separate classes, x i ,y i , i=1,…,l, where x i nR , iy 1,1 . Now, all hyperplanes in nR are parameterized by a vector w, and a constant b, expressed in the equation w.x + b = 0. The set of vectors is said to be optimally separated by the hyperplane if it is separated without error and the margin is maximal. The separating hyperplane in canonical form [10] must satisfy the following constraints,

y i [(w.x i )+b] i1 , i=1,…,l ( 4)

where i are nonnegative slack variables which measure the degree of misclassification of the data.

Hence the hyperplane that optimally separates the data (with a large margin and a small error penalty) is the one that

minimizes 21 ||w|| 2 + C

l

ii

1

(where C is a given value) under

the constrains of equation (4). The solution to the optimization problem is given by using Lagrange functional:

L(w,b, ) = 21 ||w|| 2 + C

l

ii

1

-

l

i

ii y

1

[ w.x i ) + b]-1+ i

(5 ) Where i are the Lagrange multipliers. The problem in Eq.5 is transformed to its dual problem ),,(minmax

,,

bwL

bw

Solving the dual problem,

l

i

jijl

j

iji

l

ii xxyy

1 11 21minarg

(6 )

With constraints, Ci 0 , i=1,…,l

48

l

i

ii y

1

0

The key advantage of a linear penalty function is that the slack variables vanish from the dual problem, with the constant C appearing only as an additional constraint on the Lagrange multipliers. Solving Eq. (6) with constrains determines the Lagrange multipliers, and the OSH is given by,

l

i

iii xyw

1

and jj xwyb . (7)

Where jx is one of the support vectors (the data that satisfy the equality in (4)) , and jy =1 or -1. The obtained hyperplane is called soft margin hyperplane.

For the new data point x, the classification is then f(x) = sign(w. x + b) (8)

If f(x) = +1 then x belongs to positive class, if f(x)= -1 then x belongs to negative class

B. Mapping the Inputs to other dimensions using Kernel In most of real applications, the data could not be linearly

classified. To deal with this problem, data is transformed into a higher dimensional feature space and assume that data in this space can be linearly classified [1]. In fact, determining the optimal hyperplane is a constrained optimization problem and can be solved using quadratic programming (QP) techniques. Let the mapping function be )(x , to set up new optimization problem, replacing all occurrences of x with )(x . The minimization problem (recall eq. 4) would still be

l

i

jil

j

jiji

l

ii xxyy

1 11))().((

21minarg

And w (in Eq.7) would be

l

i

iii xyw

1

)(

Equation 8 would be

l

i

iii bxxysignbxwsignxf

1

)))().((())(.()(

That is, if we knew the formula (called a Kernel) for the dot product in the higher dimensional feature space. The classifier would be:

l

i

iii bxxKysignxf

1

)),(()(

Some common Kernels include Polynomial, Gaussian Radial basic function (RBF) and Hyperbolic tangent. This paper is only implemented with RBF- SVM classifier:

2||||exp(),( jiji xxxxK for 0

C. Multiclass SVM The earliest used implementation for SVM multiclass

classification is probably the one-against-all method. It

constructs k SVM models where k is the number of classes. The ith SVM is trained with all of the examples in the ith class with positive labels, and all other examples with negative labels.

Figure 4. One-against-all (left) and one-against-one (right) multiclass

SVM [] Another major method is called the one-against-one

method, first used of on SVM was in [12]. This method constructs classifiers where each one is trained on data from two classes. There are different methods for doing the future testing after all k(k-1)/2 classifiers are constructed, voting strategy suggested in [11] is one example: if x is in the ith class, then the vote for the ith class is added by one. Otherwise, the jth is increased by one. Then predicting x is in the class with the largest vote. In case that two classes have identical votes, thought it may not be a good strategy. Platt, Cristianini, and J. Shawe-Taylor [13] proposed decision-tree-based pair-wise classification called Directed Acyclic Graph (DAG). Its training phase is the same as the one-against-one method mentioned above by solving binary SVMs. However, in the testing phase, it uses a rooted binary directed acyclic graph which has internal nodes and leaves. Each node is a binary SVM of ith and jth classes.

Figure 5. Testing phase with three classes. Fig. 5 shows the decision tree for three classes. In the

figure, i shows that associated class is not i. As the top-level classification, we can choose any pair of classes. If x does not belong to Class 2, thus it belongs to either Class 1 or 3 and the next classification pair is Classes 1 and 3. In this method,

2

3 2 1

푓13(푥)

1

3

푓12(푥)

푓23(푥)

Class 2 Class 3 Class 1

49

testing time is less than the one-against-one method. DAG method is chosen in this paper for multiclass SVM.

IV. DATABASE AND EXPERIMENTS In this work, the experiment is performed on two face

databases. One is on the Cambridge ORL face database, which contains 40 distinct persons. Each person has ten different images, taken at different times. There are variations in facial expressions such as open/closed eyes, smiling/non-smiling, and facial details such as glasses/no glasses. All the images were taken against a dark homogeneous background with the subjects in an up-right, frontal position, with tolerance for some side movements. There are also some variations in scale. The second is on the JAFFE database of 200 images of 10 individuals, containing six basic facial expressions (happiness, sadness, surprise, anger, disgust and fear) and one neutral. The performance of the SVMs based face recognition is compared with other algorithm to show its success.

(a)

(b)

Figure 6. Some images from (a) ORL [15] and (b) JAFFE [14]

databases

A. Experiments on ORL database In this face recognition experiment on the ORL

database, 200 samples (5 for each individual) are randomly selected as the training set, from which the eigenfaces and the support vector machines (SVMs) are calculated. The remaining 200 samples are used as the test set. In this experiment, the number of features will be changed up to 200. And then we take a look at the result corresponding to the number of features. In general, many feature vectors needs more computing time but gives more accuracy (Fig. 8). When the number of feature vectors is chanced, input vectors of classifier are also changed, since output passed through PCA is directly connected to the input of the classifier. Two classifications are implemented methods to conduct experiments on this database to solve face recognition problem:

1) k Nearest Neighbor (k-NN): uses distance metric L2 for classification; with the value of k set to 1

2) Support Vector Machine (SVM): Applies nonlinear SVM with the Kernel function RBF with the value of C (soft margin parameter) set to 2 , 2 , 2 ,…,2 ,…,2 and the kernel parameter gamma set to 2 , 2 , 2 , …,2 .

The reported results were obtained with Cross-Validation analysis with 2-folds on the dataset.

Figure 7. Comparison of accuracy with the changes of C and gamma of the

SVM algorithm on ORL database.

Figure 7 shows the recognition rate when changing the soft margin parameter C and the Kernel parameter . When 퐶 = 215 and 훾 = 2 11 , the recognition rate reaches a maximum value. After choosing parameters, the classifiers obtained from training are used to carry out recognition on the images in the testing samples. The change of the recognition rate with the increasing number of principle components is shown in figure 8. Initially the recognition rate increases with the increasing number of principle components, and when the principle component is increased to 80, the recognition rate reaches an optimal value of 91.75%. Fig. 8 shows the results of using SVM in comparison with k Nearest Neighbor. It is obvious that the accuracy rate of SVM is higher than that of k-NN (maximum accuracy of 85%).

Figure 8. The comparison of accuracy versus the number of eigenfaces

with NN and SVM algorithms on the ORL database.

50

B. Experiment on JAFFE database In this face recognition experiment on the JAFFE

database, 100 samples (10 for each individual) are randomly selected as the training set, from which the eigenfaces and the support vector machines (SVMs) are calculated. The remaining 100 samples are used as the test set. Two classification methods (k-NN and SVM) are used to conduct experiments on this database. The experiments are conducted using Cross-Validation analysis with 2-folds.

Since JAFFE face database overall has the face images of 10 people, the number of SVM classification functions in the experiment will be 10. In this experiment, the accuracy reach 100% with 14 principle components used for training SVMs classifier (in compare with accuracy of 99.5% using Nearest Neighbor). This is likely due to the number of individuals on this database (10 people versus 40 individuals on ORL database), the number of images are used for training (10 images each person versus 5 images for previous experiment). With the simulation result shown in Fig. 9, it is obvious that SVMs based classification still has higher accuracy rate in this experiment.

Figure 9. Comparison of accuracy versus the number of eigenfaces with NN

and SVM algorithms on the JAFFE database.

V. CONCLUSION In this paper, a multiclass recognition strategy for the use of

conventional SVMs is implemented to solve the face

recognition problem. The face recognition experiments using DAG – RBF SVMs has been developed. The effects of SVMs parameters are also evaluated via a range of different values. The simulation results show that the SVMs can be effectively trained for face recognition. As shown in the comparison with the k-Nearest Neighbor method, the use of SVMs can achieve high accuracy for face recognition.

REFERENCES [1] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning,

vol. 20, no. 3, pp. 273–297, 1995. [2] Guodong Guo, Stan Z. Li, Kap Luk Chan, “Face recognition by support

vector machines”, Automatic Face and Gesture Recognition, 2000, Proceedings. Fourth IEEE International Conference on, pp. 196 -201.

[3] L. Wiskott, J. M. Fellous, N. Kuiger, and C. von der Malsburg, “Face recognition by elastic bunch graph matching,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 19, no. 7, pp. 775–779, 1997.

[4] Len Bui, Dat Tran, Xu Huang, Chetty, G., “Classification of gender and face based on gradient faces” in Visual information Processing (EUVIP), 2011, pp.269-272

[5] M. A. Turk and A. P. Pentland, “Face recognition using eigenfaces,” in Computer Vision and Pattern Recognition, 1991. Proceedings CVPR ’91., IEEE Computer Society Conference on, 1991, pp. 586–591.

[6] M. S. Bartlett, J. R. Movellan, and T. J. Sejnowski, “Face recognition by independent component analysis,” Neural Networks, IEEE Transactions on, vol. 13, no. 6,pp. 1450–1464, 2002.

[7] R. Brunelli, T. Poggio, “Face recognition: features versus templates”, IEEE Transactions on Pattern Analysis and Machine Intelligence 15 , 1993, pp.1042-1052

[8] R. Chellapa, C. L. Wilson, S. Sirohey, “Human and machine recognition of faces: a survey”, Proceedings of IEEE 83, 1995, pp.705-741

[9] Yang Ming-Hsuan, “Kernel eigenfaces vs. kernel fisherfaces: Face recognition using kernel methods,” in Automatic Face and Gesture Recognition, 2002. Proceedings. Fifth IEEE International Conference on, 2002, pp. 215–220.

[10] V.N. Vapnik, Statiscal Learning Theory, Wiley, New York, 1998 [11] J. Friedman. (1996) Another Approach to Polychotomous Classification.

Dept. Statist., Stanford Univ., Stanford, CA. [Online]. Available: http://www-stat.stanford.edu/reports/friedman/poly.ps.Z

[12] U. Kreßel, B. Schölkopf, C. J. C. Burges, and A. J. Smola, “Pairwise classification and support vector machines,” in Advances in Kernel Methods—Support Vector Learning, , Eds. Cambridge, MA: MIT Press, 1999, pp. 255–268..

[13] J. C. Platt, N. Cristianini, and J. Shawe-Taylor, “Large margin DAG’s for multiclass classification,” in Advances in Neural Information Processing Systems. Cambridge, MA: MIT Press, 2000, vol. 12, pp. 547–553.

[14] The JAFFE database. Available: http://www.kasrl.org/jaffe.html [15] The ORL database. Available:

http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html

51

A SPEAKER RECOGNITION SYSTEM USING COMBINATION METHOD

BETWEEN VECTOR QUANTIZATION AND GAUSSIAN MIXTURE MODEL

Ngo Quoc Hung Electronic and Telecommunication Eng. Department

Danang University of Technology Danang, Vietnam

Ngoquochung.dtvt@gmail.com

Pham Van Tuan Electronic and Telecommunication Eng. Department

Danang University of Technology Danang, Vietnam

pvtuan@dut.udn.vn

Abstract— Speaker recognition is a biometric technique to recognize people’s identity based on their voice signal. A recognition system has two main requirements, which are high accuracy recognition rate and short processing time under large amount of training data. Many researchers have proposed various speaker recognition techniques; and the two most popular methods are Vector Quantization (VQ) and Gaussian Mixture Model. Each method has its own advantage. VQ method can perform simply and has fast computation time. However, the main drawback is that its recognition accuracy rate is not high, especially with large data sets. Meanwhile, the GMM-UBM has greater accuracy rate than VQ though, for long processing time, this process does not always produce satisfying result in practice. Accordingly, this paper proposes a method to solve the two above requirements by performing a combination of two advantages of each VQ and GMM model to provide a new model, which can be called a “Hybrid VQ/GMM-UBM model”. This model not only takes the advantage of high accuracy in GMM method but also improves the accuracy rate and reduces the amount of computation of the system when combined with VQ model. The efficiency of the model is evaluated by computational time and accuracy rate compared to GMM models. Experimental results showed that the hybrid VQ/GMM-UBM model had better accuracy.

Keyword - Speaker recognition, Vector Quantization, Gaussian Mixture Model

I. INTRODUCTION

Speaker recognition is a biometric technology derived from areas of speech processing. The speaker recognition field has over 50 years of research and development. The general idea of speaker identification tasks is to assume that the voice of human is unique to each individual, and it can be used as a distinguished characteristic to identify the owner of that voice among other individuals.

Speaker recognition system has two operation phases, training phase and testing phase. In both phases, speech signal is preprocessed to improve the voice quality and reduce noise. It then was extracted characteristics to obtain the set of feature vector. In the training phase, the characteristic vector is used

to train the speaker model. Many methods are used to train speaker model, from the simplest one which is used to build codebook model using vector quantization (VQ) (yet the accuracy of this method is not high) to complex methods such as Gaussian Mixture Model – Universal Background Model. The overall structure of speaker recognition system is depicted in Figure.1

Figure 1. General Speaker Recognition System [3]

II. PRE-PROCESSING AND FEATURE EXTRACTION

A. Amplitude Normalized

Voice data was obtained with the amplitude fluctuation. Even if the speaker says with a standard volume, the amplitude of obtained signal can still be unstable. This easily happens when the speaker slightly turns away or moves the microphone closer to his mouth or pulls it away more than a few centimeters. This fluctuation affects the recognition results.

The normalization is necessary. However, it does not require the signal amplitude to be good, not too small to lose its characteristics. Thus, we can simply implement by multiplying each point with an appropriate coefficient k.

(32767 / 2) 100

max ( )k

s n

52

B. Silience Removal

Speech signal usually contains many silence intervals at various points such as at the beginning points of the signal, between words of the sentence or at the end of the signal. If the signal contains silence intervals without treatment, it will occupy resources of system to process on these signal intervals. The silence intervals, however, do not have any contribution to the identification, even it can interfere to the processing. Hence, the silence intervals must be treated and eliminated before implementing feature extraction. Nowadays, a number of met-hods can effectively solve this problem as voice activity detection – VAD [9][9][9], short time energy or spectral centroid.

C. Feature Extraction

The extraction of the best parametric representation of acoustic signals is an important task to produce a better recognition performance. The efficiency of this phase is important for the next phase since it affects its behavior. MFCC is based on human hearing perceptions which cannot perceive frequencies over 1Khz. In other words, in MFCC is based on known variation of the human ear’s critical bandwidth with frequency is employed. MFCC has two types of filter which are spaced linearly at low frequency below 1000 Hz. A subjective pitch is present on Mel Frequency Scale to capture important characteristic of phonetic in speech. Logarithmic spacing is above 1000Hz.

The overall process of the MFCC is shown in Figure 2.

melcepstrum

melspectrum

framecontinuousspeech

FrameBlocking

Windowing FFT spectrum

Mel-frequencyWrapping

Cepstrum

Figure 2. Computing of mel-cepstrum

III. MODEL TRAINING

A. Vector Quantization

Figure 3. Vector Quantization based Codebook of two

speakers

Vector quantization (VQ) is a process of mapping vectors from a vector space to a finite number of regions in that space. These regions are called clusters and are represented by their centroids. A set of centroids, which represents the whole vector space, is called a codebook. In speaker identification, VQ is applied on the set of feature vectors extracted from the speech sample and as a result, the speaker codebook is generated. Such codebook has a significantly smaller size than extracted vector set and is referred as a speaker model. This codebook is generated by many algorithms such K-mean, LBG…

During the matching, a matching score is computed between extracted feature vectors and every speaker codebook enrolled in the system. In this paper, matching score is a Euclean distance [1] between feature vectors and codebook of speaker as formula:

2

1

1, min

N

j i ii

D X C x cN

where X is a set of N extracted feature vectors, C is a speaker codebook, xi are feature vectors, ci are codebook centroids.

B. Gaussian Mixture Model – Universal Background Model Gaussian Mixture Model is a type of statistical model

which was first introduced by Reynolds [7]. In this approach, UBM which is a large GMM trained to represent the speaker independent distribution of features is used. UBM can be gender independent/dependent model and use EM algorithm to training [5][6]. After UBM was trained, speaker dependent models are derived from the UBM by maximum a posteriori (MAP) adaptation [7]. To form a speaker dependent model, first, the log-likelihood of each gender dependent model given the input data is calculated. The gender is determined by selecting the gender-model with the higher score. The corresponding gender dependent UBM is used to adapt a speaker dependent model (Figure 4) [7]. Regarding speaker adaptation three EM-steps and a weighting factor of 0.6 for the adapted model and correspondingly 0.4 for the UBM are used to merge these models to final speaker dependent model [8].

Figure 4. Adaptation of speaker model from UBM [7]

53

C. The combination of VQ and GMM-UBM (VQ/GMM-UBM)

As mentioned before, VQ based solution is less accurate than that of the GMM. In this paper, a method took the superiority of VQ, which is simplicity computation to distinguish between male and female speaker. After, we use of GMM merits to identify the speaker identity in the smaller subgroup.

In this approach, a testing processing was built on three stages. In the first stage, feature vectors of testing speaker were compared with male codebook and female codebook using Euclean Distance to decide gender of testing speaker. Male codebook was trained from a large data of male speakers to represent the male speaker; the same procedure for female codebook. In the second stage, after knowing gender of testing speaker, feature vectors of testing speaker were compared to each VQ model of trained speaker in same gender group to define ten trained speaker which had the highest matching scores. In the third stage, ten trained speakers were computed the log-likelihood with feature vectors of testing speaker using GMM speaker to define a final speaker model who had the highest matching score. After, a threshold was applied to decide “accept” or “reject”. Figure 5 represents speaker identification processing with combination of VQ/GMM-UBM.

Fig 5. Speaker identification processing with

combination of VQ/GMM-UBM.

Since the idea is using both models of VQ and GMM-UBM, in training phase, two speaker model groups were built for male speaker and female speaker as figure 6. Each group will contain VQ model and GMM-UBM model for each of training speaker.

Fig 6. Building of two speaker model groups

IV. EXPERIMENTAL SETUP AND RESULT Speaker database was collected from 150 speakers (70

males, 80 females) whose voices were recorded under the low-noise environment conditions. The audio files were recorded from Adobe Audition program, using PCM, sampling frequency was16000Hz, 16bit. The recording was done because of two purposes: preparing database for the training and identification processes.

- For the training process: in this research 100 people were recorded (50 males and 50 females), each one will be 45 seconds.

- For the identifying process, testing database was taken from 150 people, including 100 people recorded in the training process who were identified to be the interested ones, the other 50.

In this paper, data was characteristically extracted with 39 characteristics per frame. For VQ, used size of codebook was 128. For GMM, model had 25 gaussian mixtures.

The time result of identifying process uses VQ, GMM-UBM and combination of VQ and GMM is shown in Figure.7. VQ based system had shortest calculating time when compared with other models. This was the main advantage of model using VQ, but its accuracy is so low (Figure. 8). Although GMM model processed the computation for a long time, merit of GMM model had higher accuracy. Therefore, with the idea of combining two merits of the two previous models, time processing of VQ-GMM model was shorten (a reduction identification time up to 26% is reached) but system performance is still improved (Figure 7).

54

Fig 7. Identifying time for each speaker with each testing database

Fig 8. DET curve of different modeling techniques

V. CONCLUSION

In this paper, the combination of two techniques has been executed. From obtained results, we observe that the combination approach between VQ and GMM is the good approach due to their different ways of classifying the data. With this combination, data was classified better in order to improve the calculating time as well as improve the system performance.

REFERENCES [1] Homayoon Beigi (2011), “Fundamental of speaker

recognition”, Spinger, New York Dordrecht Heidelberg London.

[2] Piyush Lotia, M.R. Khan (2011), “Multistage VQ Based GMM For Text Independent Speaker identification

System”, International Journal of Soft Computing and Engineering (IJSCE), Vol. 1 (No. 2), pp 21-26.

[3] Joseph Campbell (1997), “Speaker Recognition: A Tutorial”, Proceedings of IEEE, Vol. 85 (No. 9), pp 1437-1462.

[4] Rafik Djemili, Mouldi Bedda, and Hocine Bourouba (2007), “A Hybrid GMM/SVM System for Text Independent Speaker Identification”, World Academy of Science, pp 448-454.

[5] Richard O. Duda, Peter E. Hart, David G. Stork (2001), “Pattern Classification”, Willey Interscien, 2nd.

[6] Douglas A. Reynolds, Richar C. Rose (1995), “Robust Text-Independent Speaker Identification Using Gaussian Mixture Speaker Model”, IEEE Transaction on speech and audio processing, Vol. 3 (No 1), pp 72-83.

[7] Douglas A. Reynolds, Thomas F. Quatieri, Robert B. Dunn (2000), “Speaker Verification Using Adapted Gaussian Mixture Models”, Digital Signal Processing,Vol.10(1-3), pp.19-41.

[8] Tuan V. Pham, Michael Neffe, Gernot Kubin, Horst Hering (2007), “Speaker Segmentation for Air Traffic Control”,

Speaker Classification II, LNAI 4441, pp. 177-191. [9] Tuan V. Pham, Michael Neffe, Gernot Kubin (2007),

“Robust Voice Activity Detection For Narrow-Bandwidth Speaker Verification Under Adverse Environments”, Interspeech, ISSN: 1990-9772.

[10] Tuan V. Pham (2008), Wavelet Analysis for Robust Speech Processing and Applications, VDM Verlag Dr. Muller Aktiengesellschaft & Co. KG, Dudweiler Landstr. 125 a.

[11] Theodoros Giannakopoulos, “A method for silence removal and segmentation of speech signals, implemented in Matlab” Department of Informatics and Telecommunications, University of Athens, Greece

[12] Vaishali Kulkarni, H. B. Kekre (2010), “Speaker Identification by using Vector Quantization”, International Journal of Engineering Science and Technology, Vol. 2, pp 1325-1331.

[13] Archana Shende, Subhash Mishra , Shiv Kumar (2011), “Comparison of different parameters used in GMM based automatic speaker recognition”, International Journal of Soft Computing and Engineering (IJSCE), Vol. 1 (No. 3), pp 14-18.

[14] F. K. Soong, A. E. Rosenberg, L. R. Rabiner and B. H. Juang (1987), “A Vector Quantization Approach to the Speaker Recognition”, AT&T Technical Journal, Vol. 66, pp. 14-26.

[15] M.G.Sumithra, A.K.Devika (2012), “Performance Analysis of Speaker Identification System Using GMM with VQ”, International Journal of Computer Network and Security (IJCNS), Vol 4. No 1., pp 14-19.

55

DESIGN OF TELEVISION REMOTE CONTROL BY VIETNAMESE SPEECH

Nguyen Tu Ha Danang University of Technology

Da Nang, Viet Nam haspkt2000@yahoo.com

Pham Van Tuan Electronics and Telecommunication Department

Danang University of Technologyy Da Nang, Viet Nam

phamvt1976@yahoo.co.uk

Abstract- The applications of speech recognition for remote control devices have received increasing attentions from many scientists as well as manufacturers all over the world nowadays. Catching up with this trend, this paper presents a study of designing a Vietnamese speech recognition system for television remote control. In this research, the recognition of Vietnamese speech is designed based on the combination of Vector Quantization (VQ) method and Hidden Markov Models (HMMs).

Keywords- SPEECH RECOGNITION; MFCC; VECTOR QUANTIZATION; HMM; IR REMOTE CONTROL

I. INTRODUCTION Speech recognition is a process which is utilized for recognizing speech uttered by a speaker. This technology has been studied for more than five decades since 1950s. Voice communication is the most effective mode of communication used by humans. Speech recognition is an important and emerging technology with great potential. The significance of speech recognition lies on its simplicity. This simplicity, together with the ease of operating the speech recognition devices, has created a great deal of advantages in creating and using this technology. It has been applied in various fields, such as security devices, household appliances, cellular phones, ATM machines, and computers. The speech recognition technology has been recently applied in designing voice remote control for television (TV). The two popular TV producers, Samsung and LG, have presented various smartTVs which are voice controlled. However, hitherto, there is no TV production that is controlled by Vietnamese voice in the market yet. In this paper, the design of a TV remote control by Vietnamese speech recognition is presented. The speech recognition process is based on extracting the speech features by MFCC method, and recognizing by applying the combination of Vector Quantization (VQ) method and Hidden Markov Models (HMMs). A TV controlling module is designed using control signal emission principle of TV remote. This module communicates with computer by a RS232 interface.

Figure 1: Overview about remote control TV by speech

II. THEORETICAL BASIS 2.1. TV remote control signal Infra-red is frequently used in communication and control, since it is easily generated and does not suffer from electromagnetic interference. The limitation of this signal, however, is that some other light emissions could also contain infrared, and these can interfere with the communication and control of devices. In order to allow a good communication using infra-red, and avoid those "fake" signals, it is imperative to use a "key" which can detect and inform the receiver what is the real data transmitted and what is fake. Remote controls utilize the 36kHz (or the approximation) to transmit information. Infra-red light emitted by IR Diodes pulsates at 36 thousand times per second, while transmitting logic is at level 1 and silence is 0 [9]. Characteristics of remote control signal are shown in Figure 2 below:

Figure 2: Remote control signal characteristics

2.2. Speech recognition system The speech recognition based on the combination of VQ method and HMMs is illustrated in Figure 3. There is a vocabulary of V words to be recognized, and each word is modeled by a distinct HMM. The training sets consist of K utterances of each word, pronounced by one or more

56

speakers. In order to obtain a word recognizer, the following steps are performed.

