intelligent sensors and sensor networks
DESCRIPTION
intelligent sensors and sensor networks: in context of human localization and material classificationTRANSCRIPT
INTELLIGENT SENSORS AND SENSORNETWORKS
SURINDER KAUR2012CS13
M.TECH(1st Year)
July 16, 2013
Department of Computer Science and EngineeringM.N.N.I.T. Allahabad,India
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
INTELLIGENT SENSORS
Sensors that are capable of sensing and transforming thesensed data into a structured symbolic description thatsupports reasoning by artificial intelligence processes.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
NEURAL NETWORK APPROACH
Use of Neural Network in intelligent sensors for decision makingand learning in the context of the following applications:
Human localization
Discrimination of material type using:
Radial Basis Function Neural Network RBFNN.Multi-Layer Perceptron Neural Network MLPNN.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
HUMAN LOCALIZATION-SYSTEM ARCHITECTURE
Figure : System Architecture
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
HUMAN POSITION MEASUREMENT
Two complementary measurement are involved:
Global Measurement System:
Apply Laser Range Finders(LRFs).Measure distance based on time of flight principle by usinglaser.
Specific Measurement SystemSun Small Programmable Object Technology(Sun SPOT) isused.Sun SPOT has 3 sensors:
AccelerometerIlluminance sensorTemperature sensor
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
HUMAN LOCALIZATION
Spiking Neural Network(SNN) is used.
SNNs have the capability of memorizingSpatial contextTemporal context
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
OUTPUT PULSE OF THE NEURON
pi(t) =
{1 if hi (t) ≥ qi0 otherwise
where pi (t) : output pulse of the i th neuron at time thi (t) : internal state of i th neuron at time t
qi : threshold
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
INTERNAL STATE OF NEURON
hi(t) = tanh(hrefi (t) + hsyn
i (t) + hexti (t))
where hrefi (t) : is the refractoriness factor of the neuron at time thsyni (t) : includes the output pulse from the other neurons
at time thexti (t) : input to the i th neuron from the environment at
time t
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
REFRACTORINESS FACTOR OF NEURON
hrefi (t) =
{γref .hrefi (t − 1)− R if pi (t − 1) = 1
γref .hrefi (t − 1) otherwise
where hrefi (t) : is the refractoriness factor of the neuronat time t
γref : discount rateR : constant such that R > 0
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
OUTPUT FROM OTHER NEURONS
hsyni (t) = γsyn.hi(t− 1) +
N∑j=i,j 6=i
wji.hPSPj (t)
where hsyni (t) : the output pulse from the other neurons at time thi (t − 1) : internal state of i th neuron at time t-1
γsyn: temporal discount ratewji : connection weight from j th neuron to i th neuron
hPSPj (t) : presynaptic action potential of j th neuron at time t
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
PRESYNAPTIC ACTION POTENTIAL OF NEURON
hPSPj (t) =
{1 if pj(t) = 1
γPSP .hPSPj (t − 1) otherwise
where hPSPj (t) : presynaptic action potential at time t
γPSP : discount rate, such that 0< γPSP < 1pj(t) : output pulse at time t
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
ENVIRONMENTAL INPUT TO NEURON
hextij (t) = tanh(β.|dij(t)− dLTM
ij |)
where hextij (t) : input to the j th neuron of the i th LRFβ : constant such that 0 < β < 1
dij(t) : current distance value at time tdLTMij : Long-Term-Memory of the distance
It is updated using the following equation:
dLTMij = (1− α).dLTM
ij + α.dij(t)
α : constant such that 0 < α < 1
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
FEATURE POINT EXTRACTION
sij(t) =
1 if pij(t) = 1 or
hPSPij (t) > H or
rUij (t) = 1
0 otherwise
where sij(t) : feature point from i th LRF to j th neuron at time tpij(t) : pulse output of j th neuron on i th LRF at time t
H : thresholdrUij (t) : flag, it is set if the distance of the j th measurement
point form the i th LRF can be measured
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
HUMAN POSITION DETECTION
It uses nearest neighbour approach.Human position detection is based on 10 neighbourhood featurepoints.
gij(t) =
1 if5∑
k=−5si ,j+k ≥ S
0 otherwise
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
HUMAN POSITION DETECTION
Figure : SNN for Human Detection by LRF
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
WEIGHT UPDATION
If
hPSPj (t− 1) < hPSP
i (t)
Then
wji = tanh(γwhtwji + ξwgt.hPSPj (t− 1).hPSP
i (t))
where wji : connection weight from j th neuron to i th neuronγwht : Hebbian discount rateξwgt : learning rate
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
SPECIFIC MEASUREMENT
Each furniture or equipment is attached with sensor.