Figure 3: Schematic diagram of the VQ/HMM

isolated word recognition system Preprocessing After being received, voice signal is preprocessed before its features being extracted. The purpose of this pre-processing voice signal is to eliminate interference, standardize amplitude, clarify signal, determine the controlling orders, and detect the endpoint. Feature extraction Feature extraction is a process of doing analysis aimed at defining important, typical and stable data of voice signal. This process helps minimize a large set of data in training and recognition and create a significant decreasing in the number of estimations in the system. It also clarifies the differences between this voice and another, at the same time, reduce the differences between two different pronunciations of a word. In this study, we chose MFCC approach for extraction of speech feature. VQ Codebook In discrete HMM system, the continuous feature space is subdivided by a vector quantizer into M non-overlapping subsets and each subset is represented with a codeword m (1≤m≤M). The set of available codewords is called the codebook. The VQ codebook is constructed by an unsupervised cluster algorithm, the Linde-Buzo-Gray (LBG) algorithm [4] [6]. Re-Estimation of HMM For each word of vocabulary, a HMM was built. We estimated the model parameters that optimize the likelihood for the training set of observation sequences. There are many criteria that can be used for this problem. In this study, the Baum-Welch algorithm [5] [7] developed by Baum which is one of the most successful optimization methods was used. Recognition For each unknown word to be recognized, the model likelihood for all possible models was calculated, and the model with the highest likelihood was then selected. The probability calculation was performed using the Viterbi algorithm [7] [8], more precisely the logarithm of maximum likelihood.

III. DESIGN OF TV REMOTE CONTROL MODULE

Based on reception and emission of infra-red signal principle in the TV control remote, a module is designed in order to be able to “learn” TV controlling infra-red signal from remote, subsequently re-emitting the signals to control TV. Process of learning infra-red signals is operated by counting the time interval between level 1 and level 0 of infra-red signal series emitted from TV remote. It then stores into the memory. When receiving any order of controlling functional buttons from computer, this module will re-emit the corresponding series of controlling order that it learned before. Block diagram of remote control TV module is shown in figure 4 below:

Figure 4: Block diagram of remote control TV module

Figure 5: Remote control TV module

IV. RESULT A set of 18 keywords was chosen for this study, namely “tắt, bật, tivi, tăng, giảm, chuyển, âm, kênh” and “một, hai, ba, bốn, năm, sáu, bảy, tám, chín, không”. Another set for purpose of identifying TV controlling orders was also included.

Speech database was collected from 150 speakers (70males, 80 females) whose voices were recorded in low-noise environment condition. The sound files were recorded by adobe audition software, using PCM and obtaining sample at frequency of 16000 Hz, 16 bit. Sound recording was carried out with 2 purposes: preparing database for training and for recognition process. In experimental evaluation process, the commands which were spoken by 100 people were recorded. For training and evaluating the identification, the results were divided into 2 groups: 100 in trained group, and 50 in untrained group.

57

The recognition results of the words Table 1: The recognition results of the words

The identification results of the commands

Table 2: The recognition results of the commands (100 people trained)

Table 3: The recognition results of the commands

(50 people untrained)

V. CONCLUSION The study has successfully designed the Vietnamese speech recognition system by applying the combination of VQ method and HHMs. The speech recognition was indicated to perform well and rather stable in low-noise environment condition. TV control module which communicated with computer by RS232 interface also performed well. It could "learn" and control almost all kinds of TV. In order to achieve the better recognition and be able to apply in practice, it is necessary to build a bigger database with a more diverse set of vocabulary. Other noise reduce methods, moreover, should be applied in pre-processing stage.

REFERENCES

[1] Gales. M. and S. Young, “The Application of Hidden Markov Models in Speech Recognition”, Foundations and Trends in Signal Processing, Vol.1, No.2, 2007, p.p 195-304.

[2] L. R. Rabiner, “A tutorial on hidden Markov models and selected applications in speech recognition,” Proceedings of IEEE, vol. 77, no. 2, February 1989, pp. 257–286.

[3] Juang, B. H. and L. R. Rabiner, “Hidden Markov Models for Speech Recognition”, Technometrics, Vol.33, No.3, August 1991, pp. 251-272.

[4] Linde Y., Buzo A., and Gray R. M., “An Algorithm for Vector Quantizer”, IEEE Transactions on Communication, Vol.28, No.1, 1980, pp. 84-95.

[5] Segura J. C., Rubio A. J., Peinado A. M., Garcia P., and Roman R., “Multiple VQ Hidden Markov Modeling for Speech Recognition” Speech Communication, Vol.14, 1994, pp. 163-170.

[6] Balwant A. Sonkamble, D. D. Doye , “Speech Recognition Using Vector Quantization through Modified K-means LBG Algorithm”, Computer Engineering and Intelligent Systems, ISSN 2222, Vol.3, No.7, 2012, pp.137-144.

[7] Lawrence Rabiner and Biing-Hwang Juang, “Fundamentals of speech recognition”, Prentice-Hall International, Inc, 1993.

[8] Huang X. D., Hon H., Hwang M., and Lee K., “A Comparative Study of Discrete, Semi Continuous, and Continuous Hidden Markov Models”, Computer Speech and Language, Vol.7, 1993, pp. 359-368.

[9] http://www.ustr.net/

58

ECG Signal Transmission Using Wireless Technology in Patient Health-care and Monitoring

System

Duong Trong Luong, Nguyen Duc Thuan Dept. of Biomedical Engineering

Hanoi university of Science and Technology Hanoi, Viet Nam

Nguyen Hung Dept. of Biomedical Engineering

Hanoi university of Science and Technology Hanoi, Viet Nam

Abstract- The transmission and reception of biosignals from human body using wireless technology have been developed several recent years and designed mainly for center patient monitoring system that includes a center monitor and station monitors in hospital (so-called Patient Monitoring system). Due to the development of telecommunication and Internet, biosignal transmission using wireless and Internet will lead to big utilities in patient treatment and examination in real time without limitations in time and distance. With the aim of developing ECG signal monitoring system using wireless technology, this research focuses on the best method of designing an ECG signal transmission-reception system using Zigbee standard. The system was tested with simulated ECG signal at a distance of 40 m. Results obtained show that ECG signals received have quality equivalent to that of input ECG signals from transmission module. This is a basic for the author group to do research on developing and insuring the reliability, reducing the level of interference in signal transmission and implementing transmission in long distances and connected via the Internet.

Keywords- ECG signal transmission using Zigbee; patient remote monitoring system

I. INTRODUCTION ECG signals in particular and biosignals from the human

body in general are the most important physiological signals of human. According to several studies, millions of people die of cardiovascular diseases each year all over the world and also there are many patients in danger due to heart attacks and strokes due to late diagnosis and treatment. People with heart diseases and people due to late detection cannot afford to go to the hospital regularly. This leads to the need for a remote ECG signal monitoring system.

In Vietnam, the study of biological signals from the human body in general and the ECG signals in particular using wireless technology to connect to the Internet has great significance in the field of diagnosis and patient health-care. However, up to now no publication on this field has been reported. Only one project was proposed in 2012 to coordinate between the company NTT EAST-Japan and Vietnam Post and Telecommunications (VNPT) in Vietnam in remote health management services (ehealth-care management), which measure the parameters of weight, blood pressure, heart rate, blood sugar, blood fat, etc. These quantities are expressed in

text and stored in database system and transmitted over a telecommunication network of VNPT, without ECG signal processing and transmission (graphic form). In fact, there are several types of noises intermingle with ECG signals during recording and transmission. Some of the noises affecting ECG signals are AC 50 Hz noises, the others are noises due to patient’s movement, electrical interference in the propagation environment, baseline noises (low-frequency noises influence the process of receiving and analysis of ECG signals; especially when surveying the ST-T wave of the ECG signal). Therefore, in addition to system design problems on ECG signal acquisition and transmission by wireless technology, the issue of noise reduction is very important in the process of receiving and processing the ECG signal. In the framework of research and publication in this paper, we only focus on designing method of ECG signal transmission system using Zigbee standard and displaying on the computer.

II. THEORETICAL BASIS 2.1. ECG signal

ECG signal is one of the important parameters of human and is a typical signal for study and analysis of pathological and biological mechanisms of the heart. ECG signal frequency is in range of from 0.05 Hz to 100 Hz. This frequency range is used for diagnostic applications of cardiovascular diseases. In addition, the frequency range of ECG signals from 0.05 Hz to 35 Hz is used to monitor the patient's health relating to heart diseases.

Figure 1.ECG signal

[[

Ampl

itude

(mV)

Time(ms)

59

Basic characteristics of the ECG signal are shown in Table 1 and Table 2 below:

TABLE 1. THE AMPLITUDE OF WAVES IN THE ECG SIGNAL

Waves in the ECG signal Amplitude (mV) P wave 0.25 R wave 1.6 Q wave 0.4 T wave 0.1 0.5

TABLE 2. TIME INTERVALS BETWEEN THE WAVES IN THE ECG SIGNAL

Waves in the ECG signal Time intervals between the waves (S)

P-R 0.12 0.2 Q-T 0.35 0.44 S-T 0.05 0.15

P wave 0.05 0.11 QRS 0.05 0.09

2.2. Zigbee (IEEE 802.15.4 standard)

Regarding the design and application of signal transmission in wireless personal local area network (WPAN or WBAN), the basic trends are: data transfer at low speed, medium, and high speed. In particular, Bluetooth data transfer at medium speed, Zigbee at low data transfer rate. There is no available IEEE standard for high-speed data transmission in this network. Zigbee is an architecture that was developed based on the IEEE 802.15.4 standard, which has lots of advantages compared to Bluetooth. The comparison of some of the main parameters between Zigbee and Bluetooth standards is shown in Table 3. In addition, Zigbee is the perfect solution for applications based on sensor network. Zigbee’s data transfer speed in range of from 20 Kbps to 250 Kbps with different frequency bands which is shown in Table 4.

TABLE 3. COMPARISON OF ZIGBEE/ IEEE 802.15.4 WITH BLUETOOTH /IEEE 802.15.1

Standard Zigbee/ 802.15.4 Bluetooth/802.15.1 Transmission distance (m)

1-75 1-10

Time-life battery (date)

100-1000 1-7

Number of nodes in network

> 64000 7

Application Monitoring and control

Web, email, video

Stack size (Kb) 4-32 250 Data rate (Kbps) 20-250 720

TABLE 4. DATA TRANSFER RATE OF ZIGBEE WITH THE DIFFERENT FREQUENCY BANDS

Bandwidth Applicability Data transfer rate

channels

2.4GHz World 250Kbps 16 915MHz USA, Japan 40Kbps 10 868MHz Europe 20Kbps 1

Zigbee has three basic topologies: Star, mesh and Cluster-

tree topology. Each topology has its own characteristics, and is suitable for certain applications. Mesh or Cluster-tree topology is used for expanding the scope of activities of the network in about one kilometer.

Following are some characteristics of Zigbee:

Data transfer rate: 20 Kbps to 250 Kbps Low power consumption. In particular, the capacity

of the receiver / transmitter signal is between 25 and 35 mA, the capacity to "stand by” is 3 A.

Transmission distance in sight: 75 m High reliability Low costs Extended ability to 65000 nodes

III. DESIGN OF ECG SIGNAL ACQUISITION AND TRANSMISSION SYSTEM USING ZIGBEE STANDARD

3.1. Block diagram of the system 3.2. Function of each module.

Electrodes: measure the ECG signal with the necessary lead.

Amplifier and noise filter: Amplifier circuit includes: pre-amplifier and

instrumentation amplifier circuit. Pre-amplifier circuit will amplify the signal obtained from electrodes. In this circuit, OP07 IC is used to control wireless base stations. Instrumentation amplifier uses INA141 IC, which is a IC of

Figure 2. Block diagram of ECG signal transmission and acquisition system using Zigbee standard associated with the computer.

Patient

Electrodes

Amplifier & noise filter

PIC 18F26K20

Transmission module Zigbee

ECG

Acquisition

module Zigbee

IC

max 232

Computer

60

high accuracy (± 0.05%), low power consumption, power supply from ± 2.25V to ± 18V, and is used in general biosignal instrumentation amplifier.

VCC+

J3

TAY PHAI

123

-

+

U6

OP073

26

748

1

VCC-

R5 390K

VCC+

R6 390K

VCC-

VCC+

VCC-

VCC+

OUT1

VCC+

-

+

U3

OP07

3

26

74 8

1

R3

22K

R4

10K

SW1

VG

R2

22K

J4

CHAN PHAI

123

VCC-

U1

INA141

7

1

8

6

4

2

35

V+G1

G2

Vout

V-

Vin-

Vin+Vref

-

+

U2

OP07

3

26

74 8

1

-

+

U7OP07

3

26

748

1

J2

TAY TRAI

123

VCC-

Noise filter circuits include: high-pass filter circuit with cutoff frequency of about 0.05Hz, low-pass filter with cutoff frequency of about 100 Hz, band-pass filter and notch filter to eliminate noise from the AC power supply with frequency 50Hz.

PIC18F26K20: is a microcontroller from Microchip. This is a programmable IC, it performs the conversion of the signal from amplifier and noise filter module into digital format before the signal is fed to the transmission Zigbee module.

Module Zigbee includes PIC18F4620 and MRF24J40MA IC from Microchip. PIC18F4620 is a multifunction Integrated Circuit, it has digital and analog input/output pins. This IC supports RS485, RS232 standards for connections to the computer. On the other hand, it can work in Master/Slaver mode. MRF24J40MA performs reception-transmission of signal via Zigbee standard (802.15.4) with some characteristics such as operation frequency from 2.405 GHz to 2.48GHz, data transfer rate of 250Kbps, using carrier sense multiple access/collision avoid method, auto acknowledgment, check data frame transmission, ..... A Zigbee module is used as a station (node) to receive-transmit signal.

Load

-Vin

+Vin

G =10 with pin 1 no connect to pin 8. G =100 with pin 1connect to pin 8.

Vout

C6 0.047uF

J13

CON1

1

J14

CON1

1

C50.047uF

R11

0

R15

+

-

U6

OP07

3

26

7 14 8

Figure 7. Schematic of Band pass filter circuit

J13

CON1

1

Vcc-

R10

23.7K

R9

23.7K

J12

CON1

1

0

+

- OP07

3

26

7 14 8

C3

0.1uF

C40.047uF

Vcc+

Figure 6. Schematic of Low pass filter circuit

Vcc+

C7

R12

R14

R16

C8

Vcc-

Vcc+

R13

J15

CON1

1

0+

-

U7

OP07

3

26

7 14 8

Vcc-

R15

0

+

-

U6

OP07

3

26

7 14 8

J16

CON1

1

0

Figure 8. Schematic of Notch filter circuit

Figure 3.Circuit diagram pre-amplifier and ECG signal instrumentation amplifier.

Figure 4. Equivalent circuit diagram inside instrumentation amplifier IC INA141.

Vcc+

0

R8453K

J10

CON1

1 +

-

OP07

3

26

7 14 8

J11

CON1

1

Vcc-

C2

10uF

C1

10uF

R7

226K

Figure 5. Schematic of high pass filter circuit

61

When Zigbee module performs the transmission function (as shown in the block diagram of the system “Fig.2”), PIC 18F4620 will packet and receive data from 18F26K20 IC data and then create data frame in Zigbee standard and send to MRF24J40MA IC which converts received data into frequency to transmit from antenna. When Zigbee module performs the receiving function, MRF24J40MA will receive the signal from the transmission module by antenna and convert it into digital format and send to PIC 18F4620. This IC transfers and separates data to the computer by RS 232 connectivity standard which performed by IC Max 232. ECG signal measured will be displayed on the computer and can be transmitted to other terminals via the Internet.

Figure 9 describes the data transmission algorithm from the ADC inside PIC18F26K20 to the transmission module Zigbee.

IV. RESULTS The design of the proposed system was implemented in the

laboratory of biomedical signal measurement of the Department of Electronics Technology & Biomedical Engineering, School of Electronics and Telecommunication, Hanoi University of Science and Technology. The system has been tested under the following conditions: the distance between the receiver and the transmitter is 40m in sight, without any obstruction; ECG signal used in this system is

generated from the ECG signal simulation circuit. Received ECG signal is displayed on the computer. This signal has the same quality with the signal from the input of the transmission module. The result is shown in Figure 10 and Figure 11.

V. CONCLUSION The author group have researched, designed and tested a

system of ECG signal transmission using Zigbee standard. Initial result shows that the designed circuits work well with acceptable accuracy and credibility. In a subsequent study, the author group will continuously study the noise reduction solution, and carry out tests and transmission with the ECG signals from the human body in the natural environment at larger distances, in the presence of obstruction. The evaluation of results will also be made.

REFERENCES [1] Balambigai Subramanian, “ Efficient Zigbee Based health Care system

for Arrhythmia Detection In Electrocadiogram ,” European Journal of Scientific Research ISSN 1450-216X Vol.69 No.2 (2012), pp. 180-187.

[2] F. Vergari, V. Auteri, C. Corsi, C. Lamberti, “A ZigBee-based ECG transmission for a low cost solution in home care services delivery,” Viale Risorgimento, 2. 40136 BOLOGNA.

[3] H.Labiod, H.Afifi, C. Desantis, “WifiTM, BluetoothTM, ZigbeeTM and WimaxTM.” Spinger, 2007.

Figure 11. Testing of ECG signal transmission - receiving and displaying.

ECG signal monitoring

Figure 10. Testing of system.

start

Initialize

Check Stopbit

i=0;

i<=64?

ADC on, ++;

Array[i-1]= ADC_result;

Zigbee is ready?

i=0;

No

End Yes

Yes

Sent array to Zigbee module;

ADC off; Clear array;

Yes

No

No

ADC_result 10bits divided into 2 bytes, one byte 5bits +header classify high byte, low byte

Figure 9. The data transmission algorithm from the ADC in PIC 18F26K20 to the transmission module Zigbee.

62

[4] S.Chaudhuri et al., “Ambulation Analysis in wearable ECG”. Springer Science+ Business Media, LLC 2009.

[5] BURR-BROWN, “Precision, Low power Instrumentation Amplifier”. Http://www.burr-brown.com/

[6] ”28/40/44-Pin Enhanced Flash Microtrollers with 10-bit A/D and nano Watt Technology”. Microchip Technology Inc, 2004.

[7] “2.4GHz IEEE Std. 802.15.4 RF Transceiver Module”. Microchip Technology Inc, 2008.

63

Combination of analog and digital solutions for Wireless ECG Monitor

Khue Tra – Phung T.Kim Lai – Khiem Duy Nguyen – Tuan Van Pham Electronic & Telecomm. Engineering. Dept., Da Nang University of Technology, Danang

trakhue08dt3@gmail.com, phung08dt3@gmail.com, nguyendkhiem89@gmail.com, pvtuan@dut.udn.vn

Abstract— Signal processing, in various forms, is the main part of a large number of medical equipment today, and it continues to play a crucial role in the analysis of medical data. In this paper, combination of analog and digital solutions in wireless Electrocardiogram (ECG) monitor is presented. The received ECG signal from patient is fed into pre-processing unit which has amplification and analog filter such as the Sallen-Key filter, RC high pass filter and then it is post-processed with digital filter. In this way, the ECG signal is processed a better way that still maintains the compact size of the device for portable applications. Last but not least, five important parameters of ECG which are heartbeat, QRS duration, ST interval, R amplitude and T amplitude, are extracted to help doctor to diagnostic disease. Keywords— Electrocardiogram (ECG), monitoring system, MSP430, wireless ECG, digital ECG signal

I. INTRODUCTION Novel methods in medicine which are created by the

application of technology science, have large influences on characteristics of healthcare system. Nowadays, along with usefulness of digital signal in processing and storing signal, digital signal processing is applied popularly, specially in medicine [1] [2].

In older systems of Electrocardiogram (ECG) measurement, the ECG signal is processed in analog region. For example with ECG 6851, most of implementations are based on analog technology [3]. In the ECG 6851, the filters such as highpass filter, lowpass filter, Notch filter are

implemented by analog solutions. The filters contain RC filters or LC filters quite complex, cumbersome.

On the other hand, there are many implementation for ECG signal in digital region. They are quite flexible and can store in long time. However, sometime digital ECG signal processing make distortion and the received ECG parameters will be affected.

In this paper, combination of analog and digital processing is presented. With the hand held device, this is quite perfect. Our device with a moderate price of about one million VND, in accordance with the affordability of small hospitals, patients can afford an ECG monitoring circuit that they can use in their houses. Besides, device is designed with small and light circuit by using three electrode for measuring a lead. Patients carry it along and are followed while doing their normal activities without any feeling of trouble. Instead of forcing one to lie on a position to measure and record the ECG signal in a long time, transmission of signals based on wireless technology allows people to move and act freely. The ECG signal is still recorded as normal within a relatively wide range (several tens of meters). In addition, extremely low power consumption will help to save power for a long-time using, which reinforces the feasibility of our project.

The paper is organized as follows: Section II presents system architecture. Afterward, the analog solutions is shown with the sensor node in Section III. Analysis of digital method is carried out in Section IV. Finally, conclusion and future work in Section V conclude the last paper.

Implementation

Amplifier (INA118)

Filter

ADC 10bit

(MSP430F2274)

Voltage Supply

Driver Leg Circuit

(OP07D)

Wireless Trans-mission

(CC2500)

ECG signal

Sensor

Wireless Receiver (CC2500)

Interface with

CC2500 (MSP430F2274)

Interface with

computer (TUSB-3410UF) Voltage

Supply

Receiver Monitor

Receiver Process Show ECG signal

Fig. 1 Wireless ECG monitor system diagram.

64

II. SYSTEM ARCHITECTURE Three main parts of a wireless ECG monitor system are

shown in Fig.1: sensor node, receiver node and monitor node. The sensor node, worn by the person, is quite small and

portable. In this part, the ECG signal is captured by three electrodes, and then amplified and filtered by the analog circuits. Then the MSP430 in the sensor node implements analog signal to digital signal conversion. At the same time, it sends all the digitized data to the eZ430-RF2500 wireless module, which transmits data through 2.4GHz channel [4]. The received ECG data in the receiver node is sent to PC via serial port connection. In monitor part, PC can be used to implement digital processing and extract five parameters of the ECG signal. Then the PC makes unified database management and information display. Fig. 1 shows a kind of point to point communication system.

III. ANALOG SOLUTIONS The ECG signal detected at the electrode is very small. In

theory, the ECG signal at the electrode is about 1mV but in realistic working condition, it can be smaller [5]. Furthermore, the ECG signal is affected by various kinds of noise such as noise 50Hz from powerline noise (in European countries and the United States is 60 Hz), high frequency noise due to the vibration of the muscles, low frequency noise makes signal drift, etc. In this way, in ECG measurement the biggest challenge is the elimination of the body’s DC offset and the 50/60Hz hum that the human body - as an antenna - collects. From those above, ECG detection is very difficult and how to collect ECG signal properly or remove many kinds of noise more difficult. If the ECG does not collect or collect incorrectly, all continue processing will not implement. Besides, if digital processing is abused too much, the ECG signal will distortion and the received parameters will be affected. Therefore, analog ECG signal detection is very necessary and analog solutions is used in here is properly.

A. The Instrumentation Amplifier In the ECG 6851 catalog [3], amplifier is implemented

using three op-amps with other complex components. This helps the signal is amplified with amplifier of small noise. As for the purposes, the INA118 precision instrumentation amplifier was chosen. We only just with the INA118, those tasks is finished perfectly and simply. With the gain of signal amplifier in the INA118 can be changed flexibility by Rg input resistor. Besides, it has differential inputs with very high common-mode rejection ratio and integrated three op-amps for two amplifier stage. This helps cancel common noise effectively. According to the standard around the world, current load the person is less than 20µA to prevent electrical shock for patients. The INA118 FET-Input instrumentation amplifier may provide lower noise and extremely high input impedance, input current approximately zero, it is very satisfiable for medical applications [6]. When the input signal needs amplify with gain is 1000 into 1V to keep output signal within the input voltage range of the analog-digital converter (ADC). Thus, two 33 Ω resistors are enough to set the gain of the instrumentation amplifier.

B. Implementaion of Analog Filters Since the ECG signal has small amplitude under the

influence of many kind of noise such as the 50/60 Hz noise from the AC network, the DC noise in contacting body with electrode or noise for tremor. Those noise occur in the recorded environment and in the transceiver, the ECG signal is fed into filters for noise suppression before coming the analog-digital converter (ADC). The ECG signal has a wide frequency range but the useful information of ECG signal in medical is usually in the range of frequency from 0.05Hz to 100Hz [7]. In other paper, the filter is implement with a large amount of resistor and capacitor. In this paper, the Sallen-Key topology which is second order active filters is used [8]. This filter is the super-unity-gain amplifier which allows for very high Q factor and pass band gain without the use of inductors. The image of the Sallen-Key filter is shown in Figure 3.

The cut-off frequency can be chosen below the lowest

frequency component of the ECG signal, which is 0.05Hz in diagnostic mode and under the highest frequency component of the ECG signal, which is 100Hz. The cut-off frequency is calculated with the following equation:

푓푐푢푡표푓푓 = √ ∗ ∗ ∗

With the high pass filter, R1 = R2 = 10kΩ, C1 = C2 = 10µF.

C. The Microcontroller The aim of using microcontroller in this project implements

ADC and transmits ECG signals. With portable application, this design needs small and ultra-low power. Furthermore, due to the ECG signal is effect by many type of noise, specially AC noise, we use battery supply provide whole circuit to decrease noise. In this application, we use MSP430F2274 is a part of the family MSP430 – designed 16-bit RISC processors for ultra-low power applications from Texas Instruments. The MSP430F2274 has 10-bit 200ksps Analog-to-Digital Converter (A/D) with internal reference, sample and hold, autoscan, and data transfer controller. Specially the MSP430F2274 supports UART for interface with the computer and SPI protocol for RF transceiver.