The difference of current position from base value is input toSNN.
If neuron fires, it means a person uses or moves its corres--ponding furniture.
The firing pattern indicates the time− series of humanposition in the room.
The human position can be approximated to that of the corr--esponding furniture or equipment.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
TIME-SERIES OF HUMAN POSITION IN A ROOM
Figure : Transition of Human Position by SNN
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
CONNECTION WEIGHTS AFTER LEARNING
Figure : Connection Strength after Learning of SNN
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
DISCRIMINATION OF MATERIAL TYPE-SYSTEMARCHITECTURE
Figure : The Sensor with Plunger Based Probe and Optical Mouse
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
NEURAL NETWORK BASED CLASSIFIER
(A) Multi− Layer Perceptron Neural Network(MLP NN)
Feed-Forward Neural Network
Figure : MLP NN
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
MLP-NN
yi = xi
where yi : output of i th neuron in input layerxi : input signal
yj = fa(∑
wij.xi)
where yj : output of j th neuron in hidden layer or output layerfa : the activation functionwij : connection weight from i th neuron to j th neuronxi : the input from i th neuron to j th neuron
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
NEURAL NETWORK BASED CLASSIFIER
(B) Radial Basis Function Neural Network(RBF NN)
Hidden layer applies non-linear transformation from inputspace to hidden space.
The transform function is radial− symmetrical on centrepoint.
In the paper Gaussian function is chosen as basis function.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
RBF NN
yi = xi
where yi : output of i th neuron in input layerxi : input signal
yj =∑
iwi.G(|x− xi|)
where yj : output of j th neuron in hidden layer or output layerwi : connection weightxi : the input from i th neuron in the lower layer and the
centre point.x : the real valued vector
The Gaussian function is given as:
G(z) = exp(−z2/(2 ∗ σ2))
where σ : variance
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
COMPARISON OF MLP NN AND RBF NN
The comparison between the two is made on the basis of followingperformance measures:
Mean Square Error(MSE)
Percentage Classification Accuracy(PCLA)
Area under Receiver Operating Characteristic curve(AROC)
RBF NN is found better than MLP NN on following grounds:
Less training iterations are required.
It is more noise tolerant.
Topology optimization is easy.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
USIC
Universal Sensor Interface Chip
Device that has high degree of in built analogue and digitalflexibility combined with an integrated micro controller.
It includes all of the processing elements needed to producemany intelligent sensor systems.
The local intelligence is provided by the integrated RISCprocessor.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
BLOCK DIAGRAM OF USIC
Figure : Block Diagram of USIC
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
CONCLUSION
The proposed method of on− line human localization basedon SNN in intelligent sensor network can be combined withvoice recognition and visual perception for better naturalinteraction between the human and the robot caregiver.
The proposed material classifier shows satisfactoryperformance and RBF NN is proved better than MLP NN.
The proposed method is able to classify materials, attemptscan be made to classify the surface roughness of differentmaterials.
USIC provides a cost− effective solution to develop variousintelligent sensor applications using a common chip.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS
REFERENCES
T. Obo, N. Kubota, K. Taniguchi and T. Sawayama, “Humanlocalization based on spiking neural network in intelligentsensor networks”,IEEE Workshop on,Robotic Intelligence InInformationally Structured Space,2011.
Nadir N. Charniya, Sanjay V. Dudul, “Intelligent sensorsystem for discrimination of material type using neuralnetworks”, Journal Applied Soft Computing archive Volume12 Issue 1, Pages 543-552, January, 2012.
P. D. Wilson, S.P. Hopkins, R.S. Spraggs, I. Lewis, V. Skardaand J. Goodey, “Application of a universal sensor interfacechip(USIC) for intelligent sensor applications”, in Proceedingsof the IEE Colloquium on Advances in Sensors, no. 232, pp.3/13/6,December 1995.
Dekneuvel, E. and H. Medromi, “An ultrasonic intelligentsensor for a mobile robot perceptron system. Principles,design and experiments,” In IEEE Conference, 1999.
SURINDER KAUR 2012CS13 M.TECH(1st Year) INTELLIGENT SENSORS AND SENSOR NETWORKS