Fig. 2 The INA118 amplifier.

Fig. 3 The lowpass Sallen-Key filter.

65

The MSP430F2274 is integrated on eZ430-RF2500 module [4]. This is available and makes a convenient for using RF transceiver whose figure is shown in Figure 4.

D. The Driver Leg Circuit Besides two wires are connected as shown in Figure 5, we

can use isolation amplifiers for the direct patient connection. The body is connected to output of the op-amp. The OP07 is unsurpassed for low-noise, high-accuracy amplification of very-low-level signals which is applicated in this circuit. This reduces the pickup as far as the ECG amplifier is concerned and effective grounds the patient. This connection prevent accidental internal cardiac shock and protect to the circuits and equipment from damage. If abnormal high voltage should appear between the patient and ground due to electrical leakage or other means, the auxiliary op-amp saturates. Ordinary, people connects the third electrode into the right foot. Performance of elimination noise is better than the previous method. However, the purpose of this application is the ECG signal is recorded frequently and patients feel comfortable with carrying this device regularly. Therefore, the third electrode is connected as the Figure 5.

E. The Transceiver Chip The CC2500 is a low-cost 2.4 GHz transceiver designed

for very low-power wireless applications. High sensitivity of the CC2500 is –104 dBm at 2.4 kBaud and 1% packet error rate. Although it is good news for those who build high quality surface mounted boards, some of the potential customers may fall away as a result of difficulties with soldering. In the CC2500, frequency range is 2400 – 2483.5 MHz. This is a standard RF. Hence project development in the future create a network monitor is implemented easily. The CC2500 is integrated on eZ430-RF2500 module [4] as showed in Figure 3.

F. The Voltage Supply The supply is large enough to control the operation circuit.

Firstly, the INA118 needs it provides stable +2.5V for the amplifier and ensures that the negative supply stays at -1.9V. Actually, the amplifier should work well until supply absolute value is higher than 1.35V. Secondly, the ADC10 module operates properly which needs higher supply voltage (2.2V) than other peripherals in the MCU. At minimum voltage is 1.8V, CPU works, but the ADC10 did not. Failures like this are annoying and can corrupt the samples sent to the computer.

To require a dual voltage from a single battery, the voltage inverter is used as shown in Figure 6. This circuit is simple and inexpensive. We use a 6V battery. After the voltage is fed through the voltage inverter, it is divided into ±3V. The LM386 is a power amplifier designed for a saving power consumption application. With purpose creates the same output voltage, we chose R1 = R2 = 470 kΩ and C1 = C2 = 10µF.

IV. DIGITAL SOLUTIONS

The ECG signal is affected by various kinds of noise such as AC 50/60Hz noise, low frequency noise makes signal drift, etc. They are removed mostly in sensor node with analog method. However, nothing is absolute, little of noise still exist with useful signal. Furthermore, in wireless transmission, noise is added the ECG signal. Hence, digital signal processing in computer must be used and it occupy important role.

A. Notch Filter In this part there is described noise elements filtering with

digital filter. The main noise element is power supply network 50 Hz frequency. This noise has to be removed before the signal is used for next data processing like heart rate frequency determination. At computer, this aim is implemented by Notch filter [9]. The Notch filter is a filter that passes all frequencies except those in a stop band centered on a center frequency. The amplitude response form of Notch filter is drawn in Fig. 7.

Fig. 4 The eZ430-RF2500 module.

Fig. 5 The model of ECG amplifier.

Fig. 6 The model of ECG amplifier.

Fig. 7 The response of Notch filter with cutoff frequency is 50Hz.

66

In here, cut-off frequency of the Notch filter is 50Hz. After that, the ECG signal before and after Notch filter is shown in Fig. 8 and Fig. 9.

Ripples in the ECG signal as the Fig. 7 present powerline

interference 50Hz which is eliminated by Notch filter as the Fig. 8.

B. Parameters Detection 1) QRS Detection The QRS complex represents the most important

component within the ECG signal [10]. Its detection is the first step of all kinds of automatic feature extraction. QRS detector must be able to detect a large number of different QRS morphologies.

Algorithm of QRS detection is presented in the figure below [11].

The R peak is the highest peak of the ECG signal. We can

rely upon this peak to detect the rest ones. From the power spectral analysis of the various signal components in the ECG signal, a filter can be designed which effectively selects the QRS complex from the ECG. The received signal is differentiated to provide information about the slope of the QRS complex. Then we use threshold to detect the R peak. In here, adaptive threshold is implemented. We also use a time constant is 200ms to remove inappropriate R peak. If we have two R peaks in 200ms, we remove the second peak. Then adaptive threshold is adjusted automatically.

2) Heartbeat We realize that RR interval is the one heartbeat. From

detection of R peak, we calculate heartbeat by following formula.

퐻푒푎푟푡푏푒푎푡 = 60/푅푅푖푛푡푒푟푣푎푙 3) P, R, T peak amplitude For QRS detection, the R peak is detected. We get its

information to show R amplitude.

To detect P peak, we get the peak just before QRS complex. And between two consecutive R peaks, the maximum value is T peak.

4) ST interval From S peak is detected in the QRS complex detection and

T peak is detected as above algorithm, we calculate ST interval easily.

V. RESULTS AND EVALUATION

The final version of project is depicted in behind figure.

The first method in experiment results of system is

implementation with ECG simulators at different modes, such as difference for frequency, amplitude, etc. In this case, frequency we set for the ECG simulator is 60Hz, 75Hz or 35Hz, and on the tested waveform at the output, we found that the frequency is also approximately the set frequency. In figure 8, the ECG signal is set at 60 Hz and heartbeat is calculated from the wave is 5*60/4 = 75 times / 1 minute. Besides, for a standard ECG signal, the QRS, ST and PR parameters must satisfy at some specific intervals such as the QRS complex about 50-100 ms, ST interval about 200-300 ms and PR is 60-120 ms [12]. With these values, doctors can base on them for heart diseases detection. On experimental results with simulator, we also found the ECG parameters’ intervals satisfy with simulating signal. For the purpose checks amplitude, the ECG signal from the ECG simulator is set at 1mV and gain of amplifier is 1000. We see that at the ECG output after amplifying circuit 1V peak. This shows our amplifying system works fine and stable with simulating signal.

The tested results as discussed above show the capability

of our system with ECG simulator which produces ECG signal found in a normal, healthy person. The received ECG

Fig. 8 The ECG signal before Notch filter.

Fig. 9 The ECG signal after Notch filter.

Fig. 10 The QRS algorithm.

Fig. 11 Final project.

Fig. 12 The ECG signal in oscilloscope.

67

signal in oscilloscope is shown in Fig. 12. At this point, we can conclude that: the acquisition data block, wireless transmission and PC monitor are stable and accurate for ECG simulator.

The second method is testing with real persons. We conducted experiments with different scenarios including: 6 healthy students, 2 students after drinking coffee and 2 persons after waking up. This result is depicted in Table I.

TABLE I SYSTEM RESULT FOR TESTING ON REAL PERSON

ID student Heartbeat Student 1 75 Student 2 78 Student 3 70 Student 4 75 Student 5 80 Student 6 70 Student 7 72 Student 8 73 Student 9 76

Student 10 80 According to the table’s results, we can ensure that our

system works appropriately with real situations. For practical uses, one of the most challenging issues for any lab product is noise, however we was successful in solving noise cancellation on our device as shown on the table and this will secure the its application to daily life.

For the amplifier, in this project, we use ±3V rail-to-rail amplifiers. This is one advantage comparing to ECG 6851 commonly found in hospital [3], which use ±9V voltage supply. Based on our expected functions for the device, it will consume less energy, thus is more power-efficient and the analog part is less burdensome than other ECG systems. Also, our device use batteries as the main power supply instead of AC source like other ECG system sold available in the market, therefore, we do not need any isolation circuit for it and keep our device as small as possible. Furthermore, most of the current ECG systems use analog implementation for filtering [3]. This is one of the reason why the size for those systems are not small enough for portable applications. Our device, on the other hand, applies the combination of analog and digital filters according to the research of Mr Tamas Hornos [13]. This clever solution enables us to keep the device small for hand held purposes. The wave we receive at the PC after digital processing and feature extraction is shown in Figure 13.

In an indoor corridor, maximum transfer distance of more than 20 meters is fairly good. Because in a real operation situation, due to absorption of structures or more radio noise can be present, the signal can be attenuated, more attention should be paid in future into improving the reliable transfer distance.

VI. CONCLUSION AND PERSPECTIVES A new approach for ECG signal monitoring using wireless

technology was designed and implemented for homecare system of elderly person or patients in this work. The focus of this paper is to design a compact wireless ECG monitoring device using commercially available electronic components. The wave form in screen is pretty accurate and smooth, showing how well our device works. Wireless communication’s implementation in this paper is believed that it will bring benefits not only to the hospital care systems, but also to those who need 24 hours of ECG monitoring. In conclusion, the system implemented has the ability to redefine not only for current hospital care, but also for work, home and other activities.

To make the device become more effective and applicable, future work will develop this project the wireless sensor network or publish ECG on to a website or mobile phone via bluetooth connection. As a result, doctors can take care of many patients at the same time just with a portable device. Besides monitoring the ECG signal, this includes automatic ECG recognition to interpret the ECG of the user in determining the present or upcoming heart problems. These will create a larger application for this project.

ACKNOWLEDGMENT The authors would like to thank Mr Hieu V. Nguyen, Da

Nang University of Technology, Mr. Hung T. Le, Texas Instruments, Vietnam and Texas Instruments Incorporated.

REFERENCES [1] Cristian Vidal Silva, Andrew Philominraj, A DSP Practical

Application: Working on ECG Signal, Chile, November 23, 2011. [2] http://www.ntnu.edu/studies/miel/components/signal-processing-in-

medical-applications. [3] Field,Goth R., Conely J., “ECG recorder : Nihon Kohden ECG-6851

K”. [4] eZ430-RF2500 Development Tool User's Guide, Texas Instruments

Incorporated, SLAU227E – September 2007 – Revised April 2009. [5] Abhishek Joshi, Sourabh Ravindran, EKG-Based Heart-Rate Monitor

Implementation on the LaunchPad Value Line Development Kit Using the MSP430G2452 MCU, Texas Instruments Incorporated, March 2011.

[6] ECG and EEG Applications Quick Reference Guide, Texas Instruments Incorporated.

[7] Sutar,R.G. ,Electron. Dept., Mumbai Univ., Nagpur, India, ECG Feature Extraction Using LCAD, May 2012.

[8] Sonia Behra,PremprakashMourya,KamleshChaudhary,Vikas Mishra, Design of SallenKey Low Pass Filter for High Bandwidth, 2012.

[9] Piskorowski, I., Powerline interference removal from ECG signal using notch filter with non-zero initial conditions, 18-19 May 2012.

[10] Zahia Zidelmal, Ahmed Amirou, Mourad Adnane, Adel Belouchrani, QRSdetection based on wavelet coefficients, September, 2012.

[11] Valtino X. Afonso, ECG QRS Detection. [12] DSaskia A.B.E., Klaas, Monohydroxyethylrutoside as protector against

chronic doxorubicin-induced cardiotoxicity, 19 Jul 2012. [13] Tamas Hornos, Department of Electronics and Electrical Engineering

University of Glasgow, Wireless ECG/EEG with the MSP430 Microcontroller, Thesis submitted for the degree of Master of Science, 2009. Fig. 13 The ECG signal in PC.

68

Biodegradation of Phenol by Native Bacteria Isolated from Dioxin Contaminated Soils

Bui Ba Han, Nguyen Thi Lan, Dang Duc Long*

Department of Biotechnology, Faculty of Chemistry Da Nang University of Technology, The University of Da Nang

54-Nguyen Luong Bang Street, Da Nang City, Vietnam *Email: ddlong@dut.edu.vn

Abstract - In this investigation, aerobic bacteria in soil contaminated with dioxin (taken from Da Nang airport’s area in Vietnam) were isolated and selected for their ability to degrade phenol using an enrichment technique containing phenol as the sole source of carbon and energy (100 mg/L phenol in a mineral salt medium). Four strains (designated D1.1, D1.3, D1.4, and D1.6) were obtained and characterized. The results showed that these bacteria were highly effective for the removal of phenol. After 120 hrs of culture, strain D1.4 degraded 54.84% and 44.19% phenol from the initial concentrations of 100 mg/L and 1,000 mg/L, respectively; strain D1.6 degraded 66.45% of phenol from the initial concentration of 1,500 mg/L. The combination of those bacteria in the same medium had a positive effect on the phenol degradation activity. The outcome of the study can contribute new useful resources for treatments of wastewater and soils contaminated with phenolic wastes.

Keywords: dioxin-contaminated soil, phenol degradation, aerobic bacteria, isolation.

I. INTRODUCTION In the development of the world today, human health and

the environment have become the most pressing issues. Due to those concerns, biodegradation of aromatic compounds have received a great deal of attention because of their toxicity and persistence in the environment. Among all those compounds, phenol and their derivatives are amongst the most prevalent forms of chemical pollutants since they are commonly used to produce many resins, dyes, paints, varnishes, detergents, herbicides and pharmaceutical drugs. They are also by-products of many big industries such as petroleum processing, coke conversion, and steel manufacturing [1, 2, 3]. Phenol can occur naturally in some agricultural products, animal wastes and decomposition of organic materials [4]. However, it is documented to have harmful effects on human health and the environment. Phenol is a water-soluble and highly mobile neurotoxin and can cause damage to the human body and other living organisms through ingestion, inhalation or contact [1, 5]. As a potent air pollutant, phenol can damage structures and the ozone layer and may reduce visibility and the heat balance of the atmosphere [5]. Therefore, phenol has been declared to be a hazardous substance and a hazardous air pollutant by the United States Environmental Protection Agency [6].

A variety of physical, chemical and biological methods have been used for the safe removal of such a chemical from the environment. One of the cheapest and safest solutions for this is by bioremediation using microorganisms [7]. In general, it is difficult to degrade phenol by biological methods when its

concentration is above 200 mg/L and thephenol-degradation activity of microorganisms are completely deactivated at concentrations larger than 3,000 mg/L [8]. Nevertheless, microbial degradation of phenol with different initial concentrations ranging from 50-2,000 mg/L have been carried out in various reactor systems (e.g., shake flask, fluidized-bed reactor, continuous stirred tank bioreactor) using a variety of fungi and bacteria such as Candida tropicalis, Acinetobacter calcoaceticus, Alcaligenes eutrophus, Pseudomonas putida, Burkholderia cepacia G4, Bacillus stearothermophilus [9, 10]. Those microorganisms have usually been isolated from environmental samples with high concentrations of pollutants. In Vietnam, due to historical reasons, there are many sites contaminated heavily with chemical reagents, especially dioxin. In those sites there are serious environmental problems, but the sites also harbour microorganisms with the ability to metabolize toxic chemicals.

In this study we aimed to isolate naturally occurring bacterial strains present in those contaminated sites and characterizing their efficiency of phenol degradation.

II. MATERIALS AND METHODS

A. Soil Samples and Chemicals Soil samples were collected in Da Nang’s airport area in

Vietnam in the spring of 2010. This soil is heavily contaminated with dioxin from Agent Orange used in the Vietnam War (1964-1973). At present, the dioxin concentration in soil there is up to approximately 200 pg TEQ/g soil [11]. Phenol was obtained from Sigma-Aldrich Co., UK. The mineral salt medium (MM) used for bacterial enrichment and isolation was modified from the one used by Fortnagel and coworkers [12] and comprised (per liter): 3.5g of Na2HPO4.2H2O, 1.0g of KH2PO4, 0.5g of (NH4)2SO4; 0.1g of MgCl2.6H2O, 50mg of Ca(NO3)2.4H2O, 1 ml of vitamin B12, and 1 ml of trace salt solution. The final pH of the medium was 7.2. The trace salt solution contained 0.01g MoO3, 0.07g ZnSO4.5H2O, 0.005g CuSO4.5H2O, 0.01g H3BO3, 0.01g MnSO4.5H2O, 0.01g CoCl2.6H2O and 0.01g NiSO4.7H2O in 100 ml water. The MM solid medium contained 10 g of agar (Meck Co., USA) per liter.

B. Isolation procedure

Soil samples were passed through a sieve (1.7 mm mesh) to remove large pieces of debris and vegetation. A 0.5% agar medium (agar in distilled water) was autoclaved and cooled down (45oC). A 100 ml of agar medium was mixed well with

69

1 g of the sieved soil samples. Then 5 ml of the resulting suspension was spread evenly onto Petri dishes containing the sterile solid MM medium, which was supplemented with phe-nol (100 mg/L). Those dishes were cultured in an aerobic con-dition at 30°C for 120 hrs. After this incubation, individual colonies were transferred onto new MM agar plates with and without phenol (also 100 mg/L) as the carbon source. After incubating at 30°C for 120 hrs, we selected the colonies that grew on the medium with phenol, but were unable to grow on the medium without phenol. Well grown colonies were maintained on nutrient agar slants and stored at 4°C ± 1oC until used for further experiments. Basic biological characteristics of the isolated bacterial strains were carried out by standard laboratory procedures [13].

C. Growth of the bacterial strains in phenol and aromatic compounds

The isolated cultures were used to inoculate MM containing phenol at different concentrations and grown at 28°C on an orbital shaker at 150 rpm. Similar experiments for MM containing benzene, aniline and dioxin-contaminated soil as the carbon source were also conducted. Control samples of MM without phenol or the other organic substrates were prepared for reference. All the cultures were limited in exposure to light. Samples were aseptically removed at regular intervals and analyzed for growth, substrate removal and pH. The growth was measured by analysing cell numbers, total protein concentration, and biomass. To estimate the cell number, 5 ml of the culture medium was centrifuged at 4,000 rpm for 20 minutes at 4°C ± 1oC, then the obtained pellet was resuspended in 4 ml of phosphate buffer (0.05M, pH 7.2) and worked up by the method of Bedard et al. [14].

For analysis of the total protein concentration (mg/ml medium), the samples after the centrifugation were washed a few times with fresh (phenol-free) MM to remove the substrate. After cell lysis in the presence of NaOH (0.15 M) for 5 min at 95oC [14], the total protein concentration was determined by calibration with bovine serum albumin (BSA) standards according to Biure [15]. The biomass (g/L of medium) was estimated by a dried weight method [10]. Specifically, 10 ml of the culture broth was centrifuged as mentioned above, then the pellet was washed twice and finally transferred from the tube into a pre-weighed 1.2 µm pore filter paper (Whatman GF/C). The filter paper were dried in an oven at 105oC for between 72 hrs, cooled in a desiccator at room temperature and reweighed to estimate the dry weight of the biomass. The phenol concentration was determined using a 4-aminoantipyrine colorimetric approach [16]. The supernatant of the centrifuged culture medium was reacted with 4-aminoantipyrine at pH 7.9 ± 0.1 forming a brownish-orange compound. Subsequently sample absorbance was measured at 500 nm. The phenol concentration was calculated by referring to the standard curve.

D. Determination of enzyme activities in cell extracts Manganese peroxidase (MnP) is an enzyme catalyzing the

Mn(II) and H2O2-dependent oxidation of lignin and a variety of phenols. We measured the MnP activities of the isolated bacteria using an assay based on the oxidative coupling of 3-

methyl-2-benzothiazolinone hydrazone (MBTH) and 3-(di-methylamino)benzoic acid (DMAB) [17]. The reaction of MBTH and DMAB in the presence of H2O2, Mn(II), and MnP gives a deep purple-blue color with the absorption peak at 590 nm. Catalase and oxidase activities were measured by tradi-tional biochemical methods [18].

III. RESULTS

A. Enrichment, isolation and characterization of phenol-de-grading bacteria

Starting with two different dioxin-contaminated soil sam-ples at Da Nang airport (D1 and D2 location), we have been able to isolate four bacterial strains (D1.1, D1.3, D1.4, and D1.6) from the D1 sample (Figure 1). The isolates were all Gram-positive, formed yellowish and slimy colonies, and grew under strictly aerobic conditions. The strains reacted nega-tively in an acid resistance test, produced no H2S and NH3, showed catalase, but not oxidase activity (Table I). From the second location (D2), we failed to isolate any aerobic bacteria.

B. The growth of the bacteria in media containing phenol or other organics

Figure 1. Colonies of the bacterial strain on MM agar. A-D1.1 strain, B-D1.2 strain, C-D1.3 strain, D-D1.6 strain.

TABLE I. MORPHOLOGICAL AND BIOCHEMICAL CHARACT-ERISTICS OF THE PHENOL-UTILISING BACTERIAL STRAINS

Characteristics D1.1 D1.2 D1.3 D1.6

Gram reaction + + + + Cell shape Coccus Coccus Oval Oval Oxygen Requirement Aerobic Aerobic Aerobic Aerobic

Glucose test + + + + Maniltol test + + + + Lactose test + + − + Glucose fermentation + + + + H2S product − − − − NH3 product − − − −

Catalase reaction + + + + Oxydase reacton − − − − Aerobic growth + + + + Polyphosphate + − + + Acid resistance − − − −

70

The growth of the four bacterial strains on MM containing phenol as a the sole carbon and energy source are shown in Figure 2. The bacterial strains D1.3, D1.4, and D1.6 rapidly reached log phase (several hours), while strain D1.1 exhibited slower growth and reached a lower cell density with lower protein concentration. It should be noted that the amounts of protein produced by the bacteria were not tightly linked to the cell numbers, especially in the case of the strain D1.3. After 120 hrs of incubation, the biomass of strain D1.6 was highest (0.8 g/L), followed by D1.4 (0.73 g/L), and the lowest value was observed for strain D1.1 (0.24 g/L). Those growth profiles indicated that strain D1.1 grew more poorly in the phenol medium. Also, after cell incubation, the phenol amounts that were degraded by the four strains- D1.1, D1.3, D1.4, and D1.6 were 31.98%, 51.64%, 54.84% and 47.07%, respectively. Clearly, the phenol-degradation capabilities of the strains were not proportionally related to their growth. The reason is that the phenol degrading activity is dependent on amounts and characteristics of metabolic enzyme systems in those strains, not on their cell numbers. To know more about the metabolic characteristics of the bacterial strains, we grew them in various

liquid media containing other organic carbon sources, from simple sugars to toxic aromatic compounds. The strains grew well on media containing glucose, lactose and mannitol, respectively. They also showed the ability to grow in media containing benzene and aqueous extract of the dioxin-con-taminated soil, but not in aniline medium (data not shown).

C. Enzyme activities in crude cell extracts

After 120 hrs of culture, the biomass of each strain was obtained and the cells were lysed. The activities of MnP of the four cell extracts showed that strain D1.4 had the highest enzyme activity, followed by strain D1.3, then strain D1.6, and strain D1.1 had the lowest activity (Table II). This pattern was similar to the variation of the phenol-degrading capability of those strains. TABLE II. MNP ENZYME ACTIVITY (U/L)a OF THE BACTERIA AFTER

120 HRS OF CULTURE ON LIQUID MM

Bacterial strains isolated D1.1 D1.3 D1.4 D1.6

MnP enzyme activity (U/L) 2.9690 5.9791 7.4738 5.2317

0 24 48 72 96 120 1440

20

40

60

80

100

Strain D1.1

Time (h)

Phen

ol (m

g/lit

er)

0.00

0.02

0.04

0.06

0.08

0.10

0.12

Cell num

berx108/m

l

Protein (m

g/ml)

0.00

0.01

0.02

0.03

0.04

0 24 48 72 96 120 144

0

20

40

60

80

100

Strain D1.3

Time (h)

Phe

nol (

mg/

liter

)

0.00

0.05

0.10

0.15

0.20

Protein (mg/m

l)

Cell num

berx108/m

l

0.00

0.01

0.02

0.03

0.04

0.05

0 24 48 72 96 120 1440

20

40

60

80

100

Time (h)

Phe

nol (

mg/

liter

)

0.00

0.05

0.10

0.15

0.20

Strain D1.4

Cell num

berx108/m

l

Protein (m

g/ml)

0.00

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0 24 48 72 96 120 144

0

20

40

60

80

100

Strain D1.6

Time (h)

Phen

ol (m

g/lit

er)

0.00

0.05

0.10

0.15

0.20

0.25

Protein (m

g/ml)

Cell num

berx108/m

l

0.00

0.01

0.02

0.03

0.04

0.05

Figure 2. The growth of four microorganism strains (D1.1, D1.3, D1.4, and D1.6) on mineral salts liquid medium. The culturing process was conducted by culturing 1 ml of the sample of each strain in 50 ml of mineral salts liquid medium with 100 mg/L phenol as only the sole source carbon and energy. After an appropriate time, the phenol concentrations (mg/L), the total cells (cells×108/ml), and the total protein (mg/ml) were analyzed as described in Materials and Methods. Symbols: , phenol concentrations in the culture medium; , total cell numbers of bacteria; , total protein concentration.

71

aAn unit of the enzyme activity was defined as the amount of enzyme required for the oxidation of 1 µM of phenol red per min measured at the wavelength of 610 nm.

D. Ability to grow in different phenol concentrations of the bacterial strains

All four isolates grew in a series of liquid MM with the phenol concentrations varying from 100 mg/L, to 300, 1000, and 1500 mg/L. Based on the resulting cell numbers, protein concentrations, biomass, and phenol concentrations, we could observe that the bacterial strains D1.1 only grew well in 100 mg/L phenol, but grew poorly in the higher concentrations of phenol. Meanwhile, the strain D1.3 only grew poorly when the concentration of phenol was up to 1,500 mg/L, and the strains D1.4 and D1.6 grew well in the whole range of the phenol concentration. The growth of two bacterial strains (D1.4 and D1.6), which seemed to adapt best to the phenolic envi-ronment (Figure 3), growing in 1,000 mg/L (strain D1.4) and 1,500 mg/L (strain D1.6).

E. Phenol degradation ability of a mixed culture of the bacte-rial strains and the native soil microorganism systems

A synergistic action of microorganisms can be helpful in metabolizing and degrading organic chemicals [5, 19]. To test this with our isolates, we mixed the four strains together in equal numbers and grew the mixed culture in MM containing 100 mg/L. For comparison, the two initially dioxin-contaminated soil samples were also cultured in the same medium and the amounts of soil were calculated so that they carried similar numbers of bacterial cells. The phenol degradation activities of those cultures during 120 hrs of incu-bation are shown in Figure 4. Kinetics of phenol degradation of D1, D2 and D1.1-6 were different. For D1, the phenol degradation was slowly at first, however, at the latter stage (after 96 hrs of culture) it occured very quicky. Meanwhile, D1 and D1.1-6 degrade phenol very quickly at the first 72 hrs after that the degradation was slowed down greatly. However,

all the systems of the microorganisms degraded similar amounts of phenol after 120 hrs.

F. Influence of environmental factors on the phenol-degrading capability of the bacterial strains

From the results mentioned above, we found that strain D1.4 had the best ability to degrade phenol in MM medium. Therefore, we investigated the influence of environmental

0 24 48 72 96 120 1440

200

400

600

800

1000

Strain D1.4, 1000 mg/liter

Time (h)

Phe

nol (

mg/

l)

0.00

0.02

0.04

0.06

0.08

0.10

0.12

Cell num

berx108/m

l

Protein (m

g/ml)

0.00

0.01

0.02

0.03

0.04

0.05

0 24 48 72 96 120 144

0

400

800

1200

1600

Strain D1.6, 1500 mg/liter

Time (h)

Phen

ol (m

g/l)

0.00

0.03

0.06

0.09

0.12

0.15C

ell numberx10

8/ml

Protein (mg/m

l)

0.00

0.01

0.02

0.03

0.04

0.05

Figure 3. The growth of two microorganisms (D1.4 [1,000 mg/L] and D1.6 [1,500 mg/L]) in mineral salts liquid medium. Phenol concentrations and the growth of the microorganism were determined as described in the Material and Methods. Symbols: , phenol concentrations in the culture medium;

, total cell numbers of bacteria; , total protein concentrations.

Figure 4. The growth of microorganism strain D1.1-6 system and dioxin contaminated soil (D1 and D2). Mineral salts liquid medium was added to the phenol concentrations of 100 mg/liter, and was cultured at 28°C, with a shaking speed at 150 rpm. Phenol concentrations and the growth of the cultured bacteria were determined as described in the Material and Methods. Symbols: , dioxin soil sample D1; , dioxin soil sample D2;

, the mixture of microorganism strain (D1.1, D1.3, D1.4 and D1.6). Values are the means of 3 replicates and error bars represent the standard deviations.

72

factors on the phenol-degrading activity of this strain. This set of experiments would be helpful to identify optimal conditions of phenol degradation using the isolated bacteria. Figure 5 shows the effects of NaCl, glucose concentrations and pH of the growth medium on the final phenol concentration after 120 hrs of incubation of the strain D1.4 in the MM containing 100 mg/L phenol.

Low or average concentration of NaCl (from 1 to 10%) showed a small effect on the capability to degrade phenol of the bacteria (Figure 5A), only high concentration of NaCl (15%) began to inhibit the activity of the bacteria. That probably dues to sterilizing effect of NaCl at that high concentration. Figure 5B shown that glucose (in the NaCl-free mediun) had an optimum range of the concentration for the activity of D1.4 bacteria (around 0.75% w/v).

Higher or lower concentrations of glucose inhibited significantly the phenol-degrading activity of this strain. At the condition of 0% NaCl and 0.75% w/v glucose, the optimum pH value for the acitivity of D1.4 was about 8 (Figure 5C).

IV. DISCUSSIONS

Much research on biodegradation of phenol using pure or mixed microorganisms isolated from various environmental sources have been reported [2-5, 7-10, 19-22]. However, studies on phenol degrading ability of microorganisms from dioxin-contaminated soils have not been reported. In this study, we have isolated four aerobic bacterial strains from dioxin-contaminated soil that can all use phenol as the source of carbon and energy for their growth.

In addition, those bacteria can also utilize several toxic aromatic compounds, such as benzene and dioxin derivatives. Due to this kind of metabolic capability, those bacterial strains can degrade phenol and different aromatic contaminants. The phenol degradation capabilities were not the same for those isolates. Through culturing the strains in a wide range of phenol concentration, we selected two strains, D1.4 and D1.6, which can grow and metabolize a very high concentration of phenol (1,500 mg/L). Our assay suggested that the phenol-degrading activity of the bacteria were closely linked to the activity of manganese peroxidase, an important enzyme for

Figure 5. Effect of some factors on phenol degradation by D1.4. (A) The influence of NaCl, (B) glucose concentrations, (C) and pH on the ability of phenol decomposition by strain D1.4 in MM containing 100 mg/L phenol. A mineral salts medium was supplemented as Materials and Methods. After 120 hrs culturing, the remaining phenol concentrations was determined. Values are the means of 3 replicates and error bars represent the standard devations.

73

degradation of a wide variety of chemicals and polymers. These strains can be used as a biological reagent for processing phenol and aromatic chemicals in soil and wastewater. Overall, our study here demonstrates that dioxin-contaminated soils are valuable sources of microorganisms that can be beneficial in environmental protection as well as in other fields.

However, we have not been able to isolate all the phenol-degrading microorganisms existing in the soil samples. For example, the soil sample D2 showed a strong phenol-degrading activity in its native state, but we were unable toisolate any bacteria using our selection procedure. Since the procedure was designed for isolating aerobic bacteria, it is possible that in soil sample D2, the dominant phenol-degrading microorganisms are different, such as anaerobic bacteria, or fungi. The mixed culture of the strains had better phenol degradation than each individual bacterium, showing that the synergistic actions of the native organisms could be very important in the degrading capability. We have not investigated thoroughly this kind of synergy in this study, but it will be an important topic so that the best biodegradation reagent for phenol can be constructed from the soil samples.

Furthermore, identifying the best biological agent is required as well as identifying the optimal condition of the growth medium. In general, the presence of NaCl in the medium had a positive effect on the capability of phenol degradation. The bacteria grew better in MM with glucose (in the range of 0.25 – 0.75% w/v) than without glucose. pH of the medium could also influence the ability to degrade phenol of the strain D1.4.

With the results in our study, we have established the fundamental research and benefits of utilizing heavily dioxin-contaminated soil microorganisms for bioremediation of sites and the treatment of industrial wastewater contaminated with phenolic wastes.

REFERENCES [1]. Weber M, Weber M, Kleine-Boymann M, Phenol. Ullmann's

Encyclopedia of Industrial Chemistry, 2004. [2]. Ajaz M, Noor N, Rasool SA, Khan SA, “Phenol Resistant Bacteria

from Soil: Identification, Characterization and Genetical Studies”, Pakistan Journal of Botany, Vol 36, pp. 415 – 424, 2006.

[3]. Gayathri KV, Vasudevan N, “Enrichment of Phenol Degrading Moderately Halophilic Bacterial Consortium from Saline Environ-ment.”, J. Bioremed Biodegrad, , Vol 1, pp. 1-6, 2010.

[4]. Prpich GP, Daugulis AJ, “Enhanced biodecomposition of phenol by a microbial consortium in a solid – liquid two – phase partitioning bioreactor”, Biodecomposition, Vol 16, pp. 329 – 339, 2005.

[5]. Jame SA, Rashidul Alam AKM, Fakhruddin ANM, Alam MK, “Degradation of Phenol by Mixed Culture of Locally Isolated Pseudo-monas Species”, J. Bioremed Biodegrad Vol 1, pp. 1-4, 2010.

[6]. United States Environmental Protection Agency (US-EPA), “Codes of Federal Regulations”, 40 CFR 116.4, 40 CFR 372.65, 40 CFR 302.4 and 40 CER 355. http://www.epa.gov/epacfr40/chapt-1.info/chi-toc.htm, 2006

[7]. Chakraborty S, Bhattacharya T, Patel TN, Tiwari KK, “Biodegra-dation of phenol by native microorganisms isolated from coke processing wastewater”, Journal of Environmental Biology, Vol 31, pp. 293-296, 2010.

[8]. Jiang H, Fang Y, Fu Y, Guo QX, “Studies on the extraction of phenol in wastewater”, Journal of Hazardous Materials, Vol 101, pp. 179-190, 2003.

[9]. Nair CI, Jayachandran K, Shashidhar S, “Biodecomposition of Phenol”, African Journal of Biotechnology, Vol 7, pp. 4951-4958, 2008.

[10]. Agarry SE, Solomon BO (2008) Kinetics of batch microbial degrada-tion of phenols by indigenous Pseudomonas fluorescence. Interna-tional Journal of Environmental Science and Technology 5: 223-232.

[11]. Huu NB, Ha DTC, “Characterization of bacterial community structure in bioremediation treatment of herbicide/dioxin contami-nated soil at field trials”, Vietnam Journal of Biology, Vol 32, pp. 88-93, 2010.

[12]. Fortnagel P, Harms H, Wittich RM, Krohn S, Meyer H, “Metabolism of Dibenzofuran by Pseudomonas sp. Strain HH69 and the Mixed Culture HH27”, Applied and Environmental Microbiology Vol 56, pp. 1148-1156, 1990.

[13]. Brenner DJ, Krieg NR, Staley JT (Eds.), Bergey’s Manual of Systematic Bacteriology, Volume 2, 2nd ed., Springer, 2005.

[14]. Bedard DL, Haberl ML, May RJ, Brennan MJ, “Evidence for novel mechanisms of polychlorinated biphenyl metabolism in Alcaligenes eutrophus H850”, Applied and Environmental Microbiology Vol 53, pp. 1103-1112, 1987.

[15]. Layne E, “Spectrophotometric and Turbidimetric Methods for Measuring Proteins”, Methods in Enzymology, Vol 10, pp. 447 – 455, 1957.

[16]. APHA (1998) Standard Methods for the Examination of Water and Wastewater, 20th edn. American Public Health Association, Wash-ington, D.C.

[17]. Castillo MP, Stenström J, Ander P, “Determination of manganese peroxidase activity with 3-methyl-2-benzothiazolinone hydrazone and 3-(dimethylamino) benzoic acid”, Analytical Biochemistry Vol 218, pp. 99-404, 1994.

[18]. Benson HJ, Microbiological Applications: A Laboratory Manual in General Microbiology, Complete Version, 8th Edition, McGraw-Hill Science/Engineering/Math, USA, 2001.

[19]. Marrot B, Barrios-Martinez A, Moulin P, Roche N, “Biodegradation of high phenol concentration by activated sludge in an immersed membrane bioreactor”, Biochemical Engineering Journal Vol 30, pp. 174-183, 2006.

[20]. Stephen AB, Shelton DR, Berry D, Tiede JM, “Anaerobic Biode-composition of phenolic compounds in digested sludge”, Applied and Environmental Microbiology, Vol 46, pp. 50-54, 1983.

[21]. Paula M., van Schei PM, Young LY (1998) Isolation and Characterization of phenol degrading denitrifying bacteria. Applied and Environmental Microbiology 64: 2432 – 2438.

[22]. Adav SS, Chen MY, Lee DJ, Ren NQ, “Degradation of Phenol by Aerobic Granules and Isolated Yeast Candida tropicalis”, Biotech-nology and Bioengineering Vol 96, pp. 844-852, 2007

74

Biopolymer Film from Chitosan for Shelf-life Extension of Fruits

NGUYEN, Thi Xuan Lam; NGUYEN, Thi Minh Xuan; DANG, Duc Long

Biotechnology Department, Faculty of Chemistry Engineering Danang University of Technology, the University of Danang

Danang, Vietnam ntmxuan@dut.udn.vn; ddlong@dut.udn.vn

Abstract— After harvesting, fruits and vegetables deteriorate rapidly. The main causes of their deterioration are dehydration and attack of microorganisms. Therefore, the use of edible coatings on the surface, followed by a cold storage, will prolong shelf-life of fruits and vegetables. Chitosan, a derivation of deacetylated chitin, a high bioactive substance, becomes a promising alternative treatment to toxic chemical in preservation. However, the studies in Vietnam only focused on initial investigation of using chitosan film to extend storage life of these perishable commodities, but not factors boosting bioactivity of chitosan. Our research studies the effect of deacetylation degree, molecular weight of chitosan and soaking time in chitosan solution on mango preservation. Our results showed mangoes dipped into chitosan with low molecular weight at high degree of deacetylation for 10 minutes prolonged shelf –life greater than control without chitosan coating

Keywords-biopolymer film, chitosan, shelf-life extension, deacetylation degree, mango storage

I. INTRODUCTION Fruits play an important role in the Vietnam economy.

Revenues from exporting fruits during period 2004-2011, have increased growth rates that average of 20% each year and revenues reached USD 600 millions in 2011. Today, our fruits are sold in over 50 countries around the world. The largest markets are developed countries such as Japan, Germany, England, Canada, Singapore … In which, mangoes are one of the most cultivated fruit in the tropics, including Vietnam. FAO, 2009 predicted that the growth rate of mangoes will accelerate in the following years. Recent year, many potential markets have opened for importing our mangoes such as New-Zealand, South Korean and Spanish.

However, all fruits are still alive after harvesting, that means they continue to respire and transpire. This leads to lose water which usually causes shrinkage. Moreover, respiration induces transformation of organic matter to CO2 and water. Therefore, weight loss, color changes, softening and microbial spoilage occur during transportation and marketing. Treatment with chemical is usually used to extend shelf-life of fruits. However, nowadays, consumers around the world require high-quality food, without traditional preservatives, leading to increased effort in discovering new natural preservative.

Chitosan is a linear polysaccharide composed of randomly distributed β-(1-4)-linked D-glucosamine (deacetylated unit) and N-acetyl-D-glucosamine (acetylated unit). It is made by treating shrimp and other crustacean shells with the alkali sodium hydroxide. Chitosan became an ideal choice to alternate other chemical preservative because of its film-forming, biochemical properties which has led to prolonged storage life, and total natural product. Indeed, chitosan coating were demonstrated that they can decrease the respiration rate of fruits [13] and also delay the postharvest development of microbe causing diseases for fruits [10].

In Vietnam, there have been some initial achievements in using chitosan to extend storage life of fruits. Nguyen Van Toan et al. (2009), proposed a preserving-banana process which increased their shelf-life three times by chitosan. Tomatoes, grapefruits stored by films from chitosan have showed significant extension of their shelf-life compared to controls. However, these studies only investigated chitosan concentration, methods for making film (spraying or dipping) to raise storage capacity. While chitosan at higher degree of deacetylation and lower molecular weight have boosted the antimicrobial activity of chitosan rather than lower degree of deacetylation and higher molecular weight [1, 4, 11, 12, 14]. Our research considered these effect factors on bioactivity of chitosan in mango preservation.

II. MATERIALS AND METHODS Mangoes (Mangifera cambodiana) were harvested at Cam

Hai, Cam Ranh, Khanh Hoa province, Vietnam. After transported to our laboratory, their surfaces were pretreated immediately by CaCl2 and H2O2 before coating or non-coating with chitosan.

Chitosan used are a gift from Nha Trang University.

A. Determining weight loss Postharvest mangoes pretreated surfaces were numbered

and weighed. After 4, 8, 12, 14 storage days at room temperature, mangoes with or without chitosan-coating were weighed again. Weight loss was calculated by difference of weight of mangoes between before and after storage time.

75

B. Determining total sugar content Total reducing sugar content was determined by Bertrand

method. The general principle of this method is based on the oxidation reaction between sugar and metal ion under alkaline condition.

C. Determining vitamin C content Vitamin C content was measured based on ability of

vitamin C that can reduce iodine. Briefly, 5 g sample were grinded with sand in 5ml HCl 5% solution. Then, add distilled water to 50 ml. 20 ml of this solution was used to titrate by iodine with soluble starch as indicator.

III. RESULTS

A. Chemical composition of mangoes The alteration in compositions of mango fruits during

ripening in non-preservative condition was revealed in table I. These results were going to be used to compare with mangoes stored by chitosan films later.

Table I: Chemical compositions of green and ripe mango

Compositions Green mango Ripe mango

Water (%) 84.02 82.34

Dissolved solid content (Bx)

9.50 17.00

Reducing sugar content %

6.30 16.15

Total acidity % 2.59 0.17

Vitamin C (mg) 44.64 21.80

pH 2.36 5.02

B. The effect of chitosan with different degree of deacetylation on mango storability

The antimicrobial activity of chitosan was stronger at higher degree of deacetylation of chitosan [11, 12, 14]. Chitosan with a deacetylation degree of 92.5% was more effective than chitosan with deacetylation degree of 85% [15]. This may protect fruits from the attack of microbe and prevent deterioration of fruits by microbial spoilage. However there was also a report that the antimicrobial activity of chitosan did not depend on change in degree of deacetylation of chitosan (73, 84 and 95%) [7].

To clarify the effect of deacetylation degree of chitosan on shelf-life extension of mangoes, chitosan with degree of deacetylation of 93.32%, 85.8% with the same molecular weight of 200 kDa were used in our researches. Chitosan were prepared by dissolving 1g chitosan in 100 ml 1% acetic acid solvent in 2 hours with stirring until obtaining homogeneous solution. Mangoes pretreated with CaCl2 and H2O2 before coating or non-coating with chitosan. The day, when mangoes were dipped into prepared chitosan solution in 10 minutes and

stored at room temperature, was initial time of the process of preservation.

To assess storage ability of chitosan with different deacetylation degree, weight loss, sugar content and vitamin C content were examined after 4, 8, 12 and 14 days of treatment with or without chitosan. Chitosan with higher deacetylation degree prevented weight loss of postharvest mangoes better.

Weight loss is an important parameter to evaluate the quality of postharvest fruits and vegetables because of reduction of economic benefit and sensory value of products from this loss. This loss causes by evaporation and metabolism of organic matter in respiratory of fruits and vegetables.

The ability of chitosan at different degree of deacetylation in maintaining weight of mangoes was determined by weighing mangoes treated with chitosan solution after 4, 8, 12, and 14 days of treatment. Below figure has showed difference of mango weight loss between before and after dipped into chitosan solution.

Figure 1: The weight loss of mangoes after 4, 8, 12, and 14 days of treatment into chitosan solution with deacetylation degree of 93.32% (DH, blue line, diamond) or 85.8% (DL, red line, square); and mangoes without chitosan-coating (CT, green line, tangle) as control.

Figure 1 illustrated that the weight loss of mangoes coated chitosan film was significant less than the loss of mangoes non-coated. After 9 storage days, all mangoes non-treated with chitosan were rotten, while shelf-life of chitosan-coated mangoes prolonged over 14 days. Moreover, chitosan with various deacetylation degrees also conserved mango weight light differently. The more chitosan was deacetylated, the less weight loss during storing was. The hydrolysis of carbohydrate occurred slower in postharvest mangoes coated chitosan with more deacetylation.

During ripening, sugar content increase by hydrolysis of hemicelluloses, pectin and starch. Sugar content also indicates ripe degree of fruits. To prevent the loss of sugar, storage methods need to restrict this sugar produced.

Determination of reducing sugar content in 100g sample was carried out by Bertrand method. The method is based on the redox reaction of reducing sugars (glucose, fructose, maltose, etc.) with oxide copper II to generate oxide copper I under the alkaline condition. Oxide copper I, on which the

76

quantity of reducing sugar can be calculated, has red color like brick.

Figure 2: Sugar content in 100g of mangoes after 4, 8, 12, 14 days of treatment into chitosan solution with deacetylation degree of 93.32% (DH, blue line, diamond) or 85.8% (DL, red line, square); and mangoes without chitosan-coating (CT, green line, tangle) as control.

Sugar content of mangoes coated chitosan film increased markedly slower than sugar content in mangoes non-coated after 4, 8, 12, 14 days of treatment. After 14 storage days, sugar content of mangoes with chitosan film was equivalent to that of control at storage day of eighth. The similar to weight loss, the more chitosan was deacetylated, the slower hydrolysis was. Decline of vitamin C were impeded more effectively by chitosan-coating film with the higher degree of deacetylation.

Human beings cannot synthesize vitamin, a vital substance. Almost vitamins, including vitamin C, have been absorbed from fruits and vegetables. That is a reason why vitamin C content becomes an important index to indicate the quality of fruits.

Vitamin C content in mangoes treated or non-treated with deacetylated chitosan at different degree were determined by redox titration with iodine.

Figure 3: Vitamin C content (mg) of mangoes after 4, 8, 12, 14 days of treatment into chitosan solution with deacetylation degree of 93.32% (DH, blue line, diamond) or 85.8% (DL, red line, square); and mangoes without chitosan-coating (CT, green line, tangle) as control.

Vitamin C content lost during storage even mangoes treated or non-treated with chitosan solution. However, this loss was inhibited significantly by chitosan with higher deacetylation degree. Indeed, after 14 days of treatment,

vitamin C content of mangoes dipped into deacetylated chitosan of 93.32%, 85.8% were 33.12 mg, and 23.23 mg, respectively.

C. The effect of different molecular weight chitosan on mango storability

Recent researches have explored the effectiveness of different molecular weight chitosan in controlling fruit decay caused fungi as in vitro and in vivo [1, 4]. Coating fruits with chitosan of lower molecular weight was more effective in controlling the growth of fungi than chitosan of higher molecular weight. However, chitosan of low molecular weight was difficult to form film, and so hinder them from fruit preservation.

To determine appropriate molecular weight of chitosan for mango storage, chitosan of 200 kDa, 12.96 kDa, 6 kDa and mixture of 200 kDa and 6 kDa with ratio 1:1 were used in our researches. All chitosan with different molecular weight have the same degree of deacetylation and were prepared as above describe. The effect of different molecular weight chitosan on weight loss of postharvest mangoes.

Figure 4: The weight loss of mangoes after 4, 8, 12, 14 days of treatment into chitosan solution with different molecular weight of 200kDa (PH, blue line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL, green line, tangle);or into mixture of chitosan of 200kDa and 6kDa with ratio 1:1 (PHL, purple line, cross); and mangoes without chitosan-coating (CT, light blue line, asterisk ) as control

This result proved again that chitosan inhibited significantly weight loss of mangoes after harvesting. However, different molecular weight chitosan had different effect on this inhibition. Interestingly, the most success in maintaining weight of postharvest mangoes was mixture of chitosan of high and low molecular weight with ratio 1:1. Indeed, after 14 days of treatment, in 100 g of mangoes treated with chitosan of 12.96 kDa and 6 kDa were lost 5.47 g and 8.89 g, respectively, while mangoes treated with mixture were lost only 4.65 g. The effect of different molecular weight chitosan on sugar content of postharvest mangoes.

As the above explanation, reducing sugar content is an important indicator of ripe degree of postharvest fruits. The faster sugar content increase, the more difficultly fruits can be stored.

77

Figure 5: Sugar content in 100g of mangoes after 4, 8, 12, 14 days of treatment into chitosan solution with different molecular weight of 200kDa (PH, blue line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL, green line, tangle);or into mixture of chitosan of 200kDa and 6kDa with ratio 1:1 (PHL, purple line, cross); and mangoes without chitosan-coating (CT, light blue line, asterisk ) as control.

Figure 5 illustrated again that chitosan could prevent hydrolysis of carbohydrate in postharvest fruits. Chitosan of 12.96 kDa (PM) and 6 kDa (PL) showed stronger inhibition compared to chitosan of 200 kDa (PH). However, the mixture of high and low molecular weight chitosan was also the most effective during storing mangoes. The effect of different molecular weight chitosan on vitamin C content of postharvest mangoes

Figure 6: Vitamin C content (mg) of mangoes after 4, 8, 12, 14 days of treatment into chitosan solution with different molecular weight of 200kDa (PH, blue line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL, green line, tangle);or into mixture of chitosan of 200kDa and 6kDa with ratio 1:1 (PHL, purple line, cross); and mangoes without chitosan-coating (CT, light blue line, asterisk ) as control.

The results revealed that chitosan impeded the loss of vitamin C in mangoes during storing, but different molecular weight chitosan did not have significantly difference in conserving vitamin C content. Nevertheless, mixture of high and low molecular weight chitosan showed the best storability of vitamin C.

D. The effect of soaking time in chitosan solution on mango storability.

From the above results, low molecular weight chitosan was demonstrated more effective than high molecular weight chitosan in fruit preservation. However, with low molecular weight chitosan, fruits need soaking more time into chitosan solution to form film. This can damage chlorophyll in green mango peel. Consequently, sensory value of mangoes and storage life was shortened by this damage. Besides that,

different degree of deacetylation of chitosan can also affect generation of film.

To find out the appropriate soaking time, prepared mangoes were dipped into different degree of deacetylation of chitosan and different molecular weight chitosan solution for 2, 5, 10 minutes. Storage life of mangoes was from time mangoes dipped into chitosan solution until mangoes appeared phenomena such as dark yellow peel, wrinkled and soft fruits. The effect of time mangoes soaked into chitosan with different degree of deacetylation on shelf-life

Table II: Shelf-life of mangoes coated chitosan of different degree of deacetylation for different soaking time

Degree of deacetylation of chitosan

Soaking time (minutes) Shelf-life

93.32% 2 11 5 12

10 16

85.80% 2 12 5 12

10 13 Without chitosan 0 9

Storage time of mangoes dipped into chitosan with the same degree of deacetylation increased with soaking time. In all samples, shelf-life of mangoes soaked into chitosan of highest deacetylation degree of 93.32% for 10 minutes was longest. The effect of time mangoes soaked into chitosan with different molecular weight on shelf-life Table III: Shelf-life of mangoes coated chitosan at different molecular weight

for different soaking-time

Molecular weight of chitosan

Soaking time (minutes) Shelf-life

200 kDa 2 12 5 12

10 13

12.96 kDa 2 12 5 14

10 14

6 kDa 2 11 5 14

10 14

Mixture of chitosan of 200 and 6 kDa

2 14 5 16

10 18 Without chitosan 0 9 Storage time of mangoes dipped into chitosan with the

same molecular weight increased with soaking time. With chitosan at molecular weight of 12.96 and 6 kDa, shelf-life of mangoes did not show any difference between soaking time of 5 or 10 minutes. In all samples, shelf-life of mangoes soaked into mixture of chitosan of 200 and 6 kDa for 10 minutes was longest.

IV. DISCUSSION Our research demonstrated that mango storability by

chitosan with high degree of deacetylation were better than chitosan with low degree by maintaining weight, vitamin C

78

content, delaying transformation of carbohydrate to sugar, and extending storage-life of mangoes. The similar results have been documented in preserving fresh strawberries [5, 13]. Some reports suggested that effects of chitosan may be associated with direct anti-microbe properties against the postharvest pathogen [1, 2, 11]. The number of amino groups on chitosan increase with increased degree of deacetylation. As a result, chitosan with higher number of amino group dissolves completely, which increase chance of interaction between chitosan and microbial cell walls.

The molecular weight of chitosan was also showed effect on their antimicrobial activity in numerous reports [3, 6, 8, 9, 17]. Generally low molecular weight chitosan revealed greater effectively than high molecular weight. Indeed, coating fruits with chitosan of 15 kDa was more effective in controlling the growth of Penicillium digitatum and P. expansum causing fruit decay than chitosan of 357 kDa [4]. In addition, Jeon et al. [8] suggested that better antimicrobial activity of chitosan was at molecular weight more than 9 kDa. Indeed, chitosan of 9.3kDa inhibited effectively the growth of Escherichia coli, while that with molecular of 2.2 kDa enhanced growth of the same bacteria [16]. Our results displayed the same trend in mango storability and shelf-life extension. Chitosan films of 12.96 kDa presented a better effective in slowing down weight loss, production of reducing sugar, and increasing storage-life of mangoes than chitosan of 200 kDa and 6 kDa. However, the difference between 12.96 kDa and 6 kDa was not markedly, especially in maintaining vitamin C. Small molecule of chitosan may penetrate into microorganism more easily than large molecule. However, film coating on the surface of mangoes are difficult to be made from small molecules. To overcome this difficulty, we mixed chitosan of high and small molecular weight. This mixture had significantly effective on mango storage compared to individual kind of chitosan.

In conclusion, chitosan have greater effective in extending shelf-life and maintaining the quality of postharvest mangoes. However, the effect of chitosan at different degree of deacetylation and molecular weight on fruit preservation was different. Higher degree of deacetylation and lower molecular weight of chitosan have better effective. Interestingly, we find out that the mixture of chitosan at 200 kDa and 6 kDa with 93.32% degree deacetylation presented the most effective in fruit storability.

V. REFERENCES [1] Badawy, M.E.I. and E.I. Rabea, "Potential of the biopolymer chitosan with

different molecular weights to control postharvest gray mold of tomato fruit". Postharvest Biology and Technology, 2009. 51(1): p. 18.

[2] Badawy, M.E.I. and E.I. Rabea, "A Biopolymer Chitosan and Its Derivatives as Promising Antimicrobial Agents against Plant Pathogens and Their Applications in Crop Protection". International Journal of Carbohydrate Chemistry, Review article, 2011. 2011: p. 29.

[3] Chang, D.S., et al., "A development of food preservative with the waste of crab processing". Bulletin of the Korean Fisheries Society, 1989. 22: p. 9.

[4] Chien, P.J., F. Sheu, and H.R. Lin, "Coating citrus (Murcott tangor) fruit with low molecular weight chitosan increases postharvest quality and shelf life". Food Chemistry, 2007. 10(3): p. 5.

[5] Ghaouth, A.E., et al., "Chitosan coating effect on storability and quality of fresh strawberries". Journal of Food Science, 1991. 56: p. 4.

[6] Hirano, S. and N. Nagao, "Effects of chitosan, pectic acid, lysozyme, and chitinase on the growth of several phytopathogens". Agriculture and Biological Chemistry, 1989. 53: p. 2.

[7] Ikinci, G., S. Şenel, and H.A.e. al., "Effect of chitosan on a periodontal pathogen Porphyromonas gingivalis". International Journal of Pharmaceutics, 2002. 235(1-2): p. 7.

[8] Jeon, Y.J., P.J. Park, and S.K. Kim, "Antimicrobial effect of chitooligosaccharides produced by bioreactor". Carbohydrate Polymers, 2001. 44(1): p. 6.

[9] Kendra, D.F. and L.A. Hadwiger, "Characterization of the smallest chitosan oligomer that is maximally antifungal to Fusarium solani and elicits pisatin formation in Pisum sativum". Experimental Mycology, 1984. 8(3): p. 6.

[10] Li, H. and T. Yu, "Effect of chitosan on incidence of brown rot, quality and physiological attributes of postharvest peach fruit". Journal of the Science of Food and Agriculture, 2001. 81(2): p. 6.

[11] Liu, X.F., et al., "Antibacterial action of chitosan and carboxymethylated chitosan". Journal of Applied Polymer Science, 2001. 79(7): p. 12.

[12] Morimoto, M. and Y. Shigemasa, "Charaterization and bioactivities of chitin and chitosan regulated by their degree of deacetylation". Kobunshi Ronbunshu, 1997. 54(54): p. 11.

[13] Park, S.I., et al., "Antifungal coatings on fresh strawberries (Fragaria x ananassa) to control mold growth during cold storage". Journal of Food Science, 2005. 70(4): p. 6.

[14] Shimojoh, M., et al., "Bactericidal effects of chitosan from squid pens on oral streptococci". Nippon Nogeikagaku Kaishi, 1996. 70(7): p. 13.

[15] Simpson, B.K., et al., "Utilization of chitosan for preservation of raw shrimp (Pandalus borealis)". Food Biotechnology, 1997. 11(1): p. 20.

[16] Tokura, S., et al., "Induction of drug specific antibody and the controlled release of drug by 6-O-carboxymethyl-chitin". Journal of Controlled Release, 1994. 28(1): p. 7.

[17] Ueno, K., et al., "Antimicrobial activity by fractionated chitosan oligomers". Advances in Chitin Science, 1997. II: p. 6.

79

EFFECT OF CO2 UTILIZATION ON THE GROWTH OF CHLORELLA VULGARIS FOR

FOOD TECHNOLOGY Nguyen Hoang Minh, Nguyen Thi Thanh Xuan, Dang Kim Hoang, Nguyen Dinh Phu

Chemistry Faculty, Danang University of Technology, The University of Danang Danang city, Vietnam nhminh@dut.udn.vn

Abstract—CO2 fixation by Chlorella vulgaris cultivation is one of positive solutions which contribute to greenhouse effect reduction and high quality biomass production. C. vulgaris is a potential microalgae due to its great content of essential nutrition, faster growth, easier cultivation. More importantly, the microalgae are able to use high concentration of CO2, one of major contributor to greenhouse effect and global warming, as carbon source for its growth. Therefore, it is very necessary to study the effect of CO2 on growth of C. vulgaris. The results indicated that the supplying CO2 with 30 mL.min-1 of flow-rate in culture media could significantly improve biomass yield and amount of valuable nutrition (chlorophyll, protein, lipid). Optimization of growth conditions (CO2 flow-rate and light intensity) also was carried out. C. vulgaris is likely to prefer to grow in the media which is set up with 20 mL.min-1 CO2 flow-rate and 4000 Lux light intensity. Especially, isolated microalgal strains have morphological properties similar to original C. vulgaris.

Keywords- Chlorella vulgaris; CO2; food technology; microalgae; greenhouse effect; biomass.

I. INTRODUCTION Greenhouse effect causing from the increase in atmosphere

CO2 level, has become a serious worldwide problem [1, 2, 3]. According to scientific literature, the CO2 emission increased by 3% in 2011, reaching an all-time high of 34 billion tones [1]. More importantly, major reason for this CO2 increase is attributable to human activities which mostly relates to burning fossil fuels [2]. Therefore, to prevent severe damages due to excess greenhouse gases, human behavior should be seriously changed to mitigate CO2 emission in atmosphere.

Interestingly, many scientists have concentrated on converting CO2 into valuable productions through microalgae’s photosynthesis [4]. Among all microalgae, Chlorella vulgaris is the preferable object. C. vulgaris is a green spherical single celled fresh water microalgae [5, 6]. It is widely produced and marketed as a food supplement in many countries, including China, Japan, Europe and the US. Chlorella is being considered as a potential source of a wide spectrum of nutrients (protein, lipid, carotenoid, vitamins, minerals) being effectively applied in the healthy food market as well as for animal feed and aquaculture [7]. Many researches also proved that Chlorella is important as a health promoting factor on many kinds of disorders such as gastric ulcers, wounds,

constipation… [8]. Moreover, Chlorella has fast growth rate, and withstand high carbon dioxide levels in simple growth culture. Therefore, C. vulgaris has attracted an increasing amount of attention.

On the other hands, Chlorella’s growth depends on culture conditions (nutrients, CO2, light intensity…). Many studies showed that Chlorella grew better in elevated CO2 aerating culture [6, 9]. However, each microalgae requires certain CO2 concentration for highest biomass production [9]. Therefore, the aim of the research is to determine effect of CO2 utilization on the growth of C. vulgaris.

In this study, C. vulgaris strain is cultured in appropriate media. Then, microalgae’s growth was investigated in cultures containing different CO2 concentrations. After harvesting microalgae with highest productivity, concentrations of chlorophyll a, protein, lipid in Chlorella were measured and evaluated. Optimization of CO2 concentration and light intensity was performed in this study.

II. MATERIALS AND METHODS

A. Microalgae source The microalgal strain used in this study is Chlorella

vulgaris, which was bought from Hanoi Institute of Biotechnology.

B. Investigation of algal growth The seed culture was grown in 1000 ml Erlenmeyer flasks

containing 500 ml of Antoine medium and subcultured by transferring 10% (v/v) inoculum at 5-day intervals. The experiments were exposed to the 20W fluorescent lamps daylight in 12:12 circadian cycles at room temperature. In order to investigate the role of CO2 on Chlorella’s growth, growth culture was performed in 3 different cases: without aerating used as control; with aerating under air flow rate of 650mL.min-1, corresponding to ambient CO2; and with CO2-air mixture (30 mL.min-1 pure CO2 gas flow-rate mixed with 650 mL.min-1 air flow-rate, corresponding to 5% CO2).

Microalgae’s growth was tracked by measuring optical density (OD) of the culture with a spectrophotometer at 420 nm. Algal growth curve was built up with the OD values recorded at the same time every day.

80

C. Evaluation of algal quality 1) Measuring of chlorophyll-a concentration by Avigad

Vonshak method [10] Measuring chlorophyll-a concentration is carried out by

using spectrophotometer, following 3 steps: (1) separation of microalgae cells from the medium by centrifugation (15000 rpm at 4OC for 5 min); (2) extraction of pigments with acetone and (3) spectrophotometric determination of the concentration of Chlorophyll-a in the extract at 664 nm and 647 nm wavelength. Chlorophyll a was determined using the following formula:

Chlorophyll a (µg/ml) = 11.93. E664 – 1.93.E647

2) Measuring of protein concentration by Bradford assay [11]

After 14 days, microalgae biomass was collected by centrifugation (Hettich Zentrifugen centrifuge) at the speed of 6000 rpm for 10 min. In this experiment, ultra-sonication was used to release protein fractions from the whole cell. Then, protein extraction obtained by centrifugation was measured by Bradford method [8]. In briefly, 10 µL of protein extract was mixed with 90 µL of the diluted reagent. The blue mixture’s absorbance at wavelength 595nm was measured by spectrophotometer. As the result, protein concentration was determined based on standard curve plotted from concentration series (0.1-1.0 mg.mL-1).

3) Measuring of lipid concentration [12] The lipid fractions were extracted using the method which

was considered the most effective extraction method by Attilio. Following the previous method, the dried microalgae biomass samples were pulverized and extracted using mixture of chloroform: methanol (2:1, v/v). About 50 ml of solvents were used for every gram of dried sample in each extraction step. After stirring by ultra - sonication at 20 kHz frequency for 10 min, distilled water was added to the samples which were then centrifuged at 6000 rpm for 10 min. The lipid soluble in the chloroform was removed by using micropipette.

D. Optimization of growth conditions [13] 1) Optimization of CO2 flow rate

The effect of flow rate of CO2 on biomass production was investigated by varying CO2 flow-rates: 20, 40 and 60 mL. min-1, corresponding to 3, 6, and 9.2% CO2, respectively. The experiments were performed in the same conditions: the temperature of 27OC and light intensity of 7000Lux.

2) Optimization of light intensity The effect of light intensity on biomass productivity was

investigated by varying the light intensity with the values of 4000, 7000 and 13000 Lux. The luminous flux was measured by Advance Light Meter.

E. Microalgal isolation Water samples were collected aseptically from two Ham

Nghi and 29-3 lakes at Danang city. The samples were transferred to 10 tubes containing Antoine nutrient medium and incubated at 25oC under 4,000 Lux illumination and 12h light/dark photoperiods. After 10 days, 100 µL of samples was

inoculated onto petri plates containing Antoine nutrient medium solidified with 2.0% of bacteriological agar. The incubation was performed under the same condition as the tubes for 15 days. Then, colonies from the petri plates were mixed into distilled water contained in test tube. Finally, alga’s morphology was observed by using microscope with a magnification of 40x.

III. RESULTS AND DISCUSSION

A. The role of CO2 on the growth of Chlorella vulgaris Generally, phototrophic microalgal growth requires a

supply of carbon dioxide as a carbon source [3]. Microalgae grown in elevated CO2 environments typically exhibit increased rates of photosynthesis and biomass production. In this experiment (Fig. 1), without aeration, microalgae’s growth rate was likely to be very weak. With aeration, Chlorella cells slowly developed during 11 days batch cultivation. Since twelfth day, they quickly developed and achieved the maximal value after 20 days. Interestingly, C. vulgaris’s growth rate in 5% (v/v) CO2 aerated culture was exponentially elevated after first 5 days, and reached to value similar to those obtained under ambient CO2 after 12 days. The results implied that Chlorella’s growth rate depends on CO2 concentration in the culturing medium. C. vulgaris exhibited an increase in performance under elevated CO2 condition. Senthil also cultured C. vulgaris in medium supplying with varying concentration of CO2. Data revealed that the highest chlorophyll and biomass, which were 60 and 20 times more than that of C. vulgaris at ambient CO2, were recorded at 6% CO2 level [9]. Such CO2 fixation by photoautotrophic algal culture has the great potential to not only produce high biomass productivity, but also diminish the release of CO2 into the atmosphere, helping alleviate the trend toward global warming.

Figure 1. C. vulgaris’s growth at 3 different cases: without air, with air and with CO2-air mixture

Besides, in order to produce a high valuable microalgae biomass production serving for food technology, CO2 concentration in the culture should be controlled seriously. Too high CO2 concentration could produces inhibition and, on the other hand, too low concentration could limits growth [6]. These thresholds vary from one species to another and are not yet adequately known. Therefore, it is necessary to investigate

81

the role of CO2 and optimize CO2 concentration in C. vulgaris’s growth.

B. Evaluation of algal quality 1) Measuring of chlorophyll-a concentration

Chlorophyll- a is a primary photosynthetic pigment in all algae, in which Chlorella sp. is considered as a microalgae strain containing highest Chlorophyll amount. This is an indicator exhibiting microalgae’s growth. In food technology, chlorophyll- a is used as food and pharmaceutical colorants. It also can exhibit health promoting activities. Chlorophyll has wound healing and tissue regeneration properties and acts like a natural antibiotic [7]. It also plays a significant role in cancer prevention. For these reason, it is necessary to evaluate influence of CO2 in Chlorophyll- a formation of C. vulgaris. After 7 days, chlorophyll- a obtained from 5% (v/v) CO2 aerated culture (0.93 µg.mL-1) was significantly higher than that under ambient CO2 and control (0.20- 0.26 µg.mL-1) (Table I). The result indicated that ambient CO2 was not enough to support growth of C. vulgaris. Higher CO2 recorded vigorous growth. It implied that supply of elevated CO2 helped microalgae form more chloroplasts to effectively convert and use light energy.

2) Measuring of protein concentration In order to obtain valuable algal biomass applied for food

technology, protein concentration should consist of high ratio in alga. According to Dang Dinh Kim et al. [14], protein content of C. vulgaris is high, about 40 - 60 %. Razif Harun reported that the protein concentration is about 51 – 58%, compared to dried weight [15]. On the other hand, Hee sun lee et al assumed that highest protein amount of Chlorella obtained is 60.6% [16].

In this study, protein obtained under elevated CO2 was quite high (53%) and similar to Razif Harun’s results. Under ambient CO2, protein content was 31% only (table I). This reveals that the increase in the microalgal performance at higher levels of CO2 may be attributable to the enhanced availability of dissolved CO2.

3) Measuring of lipid concentration Apart from protein, lipid concentration is also important

target which should be evaluated. Previous studies showed that polyunsaturated fatty acids of algae could offer many health benefits, such as hypercholesterolemia, hyperlipidaemia and atherosclerosis [5].

C. vulgaris cultured under 5% CO2 showed significantly higher amount of lipid value (13.25 %), compared to other culture conditions (4- 9 %). This was similar to finding of Dang Dinh Kim who found that the lipid content of Chlorella was about 10-15%.

On the other hand, the collected lipid from CO2 bubbled sample was analyzed by HPLC. The result showed that the peak of triglyceride appeared at approximately 10.42 minutes retention time and the triglyceride in the mixture presents in quite high concentration (data not shown). Moreover, the triglyceride’s composition analysis results taken by GC-MS revealed that there are 2 typical peaks of methylester of fatty

acids – palmitic acid (C16:0) and oleic acid (C18:1). These 2 types of fatty acids are valuable for pharmaceutical purpose. TABLE I: Chlorophyll-a, protein, lipid concentration of C. vulgaris were

measured in the 3 different aerated cultures.

In conclusion, C. vulgaris grew well under 5% CO2 with

optimum nutritional compositions. Chlorophyll-a, protein, lipid content were dependent on the amount of CO2 in the culture, meaning that it is possible to modify the nutritional contents of Chlorella by changing culture conditions.

C. Optimization of growth conditions Previous studies indicated that it is very necessary to

increase both CO2 concentration and light intensity to obtain high biomass production [17]. In this study, the highest dry weight of C. vulgaris was obtained under growth culture supplied with 3% of CO2 and 4000 Lux of light intensity. Algal biomass decreased when CO2 and light intensity were higher than optimal values.

The growth response of C. vulgaris was studied under varying concentrations of carbon dioxide (3, 6, 9 %). The result of table II indicated that at higher CO2 concentration, algal’s biomass began to slowly reduce, but they continued to remain greater than those observed in ambient air (0.52 mg.mL-1). This is in agreement with Senthil Chinnasamy et al. [9], in which they explained that increase in the algal performance at higher level of CO2 may be because of the enhanced availability of dissolved CO2 and the toxicity of a very high CO2 concentration is due to the lowering of pH.

TABLE II: Effect of CO2 concentration on biomass production of C. vulgaris

Effect of light intensity on algal growth was showed in

table III. When using higher light intensity (7000 Lux), Chlorella’s dried biomass weight slightly reduced. However, 13,000 Lux light gave strong inhibition on algal growth. The result of light intensity in this study (table III) seems to be contradictory with that of Siranee et al. who assumed that algal culture at 8000 Lux gave significantly higher amount of algal biomass, compared to those under 3000 and 5000 Lux [18]. Moreover, according to findings of Anondho Wijanarko et al. [4], illumination using UV light indicates a decrease in cell amount. This susceptibility to photo inhibition is caused by the blockage of chloroplast-encoded protein synthesis, and a severe damage occurs in photo-system II reaction center of micro

82

algae chlorophyll a. This implied that using very high light intensity (13000 Lux) in this experiment could cause phenomenon of photo inhibition. TABLE III: Effect of light intensity on biomass production of C. vulgaris.

D. Microalgae isolation at Danang city Microalgae isolation from natural rivers at Danang city

plays an important role in discovering and possessing novel microalgae strains which are able to produce valuable biomass and tolerate to high CO2 concentration. Many reports showed the possibility of the isolation of such naturally occurring algal forms [9]. Some microalgae could quickly grow under 40% of CO2. Komada et al. reported that Chlorella littorale developed better at 30oC and 20% CO2 concentration and another marine alga could perform fast even at 60% CO2 [19]. This shows the changes of using such organisms to produce high-quality biomass.

In this experiment, natural strains isolated from Ham Nghi and 29-3 lakes have similar properties with original Chlorella vulgaris strain’s morphology, such as green color, spherical. Since we do not know whether they are Chlorella vulgaris or not, further study will focus on identifying isolated strains’ genotype.

Figure 2. Microagal morphologies were observed under microscope with a 40x magnification

IV. CONCLUSION In this study, Chlorella vulgaris could grow better under

elevated CO2. The highest chlorophyll-a, protein, lipid concentration were recorded under 5% of CO2, compared to ambient CO2 (0.036%). The highest dry weight of C. vulgaris was obtained under growth culture supplied with 3% of CO2 and 4000 Lux of light intensity. Furthermore, the first result of microalgae isolation from natural lakes at Danang city contributes to creating bank of microalgal strains which could not only produce high value molecules but also mitigate CO2 emission in atmosphere. More importantly, this study is the central foundation for carrying out further studies, such as investigation of microalgae harvesting, establishment of model using CO2 flue gas from industrial parks for large-scale Chlorella vulgaris cultivation at Danang city.

REFERENCES [1] Jos G. J. Olivier, Greet Janssens-Maenhout, Jeroen A. H. Peters. 2012.

Trends in global CO2 emissions: 2012 report. PBL Nethrlands Environmental Assessment Agency.

[2] Nhan T. Nguyen, Minh Ha- Duong, “The potential for mitigation of CO2 emissions in Vietnam’s Power sector”. Depocen. Working paper series No. 2009/22.

[3] Nguyen Thai Hoa, Kei Gomi and Yuzuru Matsuoka. “A scenario for sustainable low-carbon development in Vietnam towards 2030”

[4] Anondho W., Dianursanti. 2004. “Effect of photoperiodicity on CO2 fixation by Chlorella vulgaris Buitenzorg in bubble column photobioreactor for food supplement production”. Makara technology, vol 8, No. 2 35-43

[5] Becke W. 2004. “Microalgae aquaculture in Handbook of Microalgae culture”. Richmond A. p: 380-392.

[6] Fadhil M. Salih. 2011. “Microalgae tolerance to high concentrations of carbon dioxide”: A review. Journal of Environmental Protection, 2, 648-654.

[7] Gouveia. L A. P. Batista, Souse I. 2008. “Microalgae in Novel Food Products. Food Chemistry research development”. Nova Science Publisher. Chapter 2.pp:1-37.

[8] Gouveia L. , B. P. Nobre, F. M. Marcelo, S. Mrejen, M. T. Cardoso, A. F. Palavra, R. L. Mendes. 2007. Functional food oil coloured by pigments extracted from microalgae with supercritical CO2. Elsevier, Food Chemistry 101:717-723.

[9] Senthil Chinnasamy, Balasubramanian Ramakrishnan and Keshav C. Das. 2009. “Biomass Production Potential of a Wastewater Alga Chlorella vulgaris ARC 1 under Elevated Levels of CO2 and Temperature”. Internation Journal of Molecular Sciences. Int. J. Mol. Sci. 2009, 10, 518-532

[10] Vonshak A. (1997), “Preface in: spirulina plantesis: physiology, cell-biology and biotechnology”.

[11] Bradford, M. M. 1976. “A rapid and Sensitive-Method for the Quantitation of Laemmli”, U. K.1970. Nature. Vol 227. pp: 680-685.

[12] Attilio Converti. 2009. “Effect of temperature and nitrogen concentration on the growth and lipid content of Mannochloropsis oculata and Chlorella vulgatis for biodiesel production”. Chemical Engineering and Processing; Process Intensification, Elsevier, 48, 1148-1151

[13] Nguyen T. T. Xuan, Dang K. H. , Nguyen M. T., Le T. B. Y. , Nguyen H. M. .2012, “A study of Chlorella vulgaris microalgae cultivatoin for biodiesel production”. Science and Technology. The University of Danang. Vol. 8[57].

[14] Đặng Đình Kim. 1999. “Công nghệ sinh học vi tảo”, NXB Nông nghiệp Hà Nội.

[15] H. Razif, S. Manjinder. 2010. “Bioprogress engineering of microalgae to produce a variety of consumer products”, J. Elsevier, Renewable and sustainable Energy reviews 14:1037-1047

[16] Hee Sun Lee, Hoon Jung Park and Mi Kyung Kim. 2008. “Effect of Chlorella vulgaris on lipid metabolism in Wistar rats fed high fat diet”. The Korean Nutrition Society. 204-210.

[17] D. Sasi, E.E. Powel and G.A. Hill. 2009. “Effect of light intensity and CO2 on growth of Chlorella vulgaris”. University of Saskatchewan, Saskatoon, Saskatchewan, S7N 5A9m Canada.

[18] Siranee S. , Preeda P. 2007. “Nutrient recycling by Chlorella vulgaris from septage effluent of the Bangkok city”, Thailand. ScienceAsia 33: 293-299

[19] Komada, M. , Ikemoto, H. 1993. “A new species of highly CO2 – tolerant fast growing marine microalga suitable for high density culture”. J. Mar. Biotechnol. 1. 21-25

83

EFFECT OF CARBON SOURCES ON PROLIFERRATION OF ZEDOARY (CURCUMA

ZEDOARIA ROSCOE) CELL SUSPENSION CULTURES

Vo Chau Tuan, Tran Quang Dan

Faculty of Biology and Environmental Science College of Education, Da Nang University,

Da Nang, Vietnam Email: vochautuan@gmail.com

Abstract—Zedoary (Curcuma zedoaria Roscoe) plant, a vegetatively propagated species of the Zingiberaceae family, is an aromatic herbaceous plant with a rhizome growing mainly in China, Vietnam, India, Japan. Cell suspension culture is the prior choice in large-scale production of some valuable secondary metabolites thanks to its great advantages over other technologies. We report here the effect of different sugars on the growth of cells in a shake liquid culture condition. As a single carbohydrate source in medium, sucrose exhibited a better effect on the cellular growth when compared with fructose or glucose. Zedoary cells reached a maximum biomass of 10.44 g fresh cell (approximately 0.66 g of dry cell) after 14 days of culture at sucrose concentration of 30 g/L. The combination of two hexoses (glucose and fructose) at different concentrations were not effective for proliferation of cell suspensions. Results from this study might be a foundation for further studies on Curcuma zedoaria Roscoe in order to produce its secondary metabolites in a large scale.

Keywords- Callus, cell biomass, cell suspension, Curcuma zedoaria, medicinal plant

I. INTRODUCTION Zedoary (Curcuma zedoaria Roscoe) plant, a vegetatively

propagated species of the Zingiberaceae family, is an aromatic herbaceous plant with a rhizome growing mainly in China, Vietnam, India, and Japan [7]. Zedoary is a valuable medicinal plant; the essential oil obtained from rhizome has been reported to have antimicrobial activity and be clinically used in the treatment of cervical cancer; the water extract of zedoary demonstrated antimutagenic activity [24]. It has been also used for treatment of stomach diseases, hepatoprotection [22], treatment of blood stagnation, and promoting menstruation as a traditional medicine in Asia [15]. Furthermore, the zedoary has anti-inflammatory potency related to its antioxidant effects [22].

The presence of valuable metabolites in plants has stimulated interests of the industries in the fields of pharmaceuticals, agrochemicals, nutrition and cosmetics. The bulk of the market products, such as secondary metabolites from higher plants, are collected from plants growing in the

wild or from field cultivated sources [9]. In recent decades, interests in chemopreventive products from natural plants has grown rapidly [14]. However, there are many challenges in the production of plant-based medicines, many of which put both the consumers and the plant populations at risk [2]. Therefore, an alternative, economically viable, and environmentally sustainable production source of desired secondary metabolites is of a great interest. In this regard, plant cell cultures can be an attractive alternative as a production system, as well as a model system, to study the regulation of natural product biosynthesis in plants so as to ultimately increase yields [9].

Sugar plays a central role in plant life. They enter the metabolism pathways and the transformation of energy, which are required for growth of cell. In plant tissue cultures, sugar serves as a carbohydrate supply to provide an optimal culture condition for cells. For most of plant cell lines, sucrose is an important carbon and energy source, and its initial concentration can affect such parameters as growth rate and the yield of secondary metabolites [11]. Utilization of monosaccharides by plant cells has been extensively investigated. Many types of plant cells and tissue cultures commonly utilize sucrose, glucose, fructose and some other monosaccharides equally when they are added to the culture medium as a sole carbon source [8]. Even though carbohydrates are of prime importance for cell growth, maintenance and differentiation in vitro, the fundamental aspects of carbon utilization and metabolism in cell and tissue cultures have yet to be fully understood [3].

In this study, we investigated the effect of three kinds of sugar (sucrose, glucose, fructose) in separation or in combinations on the cell growth of Curcuma zedoaria cell suspension culture for improvement of culture medium and further aims for secondary metabolites production in large-scale.

84

II. MATERIALS AND METHODS

A. Callus formation Leaf-base explants of 0.5×0.5 cm were excised from

zedoary plants in vitro [13]. The explants were placed on the Murashigie and Skoog (MS) solid medium supplemented with 2% (w/v) sucrose, 0.25-4.0 mg/L 2,4-dichorophenoxyacetic acid (2,4-D), and 0.25-4.0 mg/L benzyladenine (BA) for callus induction. White and soft calli (primary calli) were transferred were transferred on the MS medium supplemented with 0.5-4.0 mg/L 2,4-D, and 0.5-4.0 mg/L BA for secondary calli induction. The secondary calli which were light yellow, hard and friable inducted on the MS medium supplemented with 0.5 mg/L 2,4-2,4-D and 0.5 mg/L BA after for weeks of culture were maintained in fresh medium with the same composition every two weeks. The pH of the medium was adjusted to 5.8, and then it was autoclaved at 121oC for 15 min. The cultures were incubated at 25±2oC under an intensity of 2,000-3,000 lux with a photoperiod of 10-h day light.

B. Initiation of cell suspension cultures and test carbon sources Cell suspension cultures were initiated from callus which

were light yellow, hard and friable (Fig 1A) by culturing callus (3 g) in 250 mL Erlenmeyer flasks containing 50 mL of liquid medium with the same composition as the callus inducing medium under a rotary shaker at 120 rpm, for 10 days until a suspension of free cells formed. From the resulting cell suspensions, 3 g of cells was transferred on the MS medium supplemented with 1.5 mg/L 2,4-D and 0.5 mg/L BA but varied carbon concentration and sources: sucrose, glucose, fructose (20-70 g/ L) or combination of glucose and fructose rates (20/40; 30/30 and 40/20 g/L) for biomass production. The flasks were placed on a rotary shaker (Fig 1B ) at 120 rpm under the same physical conditions as described for callus cultures except an intensity of 500 lux.

C. Growth measurement Samples were obtained after 14 days of culture to

determine the cell biomass in both fresh and dry weights. For measurement of fresh cell weight (Fig 2A), the cells in the suspension culture were filtered, washed with distilled water, collected, and weighed. The dry cell weight was determined by drying the fresh cell biomass at 50oC until a constant weight was attained (Fig 2B).

D. Statistical analysis The experiments of cell suspension culture were conducted

with a minimum of three replicates. All experiments were repeated three times. The data were analyzed by mean ± standard error followed by comparison of the means with the Duncan’s test at p<0.05.

III. RESULTS AND DISCUSSION

A. Effect of sucrose on cell growth As shown in Table I the initial concentration of sucrose

significantly affect the biomass accumulation of the cell

culture, highest fresh cell weight were attained in media containing sucrose concentration of 30 g/L, the cell biomass reached with 10.44 g of fresh cell weight (approximately 0.66 g dry cell weight). However, despite sucrose being an indisputably important carbon and energy source, increasing its concentration from 40 to 60 g/L resulted in fresh cell weight reductions and significantly reduced at sucrose concentration of 70 g/L. This decline in cell growth might be attributed to the inhibition of nutrient uptake as the osmotic potential was enhanced and the medium became more viscous. Do et al. (1991) have shown that in Vitis vinifera a higher concentration of sucrose can act as an osmotic agent, with mannitol having a similar effect on growth [4]. Additionally, this retardation in growth could be caused by a cessation in the cell cycle when nutrients are limited and sucrose concentrations are higher [6, 26].

TABLE I. EFFECT OF DIFFERENT INITIAL SUCROSE CONCENTRATIONS ON THE GROWTH OF

ZEDOARY CELL SUSPENSIONS

Sucrose concentrations (g/L)

Cell biomass (g)

fresh weight dry weight

20 7.22c 0.55d

30 10.44a 0.66b

40 8.85b 0.64b

50 8.80b 0.65b

60 8.75b 0.70a

70 6.75d 0.60d

Different letters indicate significantly different means using Duncan’s test (p<0.05)

Even though at high concentrations of sucrose (40-60 g/L), the fresh cell biomass obtained was lower than compared with the response to 30 g/L sucrose, but cells tend to increase the dry biomass accumulation. At 60 g/L sucrose concentration, fresh weight of cells reached only 8.75 g, lower than on the medium with 30 g/L sucrose (reaching 10.44 g fresh cell), but the dry weight was higher (0.70 g dry cell compared to 0.66 g dry cell). These results are in agreement with some previous reports for other plant species such as Solanum chrysotrichum [25], Centella asiatica [20]. The role of sucrose in promoting dry cell weight can be explained by its effect on tubulin, a ubiquitous protein responsible for growth and development of cells. Tubulin controls cell shape and chromosome separation and provides cytoskeletal tracks for intracellular vesicular transport via Tual and Incw1 genes. In the absence of sucrose, these genes are not expressed but addition of sucrose turns on their expression [20].

B. Effect of glucose on cell growth Results presented in the Table II shows that the change of

initial glucose concentrations had also affected on cell growth. When the sugar concentration is in the range of 20-30 g/L, the biomass accumulation of cells is relatively low. Increasing the

85

concentration of sugar to the range of 40 to 60 g/L, cell biomass accumulation increased, the cell biomass reached maximum value with 7.32 g of fresh cell weight (approximately 0.65 g dry cell weight) at sugar concentration. of 60 g/L after 14 days of culture. The growth of cells reduced with increasing sugar concentration up to 70 g/L. Generally, at different sugar concentrations, cell growth in medium containing glucose was remarkably lower than that in medium supplemented with sucrose (Table I).

TABLE II. EFFECT OF DIFFERENT INITIAL GLUCOSE CONCENTRATIONS ON THE GROWTH

OF ZEDOARY CELL SUSPENSIONS Glucose concentrations

(g/L) Cell biomass (g)

fresh weight dry weight 20 3.15e 0,22d

30 6.05d 0,42c

40 7.12c 0,53b

50 7.20b 0,62a

60 7.32a 0,65a

70 6.85c 0,57e

Different letters indicate significantly different means using

Duncan’s test (p<0.05)

The carbon source in plant cell suspension media are usually supplied as carbohydrates, with the most common sugars being sucrose and glucose. Several previous reports showed that glucose was suitable carbon source in cell and tissue cultures [8, 23, 10]. Our study showed that glucose was less effective for proliferation of zedoary cells compared with sucrose. Similar results were reported by Ling et al (2008) working with Ficus deltoidea [12] and Shinde et al (2009) with Psoralea corylifolia [21].

C. Effect of fructose on cell growth The effects of different fructose concentrations of 20,

30, 40, 50, 60 and 70 g/L on the growth of zedoary cell were shown in Table III. The initial concentration of fructose was influence on the accumulation of cell biomass. The rise of the fructose concentration from 20 to 60 g/L increased the cell biomasses for both fresh and dry weight. Zedoary cells reached a maximal biomass of 8.40 g fresh cell (approximately 0.62 g dry cell) after 14 days of culture at the initial fructose concentration of 60 g/L. However, as the concentration of fructose increased to 70g/L, the growth of cells was inhibited (only reached 7.65 g fresh cell and 0.57 g dry cell).

The absorption lines of plant cells in culture depends on the sensitivity of each species of plants to the carbon source or on differences in the formation of products of metabolic processes [16]. Soluble sugars such as glucose, fructose and sucrose are often referred to as the substances responsible for osmotic adjustment in tissues under osmotic stress and the best carbon sources for the growth of most plant cell cultures [12]. Abdullah et al. (1998) showed that 3% and 5% of fructose and mannitol were able to promote growth of Morinda elliptica

cell cultures [1]. Ling et al. (2008) working with Ficus deltoidea found that fructose treatment gave the highest growth increase followed by sucrose, sorbitol and glucose, respectively [12]. The kinetics of sugar uptake by maize endosperm suspension cultures resembled those observed for asparagus cell cultures, in that fructose was transported most rapidly, followed by glucose and then sucrose [5]. The consumption of fructose by Solanum. eleagnifolium suspension culture was found to be 27% higher than sucrose [19]. Although fructose managed to produce the highest growth rate, it was anticipated that at higher concentration, it might have a growth-inhibitory effect. Such inhibitory effect at higher concentration of fructose was well document in Nicotiana tabacum and Cinchona succirubrum with using high fructose concentration [19].

TABLE III. EFFECT OF DIFFERENT INITIAL FRUCTOSE CONCENTRATIONS ON THE GROWTH

OF ZEDOARY CELL SUSPENSIONS Glucose

concentrations (g/L)

Cell biomass (g)

fresh weight dry weight

20 3.15e 0,22d

30 6.05d 0,42c

40 7.12c 0,53b

50 7.20b 0,62a

60 7.32a 0,65a

70 6.85c 0,57e

Different letters indicate significantly different means using Duncan’s test (p<0.05)

D. Effect of glucose and fructose in combinations of different concentrations on cell growth Table IV indicated that the combinations of glucose and

fructose at different concentrations have no effect to promote on the cell growth. The cell biomass reached a highest value with 6.95 g of fresh cell (approximately 0.65 g dry cell) in media containing of 40 g/L fructose and 20 g/L glucose. Generally, cell biomass was lower than cells cultured on the medium containing sole sugar after 14 days of culture.

TABLE IV. EFFECT OF GLUCOSE AND FRUCTOSE IN COMBINATIONS OF CONCENTRATION ON

GROWTH OF ZEDOARY CELL SUSPENSIONS

Glucose and fructose rates (g/L)

Cell biomass (g)

Glucose Fructose fresh weight dry weight

40 20 6.60b 0.60b

30 30 6.75b 0.61b

20 40 6.95a 0.65a

Different letters indicate significantly different means using Duncan’s test (p<0.05)

86

Figure 1. Callus and suspension cell of zedoary. A Light yellow, compact

and friable callus, B Suspension cell

Figure 2. Cell biomass of zedoary. A: Fresh cell biomass, B: Dry cell biomass

According to Duong Tan Nhut et al. (2006), the combination of two hexoses (glucose and fructose) at different concentrations, the best proliferation of cell was obtained at the combination of 30 g/L glucose and 30 g/L fructose, higher than compared with single sugar [18]. Results of our study is completely opposite. Thus, to achieve the optimum cell growth, different plants required different carbon sources due to the different enzymatic metabolism. Moreover, the availability of sugars and its derivatives would initiate different responses and would affect plant metabolism, growth and development [12].

IV. CONCLUSION The suspension culture offers many advantages to

scale-up production of secondary metabolites in plant cells of interest. In this study, sucrose exhibited a better growth of cell when compared with fructose or glucose. The combination of two hexoses (glucose and fructose) at different concentrations were not effective for proliferation of cell suspensions. Results from this study might be a well established foundation for further studies on Curcuma zedoaria Roscoe in order to serve as a potential source for secondary metabolites production in large scale.

ACKNOWLEDGMENT This study was supported by a grant from the Basic

Research Program in Natural Science of the Vietnamese Ministry of Science and Technology (2006-2008).

REFERENCES [1] Abdullah MA, Lajis NH, Ali AM, Marziah M, Sinskey AJ, Rha CK

(2005). Issues in plant cell culture engineering for enhancement of productivity. Development of chemical engineering process 13: 1-15.

[2] Cole JB, Saxena PK, Murch SJ (2007). Medicinal biotechnology in the genus scutellari. In Vitro Cell Dev Biol-plant 43: 318-327

[3] De Mello MO, De Campos AF, Murilom (2001). Sucrose metabolizing enzymes in cell suspension cultures of Bauhinia forficata, Curcuma zedoaria and Phaseolus vulgaris. Pesq. agropec. Bras 36(9): 1085-1092.

[4] Do CB, Cormier F (1991). Effects of low nitrate and high sugar concentrations on anthocyanin content and composition of grape (Vitis vinifera L) cell suspension. Plant Cell Rep 9: 500-504.

[5] Felker FC, Miernyk JA, Crawford CG (1989). Characterization of Maize endosperm-derived suspension cells throughout the culture cycle and enhancement of tissue friability. Plant Cell, Tissue and Organ Culture 18: 153-165.

[6] Gould AR, Everett NP, Wang TL, Street HE (1981) Studies on the control of cell cycle in cultured plant cells: Effect of nutrient limitation and nutrient starvation. Protoplasma 106:1-13 .

[7] Hong CH, Noh MS, Lee WY, Lee SK (2002). Inhibitory effects of natural sesquiterpenoids isolated from the rhizomes of Curcuma zedoaria on prostaglandin E2 and nitric oxide production. Planta Med 68: 545-547.

[8] Kato A, Tohoyama H, Joho M, Inouhe M (2007). Different effects of galactose and mannose on cell proliferation and intracellular soluble sugar levels in Vigna angularis suspension cultures. J Plant Res 120:713-719.

[9] Kirakosyan A, Kaufman PB (2009). Recent advances in plant biotechnology. Springer Dordrecht Heidelberg London New York, USA.

[10] Koch K.E. (2004) Sucrose metabolism: regulatory mechanisms and pivotal roles in sugars using and plant development. Curr Opin Plant Biol 7: 235-246.

[11] Lee EJ, Mobin M, Hahn EJ, Paek KY (2006). Effects of sucrose, inoculum density, auxins, and aeration volume on cell growth of gymnema sylvestre. .Journal of Plant Biology 49(6): 427-431.

[12] Ling OS, Kiong ALP, Hussein S (2008). Establishment and optimisation of growth parameters for cell suspension cultures of Ficus deltoidea. American-Eurasian Journal of Sustainable Agriculture 2(1): 38-49.

[13] Loc NH, Duc DT, Kwon TH, Yang MS (2005). Micropropagation of zedoary (Curcuma zedoaria Roscoe)-a valuable medicinal plant. Plant Cell Tiss Organ Cult 81:119-122.

[14] Matkowski A (2008). Plant in vitro culture for the production of antioxidants-A review. Biotechnology advances 26: 548-560.

[15] Matsuda H, Ninomiya T, Yoshikawa M (1998). Inhibitory effect and action mechanism of sesquiterpenes from Zedoariae rhizoma on D-galactosamine/lipopolysaccharide-induced liver injury. Bioorg Med Chem Lett 8: 339-344.

[16] Mello MO, Dias CTS, Amarai AF (2001a). Growth of Bauhinia forficata, Curcuma zedoaria, and Phaseolus vulgaris cell suspension cultures with carbon sources. Sci Agric 58: 481-485.

[17] Murashige T, Skoog F (1962) A revised medium for rapid growth and bioassays with tobacco tissue culture. Physiol Plant 15: 473-497.

[18] Duong Tan Nhut, Nguyen Trinh Tram and Nguyen Trinh Don (2006). Effect of sucrose, glucose and fructose on proliferation of Hymalaya Yew (Taxus wallichiana Zucc.) cell suspension cultures. A potential source for Taxol production. Proceedings of International Workshop on Biotechnology in Agriculture, Nong Lam University Ho Chi Minh City, Vietnam 142-145.

[19] Nigra, HM, Alvarez MA, Giolietti AM (1990). Effect of carbon and nitrogen sources on growth and solasodine production in batch suspension cultures of Solanum eleagnifolium Cav. Plant Cell, Tissue and Organ Culture 21: 55-60.

[20] Omar R, Abdullah MA, Hasan MA, Marziah M. (2004). Development of growth medium for Centenlla asiatica cell culture via response surface methodology. American Journal of Applied Sciences 1(3): 215-219.

[21] Shinde AN, Malpathak N, Fulzele DP (2009). Studied enhancement strategies for phytoestrogens production in shake flasks by suspension culture of Psoralea corylifolia. BioresourceTechnology 100: 1833-1839.

BA

A B

87

[22] Yoshioka T, Fuji E, Endo M (1998). Anti-inflammatory potency of

dehydrocurdione, a zedoary-derived sesquiterpene. Inflamm Res 47:

476-481.

[22] Sturm A, Tang GQ (1999). The sucrose cleaving enzymes of plants are crucial for development, growth and carbon partitioning. Trends Plant Sci 4: 401-407.

[23] Syu WJr, Shen CC, Don MJ, Ou JC, Lee GH, Sun CM (1998). Cytotoxicity of curcuminoids and some novel compounds from Curcuma zedoaria. J Nat Prod 61: 1531-1534.

[24] Villarreal ML, Arias C, Feria-Velasco A, Ramirez OT, Quintero R. (1997). Cell suspension culture of Solanum chrysotrichum (Schldl) - A plant producing an antifungal spirostanol saponin. Plant Cell Tissue and Organ Culture 50:39-44.

[25] Wu CH, Dewir YH, Hahn EJ, Paek KY (2006). Optimization of culturing conditions for the production of biomass and phenolics from adventitious roots of Echinacea angustifofia. J Plant Biol 49:193-199.

88

Use of BMWPVIET and ASPT indices as bioindicators for testing water quality of rivers in Danang city

Nguyen Van Khanh, Vo Van Minh, Kieu Thi Kinh Danang University of Education, Danang, Vietnam

vankhanhsk23@gmail.com

Tran Duy Vinh(1), Phan Thi Hien(2) (1) Okayama University, Japan

(2) Danang Dept. of Natural Resources and Environment, Vietnam

Abstract— As a consequence of rapid urbanization and industrialization, pollutants have increased in both quantity and diversity in the water environment. Monitoring plays a key role in prevention of negative impacts from potential environmental risks. Currently, physical and chemical analysis is considered as the popular methods for monitoring. Yet it is obvious that these methods require costly expenditure, high sampling frequency and difficult to evaluate the environmental status as well as to forecast the movement of chemical concentration. Using macro-invertebrates as bio-indicators has been evaluated as the advantageous method for testing water quality since the processed data can mirror water quality based on the ecological and biological characteristics. Biological indices named BMWP (biological monitoring working party) and ASPT (average score per taxon) have been initialized and adopted in temperature-countries and have shown remarkable advantages in assessment of water quality. In this paper, we present the initial results in applying BMWP and ASPT to monitor water quality of rivers in Central Region during 5 years from 2007 to 2011.

Keywords: BMWP, ASPT, macroinvertebrate, water quality,

biomonitoring.

I. INTRODUCTION Human activities have led to series adverse impacts the

environment. It is widely recognized that the increase of chemical concentration in water resulted by receiving effluents from industrial production, agricultural fertilizers, and domestic wastes have significantly contributed to environmental quality deterioration. Environmental risks should be prevented in order to avoid negative impacts to ecological health and human by betimes identifying the state of environmental quality. In terms of monitoring methods, currently, popular methods for testing water quality are majorly based on physical and chemical analyses [7], [9]. Conversely, these methods require costly expenses and high sampling frequency. In addition, effects of pollutants on species composition and community structure in aquatic ecosystem could not correctly be reflected and it is difficult to forecast aquatic-environmental changes [1], [2], [7], [9], [10].

One of the most practical methods to assess water quality that has highly been considered as an advantageous type is biological indices. Biological index called BMWP (Biological Monitoring Working Party) has been originated from Great Britain and adapted in Spain and Brazil, associated with ASPT

(average score per taxon) to evaluate water quality of streams [10]. The principle of this approach is based on presence or absence of macroinvertebrate families and their relative proportion may be indicative of the state of the water bodies [9], [11]. The mentioned biological indices have shown significant advantages in water quality evaluation, great abundance, for example, restricted mobility, predictable community structure, easy and cost-effective sampling and identification [9].

Vietnam has a continuous and stable economic growth rate, the remarkable improvement of social lives and living standard, however, negative impacts from urbanization on aquatic life may be inevitable and Danang city is not an exception. So far, in Vietnam, a practical procedure associated with data analysis and score system in term of biological indices has been established that called the BMWPVIET [9].

This paper aims at presenting the research results on the BMWP and ASPT for testing water quality of rivers in the period of 5 years from 2007 to 2011 and evaluates the applicability of the BMWP index.

II. METHODOLOGY In the period of 5 years from 2007 to 2011, researches were

carried out using the BMWP and ASPT for testing water quality of Phu Loc river, Han river, Cu De river and Cau Do – Tuy Loan river system in Danang city. Macroinvertebrates were sampled by using dredge bucket (500 nm mesh) and pond-net (500 nm mesh). Kick sampling technique was utilized from shallow sites, then the pond net was held vertically on the river bed against the current to catch macroinvertebrates [9]. From deeper sites, sweep sampling in which pond net and dredge bucket were used. In some cases, we searched the habitat of macroinvertebrates when the above basis techniques could not be done. After caught, samples were classified and stored in alcohol 70 degrees in the university laboratory. In the laboratory, macroinvertebrate samples were taxonomically identified as separated groups or families allocated a score between 1 and 10 in accordance to the method based on the BMWPVIET system established by Nguyen Xuan Quynh, Clive Pinder, Steve Tilling and Mai Dinh Yen [8], [9]. The scores for macroinvertebrate samples were then summed to give the BMWP score. After that, the Average Score Per Taxon (ASPT) were so calculated by dividing the BMWP score by the total number of taxa (families) in the sample [9].

89

Figure 1. Map of research sites

The extent of water pollution was classified in accordance to water quality identification system based on ASPT index of Richard Orton, Anne Bebbington, Jonh Bebbington (1995) and Stephen Eric Mustow (1997) as shown in table I [9].

TABLE I ASPT INDEX AND WATER POLLUTION LEVEL

ASPT index Category 0– 0.99 Extremely severe pollution

1.00 – 2.99 Severe pollution (Oligosaprobic) 3.00 – 4.99 Moderate pollution (α -mesosaprobic) 5.00 – 5.99 Moderate pollution (β-mesosaprobic) 6.00 – 7.99 Relatively clean water (Polysaprobic)

8 - 10 Clean water

III. RESULTS

A. Taxon composition in rivers in Danang city. Evaluation of macroinvertebrate composition provides

significant information in terms of ecological health and structure of aquatic organism community. Regardless of macroinvertebrates out of the BMWPVIET score system, from 2007 to 2011 in 4 rivers in Danang city, the research has resulted in 16 families belonging to 6 orders and 1 sub-class in Phu Loc river (2007 – 2008) [6]; 24 families belonging to 18 orders in Cu De river (2009 – 2010)[5]; 20 families belonging to 14 orders and 1 sub-class in Cau Do – Tuy Loan system river (2009 – 2010) [4]; 16 families belonging to 11 orders and 1 sub-class in Han rivers (2010 – 2011)[3]. The taxon compositions are specified in the following tables (from I to V).

TABLE II. TAXON COMPOSITION IN PHU LOC RIVER (2007 – 2008)

No. Order Family 1 Odonata Lestidae 2 Odonata Coenagrionidae 3 Coleoptera Chrysomelidae 4 Coleoptera Dytiscidae 5 Coleoptera Hydrophilidae

6 Coleoptera Hygrobiidae 7 Heteroptera Nepidae 8 Heteroptera Pleidae 9 Heteroptera Belostomatidae 10 Architaenioglossa Viviparidae 11 Basommatophora Lymnaeidae 12 Basommatophora Planorbidae 13 Decapoda Palaemonidae 14 Decapoda Parathelphusidae 15 Neotaenioglossa Thiaridae 16 Diptera Chironomidae 17 Oligochaeta*

*: sub-class

TABLE III. TAXON COMPOSTION IN CU DE RIVER (2009 – 2010)

No. Order Family 1 Odonata Amphipterygidae 2 Odonata Petaluridae 3 Decapoda Potamidae 4 Decapoda Palaemonidae 5 Decapoda Atyidae 6 Mesogastropoda Fairbankiidae 7 Mesogastropoda Fluminicolidae 8 Coleoptera Ptilodactiliidae 9 Coleoptera Psephenidae 10 Hypsogastropoda Assimineidae 11 Hypsogastropoda Bithynidae 12 Mytiloida Mytilidae 13 Aciculata Nereidae 14 Veneroida Corbiculidae 15 Unionoida Unionidae 16 Amphipoda Gammaridae 17 Basommatophora Ancylidae 18 Neotaenioglossa Thiaridae 19 Isopoda Corallanidae 20 Sorbeoconcha Pachychilidae 21 Rhynchobdellida Piscicolidae 22 Plecoptera Perlodidae 23 Diptera Ptychopteridae 24 Hemiptera Aphelocheiridae

90

TABLE IV. TAXON COMPOSITION IN CAU DO – TUY LOAN RIVER SYSTEM (2010 – 2011)

No. Order Family 1 Basommatophora Ancylidae 2 Basommatophora Lymnaeidae 3 Basommatophora Planorbidae 4 Neotaenioglossa Bithyniidae 5 Neotaenioglossa Thiaridae 6 Decapoda Palaemonidae 7 Decapoda Potamidae 8 Mesogastropoda Fluminicolidae 9 Mesogastropoda Trichotropidae 10 Coleoptera Psephenidae 11 Trichoptera Philopotamidae 12 Mytiloida Mytilidae 13 Heteroptera Hydrometridae 14 Veneroida Corbiculidae 15 Odonata Platycnemiidae 16 Diptera Ptychoteridae 17 Hypsogastropoda Assimineidae 18 Aciculata Nereidae 19 Sorbeoconcha Pachychilidae 20 Polychaeta**

*: sub-class

TABLE V. TAXON COMPOSITION IN HAN RIVER (2010 – 2011)

No. Order Family 1 Decapoda Potamidae 2 Decapoda Palaemonidae 3 Decapoda Parathelphusidae 4 Basommatophora Ancylidae 5 Basommatophora Lymnaeidae 6 Hemiptera Hydrometridae 7 Hemiptera Pleidae 8 Mesogastropoda Pilidae 9 Unionoida Unionidae 10 Odonata Lestidae 11 Architaenioglossa Viviparidae 12 Heteroptera Belostomatidae 13 Coleoptera Chrysomelidae 14 Basommatophora Planorbidae 15 Neotaenioglossa Thiaridae 16 Polychaeta**

*: sub-class

B. Results of testing water quality of rivers in Danang city. Macroinvertebrate community structure contains data

significant for reflecting the state of water body, the presence of pollution-intolerant families within high BMWP score suggest great freshwater biodiversity and good water quality, whilst low BMWP score families as pollution-tolerant families provide signs of environmental contamination.

TABLE VI. WATER QUALITY OF PHU LOC RIVER BASED ON ASPT INDEX

Research site

ASPT Rank ASPT Rank June 2007 September 2007

PL1 3.73 α-mesosaprobe 4.09 α-mesosaprobe PL2 3.83 α-mesosaprobe 3.98 α-mesosaprobe PL3 4.33 α-mesosaprobe 4.08 α-mesosaprobe PL4 3.72 α-mesosaprobe 4.61 α-mesosaprobe

December 2008 March 2008 PL1 3.67 α-mesosaprobe 3.60 α-mesosaprobe PL2 3.67 α-mesosaprobe 3.77 α-mesosaprobe PL3 3.76 α-mesosaprobe 3.78 α-mesosaprobe PL4 4.40 α-mesosaprobe 3.89 α-mesosaprobe

Table VI described water quality in Phu Loc river in 4 sampling times (2007 – 2008) within 4 seasons. The samples were taken to investigate APST index for testing water quality. The analytical results showed that APTS index ranged from 3.60 to 4.40 (Table VI). This indicated that Phu Loc river had signs of contamination, specifically, the water quality of Phu Loc river was identified as moderate pollution levels (α-mesosaprobe). Also, the research at Phu Loc indicated that no significant differences of seasonal variation were found, in other words, changes of the water quality investigated in Phu Loc river throughout 4 seasons (2007 – 2008) was inconsiderable [6].

TABLE VII. WATER

June 2009 January 2010 ASPT Rank ASPT Rank

CD1 3.25 α-mesosaprobe 5.25 β-mesosaprobe CD2 5.50 β-mesosaprobe 5.25 β-mesosaprobe CD3 5.00 β-mesosaprobe 5.50 β-mesosaprobe CD4 6.67 Polysaprobe 7.00 Polysaprobe CD5 7.00 Polysaprobe 6.40 Polysaprobe CD6 7.00 Polysaprobe 5.67 β-mesosaprobe

It can be seen that the water biological indices for each site

can be seen. ASPT scores ranged from 3.25 to 7.00 recorded at Cu De river in 2 occasions (June 2009 and January 2010). The water quality of Cu De river was ranged within levels from polysaprobe to α-mesosaprobe, especially, some research sites had high ASPT score (up to 7.00) indicated relatively good quality (Table VII) [5].

TABLE VIII. WATER QUALITY OF CAU DO – TUY LOAN RIVER SYSTEM BASED ON ASPT INDEX

Research site

August 2009 February 2010 ASPT Rank ASPT Rank

TL1 3.00 α-mesosaprobe 4.33 α-mesosaprobe TL2 3.00 α-mesosaprobe 3.50 α-mesosaprobe TL3 4.00 α-mesosaprobe 4.75 α-mesosaprobe TL4 3.00 α-mesosaprobe 3.83 α-mesosaprobe TL 5 5.00 β-mesosaprobe 3.83 α-mesosaprobe TL6 3.88 α-mesosaprobe 3.00 α-mesosaprobe

In Cau Do – Tuy Loan river system (2009 - 2010), ASPT scores ranged from 3.00 to 5.00 at research sites. According to classification system for freshwater quality [9], the quality of Cau Do – Tuy Loan river system was classified from β-mesosaprobic to α-mesosaprobe, in other words, the river water quality had signs of deterioration. Also, it implied that aquatic ecosystem at Cau Do – Tuy Loan river system was suffered negative impacts from pollution [4]. Specifically, ASPT scores for each site can be seen at the Table VIII.

TABLE IX. WATER QUALITY OF HAN RIVER BASED ON ASPT INDEX

Research site ASPT Rank November 2010

H1 3.25 α - mesosaprobe H2 3.50 α - mesosaprobe H3 4.00 α - mesosaprobe H4 3.25 α - mesosaprobe H5 3.50 α - mesosaprobe H6 4.00 α - mesosaprobe

January 2011 H1 4.40 α - mesosaprobe H2 4.40 α - mesosaprobe H3 4.40 α - mesosaprobe H4 4.57 α - mesosaprobe

91

H5 4.10 α - mesosaprobe H6 4.67 α - mesosaprobe

March 2011 H1 4.00 α - mesosaprobe H2 4.71 α - mesosaprobe H3 3.67 α - mesosaprobe H4 4.00 α - mesosaprobe H5 4.17 α - mesosaprobe H6 5.60 β- mesosaprobe

The research conducted in Han river (2010 – 2011) showed that ASPT indices ranged from 3.25 to 5.60. Hence, the water quality of Han river was classified from β-mesosaprobe to α-mesosaprobe. Similar to state of water body of Cau Do – Tuy Loan river, Han river had signs of contamination by receiving polluted effluents [3]. ASPT scores for each site can be seen at the table 9.

IV. DISCUSSION In the life circle, aquatic organisms frequently interact with

their habitat. Their numbers and compositions contain data being useful to determine environmental quality based on ecological features (e.g. chemical, physical, and biological). Based on that, broad measurements of the combined impacts of contaminations could be identified. Specifically, the ecological-based score of each family of aquatic species reflects their susceptibility to pollution, which is based on the principle that aquatic macroinvertebrates have different tolerance to pollutants, ecological health and state of water body hence they change in the way, which could be recognized.

Biomonitoring has not been accepted and utilized for governed monitoring programs by authorities in Vietnam. The results of this long-term research have enhanced and provided scientific data that it is applicable to adopt the BMWPVIET and ASPT for freshwater quality monitoring. The results have proved the advantages of biomonitoring which is a cost-effective and easy applicable method with easy identification and simple practical procedure. Particularly, it can rapidly indicate of the state of the water body and combined impacts of pollutants on community structure and ecological health which physical and chemical analyses implied limitation. Thus, the BMWPVIET and ASPT indices should be considered as an typical indicator for river quality monitoring.

Nguyen Xuan Quynh et al (2002) proposed a practical procedure with sampling which should be conducted 4 times in 4 seasons per year in the North of Vietnam. In the South of Vietnam, sampling may be implemented in 2 times (representing wet season and dry season) per year or over. Concerning to aquatic life, macroinvertebrates may be impacted by benthic structure, flow rate, chemical and physical parameters of the water body, specially, sudden weather changes. From the features of climatic condition in central Vietnam which could be divided to 2 seasons, dry season and wet season and the results of Phu Loc river, we propose that sampling could be taken in 2 times.

However, the adaption of BMWP score system to natural conditions in Danang city is required in order to enhance applicability of the BMWPVIET and ASPT. In fact, natural condition in Danang is different with natural condition in the

North of Vietnam with features that prolonged rainy season and high frequency of storm and other inordinate weather events, leading to reduce sudden disturbance of environmental parameters. From empirical results of the research, we propose that the best time for sampling per year is the end of wet season (from February to March) and mid dry season (from July to August) that are the moments when rainfall and river flows are relatively stable. Benthic structure of rivers, a factor influencing on aquatic ecosystems, such as Cu De river and Tuy Loan – Cau Do river system are complicated, therefore, to accurately evaluate the state of river water quality, distance between sampling sites should be from 1000 to 3000 m, depending on the length of each river. We identified and sampled in 4 rivers, the research results showed that the mentioned distance was appropriate.

A suitable adjustment associated with proper practical procedure demonstrated that it is significantly applicable to adopt BMWP and ASPT score system water quality testing in central Vietnam. It can be used as an independent method for rapid assessment of water quality, alternatively, as a advantageous tool combinative with physical and chemical analyses to elevate efficiency of environmental monitoring activities.

V. CONSLUSION The research has found 16 families and sub-class belonging

to 10 orders in Phu Loc river (2007 – 2008); 24 families belonging to 18 orders and 1 sub-class in Cu De river; 20 families belonging to 16 orders and sub-class in Cau Do – Tuy Loan system river; 16 families belonging to 11 orders in Han rivers in the BMWPVIET system.

The analytical analyses of ASPT index showed the water quality was moderately polluted (α-mesosaprobe) in Phu Loc river (2007 – 2008); classified from relatively clean water to moderately polluted water in Cu De river (2009 – 2010); moderately polluted (from β-mesosaprobic to α-mesosaprobe) in the Cau Do – Tuy Loan river (2009 – 2010) and in Han river (2010 - 2011). Biological indices reflected the state of water body of rivers in Danang city and provided broad measure of combined impacts of pollutants on ecosystem and organism life. The initial results also indicates that BMWPVIET and ASPT indices can undoubtedly be applied to monitor rivers in Danang city in particular and in central Vietnam in general.

REFERENCES

[1]. Truong Thanh Canh and Ngo Thi Tram Anh, "Application of macroinvertebrates for assessment of water quality in 4 main canal systems in Hochiminh city", Journal of Science and Technology Development, Vietnam 10, 25 - 31,2007.

[2]. L. Cota, M. Goulart, P. Moreno, and M. Callisto,"Rapid assessment of river water quality using a adapted BMWP index: a practical tool to evaluate ecosystem health", Verh. Internat. Verein. Limnol. 28, 1 - 4,2002.

[3]. Phan Thi Hien, Nguyen Dinh Anh, Vo Van Minh, and Nguyen Van Khanh, "Use of macroinvertebrates for water quality monitoring in Han river, Danang city", Journal of Science and Education, Danang University of Education 3 (02), 25 - 30,2012.

[4]. Nguyen Van Khanh, Pham Thi Hong Ha, and Dam Minh Anh,"Assessment of water quality of Cau Do - Tuy Loan river system in Danang city using BMWPVIET", Journal of Science and Technology, Danang University Vol 1. No. 5(40),2010.

92

[5]. Nguyen Van Khanh, Vo Van Minh, and Vo Huy Cam, "Assessment of water quality of Cu De river, Danang city using BMWPVIET score system" in National Scientific Conference for Union Officials. 2012: Hanoi University of Education.

[6]. Nguyen Van Khanh, Tran Duy Vinh, Duong Cong Vinh, and Ung Van Thach,"Use of macroinvertebrates for assessment of water quality in Phu Loc river", Journal of Science and Technology, Danang University 2 (37),2010.

[7]. Vo Van Phu, Hoang Dinh Trung, and Le Mai Hoang Thy,"Using large-size invertebrates to estimate water quality in Bo river, Thua Thien Hue province", Journal of Science, Hue University 57, 129 - 139,2010.

[8]. Nguyen Xuan Quynh, Clive Pinder, and Steve Tilling, Identification of commonly encountered macroinvertebrates in freshwater in Vietnam. 2001: Vietnam National University, Hanoi.

[9]. Nguyen Xuan Quynh, Clive Pinder, Steve Tilling, and Mai Dinh Yen, Freshwater environment biomonitoring using macroinvertebrates. 2002: Vietnam National University, Hanoi.

[10]. Kennedy Francis Roche, Ednilson Paulino Queiroz, and Karina Ocampo Righi Glaucio Mendes de Souza,"Use of the BMWP and ASPT indexes for monitoring environmental quality in a neotropical stream", Acta Limnologics Brasiliensia vol. 22, no. 1, p.105-108,2010.

[11]. Caroline L. Wenn,"Do freshwater macroinvertebrates reflet water quality improvement following the removal of point source polution from Spen Beck, West Yorkshire? " Earth and Envionment Vol. 3, 369 - 406,2008.

93

WHOLE CELL IMMOBILISATION OF BACILLUS SUBTILIS ON CELLULOSE CARRIERS AND WASTEWATER TREATMEMT APPLICATION

TRAN, Thi Xo Dept. of Biotechnology, Faculty of Chemistry Engineering, Danang University of Technology. NGUYEN, Thi Diem Quynh

Dept. of Medical and Pharmaceutical Science, Duy Tan University, Da Nang

Danang city, Vietnam

Abstract - In this study, we carried out fixing Bacillus subtilis on cellulose-rich carriers (filter-paper, coconut fiber). B. subtilis is capable of producing extra-cellular amylase, cellulase and protease effectively. After immobilization, bacteria have still maintained its original properties. Bacteria-fixed materials could be stored at 4C, for 3 months. Moreover, application of these carriers in treating wastewater which contains a large amount of protein, starch, and cellulose showed that COD and BOD5 significantly reduced (12.8% and 10.4%, respectively) in comparison with treating wastewater by active sludge.

Key words: Bacillus subtilis, immobilization, cellulose carriers, waste water treatment.

I. INTRODUCTION Bacillus subtilis is a rod-shaped, Gram-positive bacterium

which can be found in soil, water, air, dried grass, skin…This anaerobic or aerobic, non-pathogenic strain could prosperously grow at 5 – 500C. At higher temperature, B. subtilis is able to produce an endo-spore that allows it to endure extreme conditions of heat and desiccation in the environment. Bacillus subtilis is one of the most widely used bacteria for the production of enzymes (amylase, glucoamylase, glucanase, cellulase, dextranase, protease), and special chemical compounds, such as riboflavin. Therefore, B. subtillis has been applied in many industrial fields [1].

In 1978, Toshinori Kokubu immobilized Bacillus subtilis on polyacrylamide gel to produce α-Amylase [2]. In 2005, Kunamneni Adinarayana [3] attached Bacillus subtilis on others carriers such as calcium alginat, k -Carrageenan, ployacrylamide, agar-agar and gelatin to evaluate protease producing ability of B. subtilis.

Cellulosic materials also were choosen to immobilize bacteria. Aijun A. Wang, Ashok Mulchandani and Wilfred Chen (2001) [4] who studied on Eshcherichia coli immobilization on cellulosic carriers, concluded that E. coli was capable of binding on surface of cellulosic materials and

its binding level depended on surface structure. In 1980, Mogilevich et al. carried out Bacillus subtilis fixation on cellulosic carrier. They indicated that bacterial activity was not changed as fixing on this material [5].

Based on the previous research, we performed Bacillus subtilis immobilization on cellulosic surface (filter-paper, coconut fiber) for preservation as well as wastewater treatment application

II. MATERIALS AND METHODS

A. Materials:

Bacillus subtilis was provided by Biotechnology Lab, Danang University of Technology. Filter-paper with high thickness was taken from Advantec company, China. Coconut fiber bought from supermakets was pretreated before performing the immobilization.

B. subtilis medium ingredients: Beef extract 3g/L, Yeast extract 10g/L, Pepton 10g/L, NaCl 5g/L.

B. Methods

1) Method of evaluating extracellular enzyme production ( amylase, cellulose and protease).

Bacteria were cultured on medium containing 1% substrates (starch, Carboxymethyl cellulose - CMC, and gelatin). Then, the culture was grown at 30oC for 24h. We used Lugol reagent to determine amylase and cellulase’s activity and amido black dye to determine protease’s activity.

2) Method of whole cell immobilization

We chose absorption method to fix whole cell on carriers. Cellulose carriers autoclaved were put into the growth culture of B.subtilis, then incubated at 30oC for 12h with shaking (150 rpm) for attachment between bacteria and cellulosic surface. After that, these materials were taken out of culture and dried

94

at appropriate temperature. Finally, bacteria-immolized carriers were stored at 4C.

III. RESULTS AND DISCUSSION

A. Carrier preparation

When soaking coconut fibres into water, water color became brown-red. Therefore, we pretreated coconut fiber by soaking it into NaOH 1% for 4h. Then, they were washed many times by water until pH of water was neutral (Fig. 1).

Figure 1. Water color before (a) and after (b) treatment.

Treated fiber dried to moisture lower than 15% and filter-paper were cut into 4 cm lines and 1x1.5 cm piece, respectively. Before fixing bacterial cells on surface, we evaluted cellulose and lignin concentration of cellulosic surfaces. The result (table 1) showed that treated fiber consist of not only cellulose, but also lignin (37.2%).

Table 1. Cellulose and lignin concentration of coconut fibres and filter-paper.

Sample Fiber Filter-paper

Cellulose (%) 40,0 63,3

Lignin (%) 37,2 -

B. Evaluation of exoenzyme production of Bacillus subtilis

After activating the bacteria in media containing 20 g/l agar, we checked extracellular enzyme producing ability of the bacteria by measuring hydrolyzed circle diameter. The results indicated that circle diameters produced by amylase, cellulase, and protease’s hydrolyzation on appropriate substrates (starch, cellulose, protein) were 0.9, 2.8 and 1.0 cm, respectively (Fig 2). This result indicated that B. Subtilis is able to synthesis all 3 types of these enzymes.

Figure 2. Evaluation of exoenzyme production of Bacillus subtilis: amylase (a), cellulase (b), protease (c)

C. Whole cell immobilization of Bacillus subtilis on filter-paper and coconut fiber

Firstly, the bacteria were grown in broth at 30oC, with shaking (150 rpm) for 24h until the culture reached to log phase. Then, cellulosic carriers were incubated into the culture for 12h. This incubation combined with discontinous shaking which performed shaking for 15 min after 1h stationary incubation. After that, the carriers were taken out of the culture and dried at 38.5oC by using vacuum drying cabinet. In order to achieve initial dryness, bacterial cell-immobilized filter paper and coconut fiber were heated for 32h and 35h, respectively.

After immobilization, bacterial activity was examined at 2 periods: after drying and after 3- month storage. In order to activate the bacteria, we soaked dried carrier in liquid medium for 15 min and then asepticly put on agar medium. The plate was incubated at 30oC. After 24h, typical colonies of Bacillus subtilis were formed (Fig. 3)

Figure 3. Evaluation of growth of B. Subtilis after immobilization on carrier

(a) filter paper (b) Coconut fibre

After 3-month storage, we evaluated extracellular enzyme production of B. subtilis. The result showed that the bacteria still maintained original properties (Fig.4).

(a) (b) (c) Figure 4: Evaluation of exoenzyme production of Bacillus subtilis: amylase (a), cellulase (b), protease (c) after 3 months storage

Based on these results obtained, we realized that:

- Bacillus subtilis is able to immobilize on cellulose rich carriers by using absorption method.

- The present of lignin in coconut fiber did not badly affect the binding of bacterial cell on carrier.

B A

A C B

95

- Immobilization manipulation was performed in aseptic conditions. Therefore, it is possible to use this method for B. subtilis storage purpose.

- After 3 months storage at 4oC, B. Subtilis still conserves characteristic properties, such as exoenzyme producing ability (amylase, cellulase and protease).

D. Application of Bacillus subtilis immobilization on carriers in wastewater treatment.

The bacteria immobilized on cellulosis carrier is used as a potential source for wastewater treatment. Wastewater is a nutrient-rich medium that is suitable for bacterial growth. After treatment, bacterial biomass in the wastewater is fixed on carrier. By that way, it is completely possible to use bacteria-immobilized carriers for many times.

We attached the bacteria on long coconut fibres with 8 cm length, 5 cm width and 1.5 cm thickness. Wastewater samples were taken from wastewater pre-biotreating area of Dien Nam-Dien Ngoc Industrial Park, Quang Nam province. This Industrial Park hosts seafood factories, breweries, paper-mills, so wastewater from these factories facilitate the bacteria’s growth to produce enzymes for wastewater treatment.

Then we carried two procedures for the treatment of that wastewater. The first procedure is aerobic treatment with actived sludge for 12h. The effuluent sample was considerd as a control. In the second procedure, the wastewater was passed through bacteria-immobilized coconut fibre, then the output wastewater was aerobically treated with actived sludge for 12h as in procedure 1. BOD5 and COD of pretreated and treated wastewater were evaluated. The results (Table 2) showed that after passing wastewater through bacteria immobilized in the carrier, COD and BOD5 significantly reduced. BOD5 and COD was 10.4% and 12.8% lower than controls, respectively.

Table 2: COD and BOD5 value of pretreated and treated wastewater

Sample COD (mg/l) BOD5 (mg/l)

Pretreated 126.70 74.53

Treated (Procedure 1) 83.30 47.60

Treated (Procedure 2) 67.10 38.90

This indicated that it is possible to fix non-pathogenic bacteria on natural carriers for environmental treatment.

4. Conclusion

Bacillus subtilis, non-pathogenic, enzymes producing baterium, could be fixed on cellulosis carriers, such as filter-paper, coconut fibre. After immobilization, bacteria have still maintained its original properties. Bacteria-immobilized carriers could be stored 4C for 3 months. Application of these carries fixed with bacteria in protein, starch and cellolose rich wastewater gave lower COD và BOD5 value than treatment wastewater with active sludge.

REFERENCES

[1]. Nguyễn Lân Dũng, Nguyễn Đình Quyến, Phan Văn Ty (2002), Vi sinh vật học, Nhà Xuất Bản Giáo Dục.

[2].Toshinori Kokubu, Isao Karube and Shuichi Suzuki, (1978), “α-Amylase production by immobilized whole cells of Bacillus subtilis”, European J. Appl. Microbiol, Volume 5, pp. 233-240.

[3].Kunamneni Adinarayana, Bezaada Jyothi, and Poluri Ellaiah, (2005), “Production of Alkaline Protease With Immobilized Cells of Bacillus subtilis PE-11 in Various Matrices by Entrapment Technique”, AAPS PharmSciTech, Volume 6, pp. 391-397.

[4]. Aijun A. Wang, Ashok Mulchandani and Wilfred Chen, (2001), “Whole-Cell Immobilization Using Cell Surface-Exposed Cellulose-Binding Domain”, Biotechnol. Prog, Volume 17, pp.407−411.

[5]. Mogilevich NF, Garbara SV, (1980), “Effect of electro-immobilization on Bacillus subtilis”, Mikrobiologiia, Volume 49, No 06, pp. 876-879

96

LACTIC ACID FERMENTATION FROM JACKFRUIT SEED FLOUR

Trương Thi Minh Hanh Faculty of Chemistry, Danang University of

Technology, email: tminhhanh2001@yahoo.com

Ho Thi Hao

Agriculture and Forestry Faculty, Tay Nguyen University, email: hoahaoanh123@gmail.com

Abstract— In this paper, we present our research results on lactic acid fermentation process from jackfruit seed flour. The hydrolysis process of jackfruit seed starch using two enzymes, α-amylase and glucoamylase with the percentage of 0.10% and 0.18%, respectively, compared to the substrates, yields the hydrolysis efficiency of 91.436%. Four factors that affect the lactic acid fermentation are investigated in this study, including reducing sugar concentration, Lactobacillus casei bacteria percentage, temperature, and fermentation duration. Based on experimental results, we propose parameters for lactic acid fermentation: (i) initial reducing sugar concentration of 5%, (ii) Lactobacillus casei bacteria percentage of 5% with cell density of 51.103CFU/ml, (iii) temperature of 37OC, and (iv) fermentation duration of 72 hours. Under the above conditions, the obtained lactic acid concentration is thus 18.25g/l we propose parameters for lactic acid fermentation: (i) initial reducing sugar concentration of 5%, (ii) Lactobacillus casei bacteria percentage of 5% with cell density of 51.103CFU/ml, (iii) temperature of 37OC, and (iv) fermentation duration of 72 hours. Under the above conditions, the obtained lactic acid concentration is thus 18.25g/l .

Keywords- Jackfruit seed flour, agricultural residue, liquefaction, saccharification, lactic acid fermentation medium

I. INTRODUCTION Lactic acid is a product of a lactic fermentation process,

which is an anaerobic biological metabolism that transforms sugar compound mainly into lactic acid and other by-products. Lactic acid is widely used in the food industry and others such as cosmetic, chemical, and pharmaceutical industries. In novel material technologies, lactic acid is a raw material to produce poly lactic acid (PLA), which is an important biomaterial for many other industries, through polymerization reactions. Raw materials for producing lactic acid conventionally come from carbohydrate compounds such as saccharose, glucose and lactose as well as starch such as corn, potato, cassava, sugarcane molasses, starch hydrolyzate and timber hydrolyzate. Recently, many scientists have increasingly focused on using agricultural and industrial residues as a new approach for lactic acid production [3].

Jackfruit, a delicious and nutritious tropical fruit, is an important material for the production of jackfruit jam, dried jackfruit, jackfruit juice, jackfruit preserve, candied lanka [5]. Interestingly, Daklak is a province that has relatively high

jackfruit production in Vietnam. In 2010, the province’s total production of jackfruit is 280667.8 tones, in which Vinamit Company’s need is 132000 tones. According to Daklak Agricultural Department, the amount of jackfruit seeds released into the environment is 17160 tones. These seeds will become pollution unless they are treated effectively. Since jackfruit seeds contain high concentration of starch and protein, it is possible to use the jackfruit seeds as a good substrate source for lactic acid fermentation. In other words, lactic acid fermentation from jackfruit seed starch not only offers significant economic value but also contributes to environmental improvement.

II. MATERIALS AND METHODS

A. Materials We used the following materials:

Jackfruit seeds were collected from a small jackfruit factory of Vinamit Company in Buon Ho, Krongbuk, Daklak Province.

Starch hydrolysis enzymes used in this study were α-amylase and glucoamylase (γ-amylase), provided by Novozyme Company (Denmark) and Genencor Bio-products Co. Ltd (China), respectively.

Lactobacillus casei were provided by Nanogen Biopharma, Hochiminh City, Vietnam.

B. Methods 1) Analysis methods

After being transformed into reducing sugar by using an acid solution 2%, starch was quantitatively determined using the Graxianop method [2]. The reducing sugar concentration in the obtained hydrolyzate was determined by the colorimetric method using dinitrosalicylic acid (DNS) [2]. In addition, lactic acid was determined by the colorimetric method using paraoxidiphenyl [2]. The activities of α-amylase and glucoamylase were determined by using the Rukhliadeva method and the V.Y.Rodzevich - O.P.Korenbiakina method, respectively [2].

2) Technological methods

97

a) Jackfruit processing and jackfruit flour production

b) Starch hydrolysis method

c) Lactic fermentation method:

- Medium preparation: The production medium contained 8.0g/l yeast extract, 0.5g/l dipotassium hydrogen phosphate, 0.5g/l potassium dihydrogen phosphate, 1.0 g/l sodium acetate 3H2O, 0.6g/l magnesium sulfate 7H2O, 0.03g/l manganese sulfate 4H2O, and a 1000-ml hydrolysate of jackfruit starch [34]. The pH of the medium was adjusted by H2SO4 20% and NaOH 5N [34] to pH=6. The medium was sterilized at 121oC for 20 minutes

- Fermentation procedure: Fig. 1 summarizes the fermentation procedure and the processing of the fermented solution in order to quantitatively determine lactic acid [1]

Figure 1. Fermentation procedure and fermented solution processing

III. RESULTS AND DISCUSSION

A. Investigation of material characteristics before lactic acid fermentation

Table I showed that the content of jackfruit starch in this work was lower than that reported in previous study (77.76 ± 0.965%) [6]. It is probably due to partially ripe jackfruit. Nevertheless, with humidity under a certain threshold (< 13%), this jackfruit flour is still suitable for our research purpose and even for production. In addition, we observe that the activities of the enzymes at the experimentation time were high. However, since the ratio of glucoamylase enzyme to the substrate had not been determined, we carried out further study to find the best enzyme concentration for the saccharification process.

TABLE I. MATERIAL CHARACTERISTICS BEFORE LACTIC ACID FERMENTATION

Feature Jackfruit flour

Enzyme

Humidity (%) 11.06

Starch (%) 65.62

Activity of α-amylase enzyme product

3623.4 unit

Activity of glucoamylase enzyme product

4752.0 unit

Note:

Unit of Activity of α-amylase enzyme product is the amount of enzyme (ml) that transforms 1g of dissolved starch into dextrin with different molecular weights at 30oC in 1 hour.

Unit of Activity of glucoamylase enzyme product is the amount of enzyme (ml) that applies on dissolved starch solution at pH = 4.7 at 30oC in 1 hour to release 1mg of glucose.

B. Effects of the glucoamylase enzyme percentage on the hydrolysis of jackfruit seed starch Experiments were carried out according to the flow chart of

starch hydrolysis in Section 2.b. We used 0.1 % of α-amylase and 33% of jackfruit seed flour and γ-amylase varied from 0.12 to 0.20%. The results of Table II presented reducing sugar concentrations obtained from jackfruit seed hydrolysate under different glucoamylase amounts. TABLE II: REDUCING SUGAR CONCENTRATION OBTAINED AFTER

HYDROLYSIS

No. Supplemented

enzyme γ – amylase, %

Obtained reducing sugar,

g/100ml

Hydrolysis efficiency, %

1 0.12 10.769 49.233

2 0.14 11.667 53.339

3 0.16 14.000 64.005

4 0.18 20.000 91.436

5 0.20 15.556 71.119

The data in Table II showed that elevated γ–amylase

gave the increase of reducing sugar amount in the hydrolyzed solution. However, when the percentage of γ–amylase was 0.20 %, the concentration of reducing sugar decreased. It is maybe due to competing effects between the enzyme and the substrate. Moreover, using 0.18% of γ–amylase offered highest hydrolysis efficiency (91.436 %). It is in the same magnitude of results reported by Nguyen Minh Hien and

98

Nguyen Thuy Huong [7]. Therefore, we chose 0.18% γ–amylase for next experiments.

C. Investigation of factors affecting lactic acid fermentation

1) Effect of fermentation temperature This experiment was carried out with the following

parameters: (i) the initial reducing sugar percentage 4%, (ii) fermentation duration 72h, (iii) stirring speed 150 rpm, (iv) cultivated Lactobacillus casei strain percentage that was added 5% (v/v) compared to the fermentation medium, with cell density 51.103CFU/ml and (v) fermentation temperature 35OC - 40OC. The result was shown in Fig. 2.

Figure 2. Effects of temperature on lactic acid fermentation

Accoding to Fig. 2, the Lactobacillus casei bacteria grew well at the temperature of 37oC in the medium containing hydrolyzed jackfruit starch solution. At this condition, the concentration of lactic acid achieved the maximal value of 18.20 g/l. Hence, we used 37OC as a suitable temperature in next experiments.

2) Effects of bacteria strain percentage Lactic acid fermentation process was carried out with the

initial reducing sugar 4% at 37OC for 72 hours. The solution was shaken with 150 rpm. The concentration of bacteria strain inoculated in medium was changed from 2.0 to 8.0 %. The results of Fig. 3 showed that the initial bacteria strain percentage of 5 % was the most suitable for lactic acid fermentation.

Figure 3. Effects of the bacteria strain percentage on lactic acid fermentation

3) Effects of initial reducing sugar percentage The lactic acid fermentation was carried out by culturing

bacteria in medium containing 5 % of bacteria strain and 2.0 – 8.0 % reducing sugar. The solution was shaken with rate of 150 rpm at 37OC for 72 hours. The result is shown in Fig. 4.

.

Figure 4. Effects of initial reducing sugar percentage on lactic acid fermentation

4) Effects of fermentation duration: Similarly, the lactic acid fermentation was carried out

with bacteria strain of the 5% concentration and initial reducing sugar of 4%. We used different durations in this experiment. In particular, the solution was shaked with rate of 150 rpm/min at 37OC from 24 to 88 hours. The result is shown in Fig. 5..

By varying three effecting factors, namely bacteria strain percentage, initial reducing sugar concentration and duration, we determined that lactic acid fermentation is best when Lactobacillus casei was grown in the medium containing 5% initial reducing sugar concentration at 37OC for 72 hours.

Figure 5. Effects of fermentation duration on lactic acid fermentation

IV. CONCLUSION In summary, we have studied the hydrolysis of jackfruit

seed starch using the α-amylase enzyme with percentage of 0.1% and the γ-amylase enzyme with percentage of 0.18%

99

compared to the substrate. The results showed that we achieved the hydrolysis efficiency of 91.44%, reducing sugar concentration of 200 g/l, and the quality of hydrolyzed solution was good. In addition, we have determined optimal conditions for lactic acid fermentation as following: temperature of 37OC, Lactobacillus casei inoculum of 5%, initial reducing sugar concentration of 5%, and fermentation duration of 72h. Under the above conditions, the lactic acid concentration obtained was 18.25 g/l. These results have shown that utilizing jackfruit seeds in lactic acid fermentation technologies has promising economic value and reduces significantly harmful wastes in the environment.

REFERENCES [1] Nguyễn Đức Lượng, “Công nghệ vi sinh vật - tập 2”, Trường Đại học

Kỹ thuật, 1996 [2] Lê Thanh Mai, “Các phương pháp phân tích ngành công nghệ lên men”,

Nhà xuất bản Khoa học và Kỹ thuật Hà Nội, 2005. [3] P.Singh.nee’Nigam, “Production of Organic acids from Agro_Industrial

Residues”, Faculty of life and Health Sciences, School of Biomedical Sciences, University of Ulster, Coleraine BT52 1SA, Northern Ireland, UK. 2009

[4] Sakhamuri Sivakesava, Joseph Irudayaraj Demivci Ali, “Simultaneous determination of multiple components in lactic acid fermentation using FT_MiR, NiR and FT_Raman spectroscopic techniques”, Process Biochemistry 2001

[5] Technoguide Series, “Jackfruit DA-RFU 8”, eastern Visayas Integrated Agricultural Research Center (EVIARC)

[6] Vanna Tulyathana, Kanitha Tananuwonga, Prapa Songjinda and Nongnuj Jaiboon, “Some Physicochemical Properties of Jackfruit (Artocarpus heterophyllus Lam) Seed Flour and Starch”, 2001, Department of Food Technology, Faculty of Science, Chulalongkorn University, Thailand 10330.

[7] http://www.cesti.gov.vn/kh-cn-trong-n-c/nghi-n-c-u-qua-trinh-th-y-phan-tinh-b-t-h-t-mit-b-ng-enzyme-termamyl-120l-va-amg-300l-d-l-n-men-r-u-ch-ng-c-t.html

100

GOLD NANOPARTICLES BASED LOCALIZED SURFACE PLASMON RESONANCE IN COMBINATION WITH MICROFLUIDIC SYSTEM FOR BIOMOLECULE

DETERMINATION

Nguyen Ba Trung, Le Tu Hai, Danang University of Education, Danang University

Yuzuru Takamura Japan Advanced Institute of Science and Tecnology

Abstract—In recent years, metallic nanoparticles have been studied extensively for the nanoelectronics, nanophotonics, nano-catalyst, and biosensor applications. The unique optical property of metallic nanoparticles is suitable to be employed as a marker for the label-free optical detection based on localized surface plasmon resonance. In this work, we investigate a simple method for monitoring the interaction of antigen-antibody on the PDMS sensing surface based on the localized surface plasmon resonance of immobilized gold nano particles. The interaction between antigen-antibody was examined by recording the absorbance intensity, as well as the peak wavelength shift of the LSPR band. The results showed that our device can be employed to qualitatively and quantitatively analyze the antigen presenting in sample. The developed technique is also hopefully expanded the potential applications of PDMS microfluidic chip on studying immunoassays and other biochemical analyses

Keywords-component; surface plasmon resonance; biosensing; gold nano particles; immunoassay; microfluidic

I. INTRODUCTION In recent years, metallic nanoparticles have been studied

extensively for the nanoelectronics, nanophotonics, and biosensor applications.1) The unique optical property of metallic nanoparticles is suitable to be employed as a marker for the label-free optical detection based on localized surface plasmon resonance (LSPR) which is a collective oscillation of free electron on metal nanoparticles induced by incident light. The absorbance peak wavelength and the spectrum shape of AuNPs are sensitive to the size, shape, and aggregation of AuNPs, as well as the surrounding medium. The principle of LSPR sensor is to measure changes in a characteristic of the LSPR spectrum from alteration in the effective refractive index of the surrounding medium, leading to a shift in the absorbance peak wavelength. Nowadays, LSPR sensors have attracted increasing attention to use for detecting chemical and biological substances related to medical diagnostics, environmental monitoring, agriculture pesticide, antibiotic monitoring, food additive testing, chemical agent testing, and drug screening.

Micro total analysis system (µ-TAS) has been established itself at the forefront of analytical chemistry, especially biosensing microfluidic devices. Recently, soft polymer materials are known as the chosen materials for fabricating multifunctional microfluidic devices instead of glass, quartz or silicon. Among them, Polydimethyl siloxane (PDMS) is one of the most widely-used polymer materials due

to its transparency, good biocompatibility, facile bonding ability, high transparency for UV and fluorescence detection, and cost-effectiveness for the production.2)

The combination of microfluidic chip made of PDMS material and unique property of localized surface plasmon resonance of noble metal NPs could be efficiently applied for the label free biosening microfluidic device. In this work, AuNPs were directly immobilized on the sensing surface of PDMS microfluidic system for biosensing applications. Immunno assay for the detection of C-reactive protein (CRP) antibody-antigen interaction was carried out using the developed device. The sensitivity is comparable to the reported method3). This method contributes to the development of label free biological molecule analyses.

II. EXPERIMENTAL

A. Chemicals The Sylgard 184 including PDMS monomer and curing

agent were purchased from Dow Corning (Midland, MI, USA). Aminopropyltriethoxysilane (- APTES) and D-(+) Glucose were purchased from Sigma-Aldrich. Citrate stabilized AuNPs 100 nm in average diameter distribution was purchased from British Bio Cell and used as received. 10-carboxyl-1-decanethiol and 1-ethyl-3-(3-dimethyl aminopropyl) carbodiimide hydrochloride (EDC) were bought from Dojindo Laboratories, Japan. N-hydroxysuccinimide (NHS) was purchased from Wako, Japan. Goat Anti – Human C – Reactive Protein (CRP) and C - reactive protein were bought from Bethyl Laboratories, Inc and Sigma-Aldrich, respectively.

B. Absorbance measurement and surface morphology characterization

1) Absorbance measurement Absorbance measurements of AuNPs patterned on PDMS

surface in the visible region were performed using Shimazu UV-3100. LSPR measurements on the chip were performed using a set of instruments as shown in Figure 1: spectrophotometer (USB-4000-UV-Vis, wavelength range: 200 –1100 nm), tungsten halogen light source (LS-1, wavelength range: 360 – 2000 nm), and optical fiber probe (R-400-7 UV/Vis, fiber core diameter: 400 m, wavelength range: 250 – 800 nm) with collimating lens that were all purchased from Ocean Optics. The sample was placed proximately to the optical fiber probe bundle surface so that the incident light could be able to pass through the sample. The intensity of the

101

transmitted light was detected by another optical fiber probe, and then analyzed by UV–Vis spectrophotometer in a wavelength range of 400 – 900 nm at room temperature. The absorbance peak wavelength of AuNP was determined using Gaussian curve fitting method.

2) Surface morphology characterization: The gold colloid distribution on the substrate surface was

characterized by a commercial Atomic force microscope (Digital Instrument D3100). Measurements done in tapping mode with a force constant of 18N/m and scan rates in the range 0.25–0.50 Hz allowed us to observe the gold colloids on the substrate surface

Figure 1. Experimental setup for absorbance spectrum measurements

3) Immobilization of AuNPs on the PDMS substrate AuNPs were immobilized ion the PDMS sensing surface as

our earlier description. For the immobilization procedure of AuNPs on the sensing surface of microfluidic chip, mask tightly sealed with PDMS were used to protect the desired region on PDMS surface during oxygen plasma treatment. Si–H group on the surface would be consumed by the oxidation with oxygen plasma and SiOx silica-like layer would be formed on PDMS surface.4,5) The silanisation exclusively occurred on the oxygen plasma treated area, which allows us to make spatially selective deposition layer of AuNPs onto a PDMS surface. Plasma treated PDMS surface was then washed with ethanol before loading -APTES on the PDMS sensing surface and kept in the desiccators for 15 minutes. The sample was then washed again with ethanol and deionized water to remove any remaining residual -APTES molecules. The samples were annealed at 1200C for two hours. The silanised samples were immersed into an aqueous solution of citrate stabilized gold colloids at proper time and thoroughly rinsed with deionized water. Finally, the samples were then annealed for about 30 min at 1200C

4) Microfluidic LSPR chip for monitoring the real time interaction between Antigen – Antibody

The LSPR PDMS chip is composed of two layers: a flow layer and a cover layer bonded by oxygen plasma. The microfluidic device was fabricated using PDMS by standard soft-lithography techniques8). The flow layer contains a sensing chamber connecting with inlet and outlet microchannels. The channels were created with the width of 200 µm, the height of 250 m, and 20mm in length. The circular shaped chamber is 5 mm in diameter. The inlet and outlet of the microfluidic device was formed by connecting with the fluorinated ethylene

propylene (FEP) tube (0.15 ± 0.05mm i.d.) and sealing with PDMS to prevent leak. Flow through the microfluidic LSPR chip was performed using micro-syringe pump (KD Scientific). The flow rate of 60 L/min was set in all experiments

5) Antibody- antigen interaction experiment The immobilization of CRP antibody on the surface of

AuNPs chip is shown in Figure 2. At first, the formation of self-assembly monolayer (SAM) was archived by introducing 1mM of 10-carboxyl-1-decanethiol in ethanol onto the surface of AuNPs layer on chip and keeping overnight at room temperature. SAM functionalization was carried out with 1:1 in volume mixture of 400mM of 1-ethyl-3-(3-dimethyl aminopropyl) carbodiimide hydrochloride (EDC) solution and 100mM of N-hydroxysuccinimide (NHS) solution. The mixture was added to the SAM functionalized surface for 30 minutes. The 100 g/mL of CRP antibody was immobilized onto activated SAM modification surface for 30 minutes. After the immobilization of CRP antibody, the chip surface was rinsed thoroughly with 20mM phosphate buffered saline (PBS, pH 7.4), and then blocked with 1% BSA to prevent non-specific absorption. After rinsing with PBS to eliminate the unbound molecules, the initial absorbance spectrum was recorded. Next, CRP antibody immobilization procedure was performed to monitor the interaction between CRP antibody-antigen. A desired concentration of CRP antigen solution in PBS was introduced onto the CRP antibody immobilized Au NPs surface. The interaction was allowed while continuously incubating for 30 minutes at room temperature. After a rinsing and drying up the surface, the changes in the absorption spectrum caused by the interaction were observed.

Figure 2. The schematic of antibody immobilization process on the sensing

surface and the interaction between the antibody and its relevant antigen

III. RESULTS AND DISCUSSION A) Immobilization of AnNPs on to PDMS surface

For the introduction of active groups on the inert PDMS surface, oxygen plasma was applied in order to replace the silane (Si-Me) groups on the PDMS surface with silanol groups (Si-OH). After treating with oxygen plasma, PDMS surface can react with 1% -APTES in ethanol to generate the -NH2 group on its surface. The positively charged amino groups can help to immobilize AuNPs onto silanised substrates by the strong electrostatic interaction with the negatively charged gold core surrounded by ionic double layer. The formation of AuNPs onto the surface of PDMS was monitor using UV-Visible absorption spectroscopy. Figure 3 shows the surface plasmon band of AuNPs on the PMDS sensing surface with and without

102

oxygen plasma. From the results, it can be inferred that treatment of PDMS with oxygen plasma is crucial for the formation of -APTES film which is necessary for the AuNP immobilization

Figure 3. (a) UV-Vis spectra of 100nm AuNP patterned on PDMS substrates, performed on Shimadzu UV-310 0 UV-Vis- NIR recording spectrophotometer. The silanisation was carried out with 10% - APTES solution in ethanol for 15 minutes. The incubation time for AuNP deposition at initial concentration was 4 hours. (b) AFM images of AuNPs on 10% -APTES modified PDMS surface.

B/ Localized surface plasmon resonance of AuNPs with the changes in refractive index

400 500 600 700 800 9000.06

0.08

0.10

0.12

0.14

0.16

0.18

Abso

rptio

n

Wavelength (nm)

Glucose 10% Glucose 20% Glucose 30% Glucose 40% Glucose 50%

Figure 4. Absorbance spectrum of a monolayer of immobilized gold colloids on PDMS in the following glucose concentrations: Glucose 10% (n = 1.347);

Glucose 20% (n = 1.361); Glucose 30% (n = 1.375); Glucose 40% (n = 1.389); Glucose 50% (n = 1.403)

We examined the ability of the immobilized monolayer of AuNPs to transduce changes in the surrounding refractive index of different glucose concentrations into the absorbance spectrum. As shown in the Figure 4 there are two plasmon resonance peaks of AuNPs on the sensing surface. The first peak around 600 nm in wavelength characterized for mono AuNPs with high absorbance intensity and another lower absorbance intensity peak at longer wavelength around 890 nm characterized for cluster AuNPs. The Plasmon of single AuNPs is sensitive enough with the changing of RI of surrounding medium, so that for the discussion on sensing application of this material, we would only mention the first peak.

In the absorbance spectrum of AuNPs shown in Figure 5, it exhibited a red shift in the peak wavelength, along with an increase in the absorbance intensity as increasing of the concentration of D-(+) glucose solution in the range of 10% to 50%. It was found that a linear relationship between the LSPR peak (max) and the refractive index of glucose solution. A

least-square gave a sensitivity of 233 nm/refractive index unit (RIU) with linear regression value R2 = 0.93 and 0.38 absorbance unit/RIU with R2 = 0.97. The result indicates that the optical property of an immobilized monolayer of AuNPs is sensitive to the refractive index of the surrounding bulk medium. This sensitivity is three times higher compared to the reported paper (76.4nm/refractive index unit) [7].

Figure 5. The dependences of the absorbance intensity and the wavelength shift on different refractive indexs of the glucose solutions in water.

C/ Localized surface plasmon resonance of antibody-antigen interaction

400 500 600 700 800 900

0.08

0.12

0.16

0.20

0.24

Abso

rban

ce

Wavelength (nm)

SAM Antibody BSA Antigen

Figure 6. The superimposed absorbance spectra curves obtained during

the experiment of CRP antibody-antigen interaction at 1ng/mL of CRP antigen.

We examined whether the refractive index change at the surface of AuNPs due to bio-molecular binding events could also be transduced into an experimentally detectable change in the absorbance spectrum. CRP antibody and CRP antigen interaction was used to monitor the characteristic of this sensing device. Figure 6 shows the superimposed absorbance spectrum curves obtained during CRP antibody immobilized process and the interaction of CRP antibody with 10 ng/mL CRP antigen. The significant increase in absorbance intensity as well as the red shift in maximum wavelength absorbance was observed after each step. The dependence of wavelength shifts on the CRP antigen was also checked. Figure 7 shows the linear regression line fitted the experimental data with linear regression value R2 = 0.9875 for the concentrations ranging from 10 ng/mL to 0.1 µg/mL. This result, one more

103

time, could point out/ prove the sensing characteristic of AuNPs patterned on PDMS to the biomolecular binding events. This is useful information for the application of this device on label free bio-molecules detection.

Figure 7. The relationship between the peak shift and CRP antigen concentration

IV. RESULTS AND DISCUSSION We successfully demonstrated the development of PDMS

microfluidic device for bio-analytical applications. The use of LSPR phenomena of AuNPs in combination with microfluidic system was successfully monitored the interaction of antigen-antibody. CRP antigen could be found in a very small concentration down to 10ng/ml. With our primarily results on the sensing applicability or bio-molecules, we have demonstrated that the integration of LSPR phenomena of AuNPs with the remarkable advantages of microfluidics can be used to develop a label-free optical sensing device for biomolecule analysis in medical uses and bioscience application.

V. REFERENCES [1] L. He, M. D. Musick, S. R. Nicewarner, F. G. Salinas, S. J. Benkovic,

M. J. Natan, and C. D. Keating: J. Am. Chem. Soc. 122 (2000) 9071 [2] J. C. McDonald, D. C. Duffy, J. R. Anderson, D. T. Chiu, H. Wu, O. J.

A. Schueller, and G. M. Whitesides: Electrophoresis. 21 (2000) 27. [3] D. K. Kim, K. Kerman, S. Yamamura, Y. S. Kwon, Y. Takamura, and E.

Tamiya: Jpn. J. Appl. Phys, 47 (2008)1351. [4] Q. Zhang, J. J. Xu, Y. Liu, and H. Y. Chen: Lab Chip. 8 (2008) 352. [5] Toth, I. Bertoti, M. Blazio, G. Banhegyi, A. Bognar and P. Szaplonczay,

J. Appl. Polym. Sci. 52 (1994) 1293. [6] D. C. Duffy, J. C. McDonald, O. J. A. Schueller, and G. M. Whitesides:

Anal. Chem. 70 (1998) 4974. [7] N. Nath and A. Chilkoti: Anal. Chem, 74 (2002) 504.