presented at njdot quarterly meeting, january 9, 2015
TRANSCRIPT
LASER SCANNING AGGREGATES FOR REAL-TIME PROPERTY IDENTIFICATION
Andrew Branin, Bless Ann Varghese, Dr. Michael Lim, Dr. Ravi Ramachandran and
Dr. Beena Sukumaran
Presented at NJDOT Quarterly Meeting, January 9, 2015
Background
Use of unacceptable aggregates such as mica schist and carbonate rocks can reduce the quality and
life of roadway pavement.
Currently used analysis techniques such as XRF analysis for
determination of mineralogy do not provide real-time data.
Objectives of the Study Obtain laser spectra models for various
aggregate sources from New Jersey Calibrate laser-spectra models to identify real-
time aggregate properties such as mineralogy Determine the feasibility of laser technology as
a portable tool for identification of real time aggregate properties such as mineralogy and particle morphology
Determine the feasibility and affordability of adapting laboratory based laser technology applications for field use
Approach
LIBS
Experimental Procedure
Spectrum Preprocessing
PLS Analysis
Determine Chemical
Composition
Approach
LIBS
Experimental Procedure
PLS Analysis
Spectrum Preprocessing
Determine Chemical
Composition
Overview of LIBS(Laser-Induced Breakdown Spectroscopy)
http://www.arl.army.mil/www/default.cfm?page=247http://industrial-lasers.net/yag.html LIBS Handbook
Previous Geological Applications
NY State DOT used LIBS to:1. Determine Acid
Insoluble Residue (AIR) in an aggregate sample.
2. Determine the percent noncarbonated stone in an aggregate blend.
PLS Analysis technique employed for data analysis
(Chesner 2012)
Previous Geological Applications KSDOT
LIBS was used to:1. Predict D-cracking
likelihood: pass/fail.2. Determine the
aggregate source. PCA and PLS
analysis technique used
(Chesner 2012)
Previous Geological Applications TXDOT
LIBS was used to:1. Determine the
amount of chert.2. Predict result of
state testing.3. Differentiate
cherts. PCA and PLS
Analyses(Chesner 2012)
Other Noteworthy Applications
Quality control for: Cement powder composition Concrete repair by modeling chloride and
sulphur contamination at varying depth Recycling demolished concrete
Analysis of micro-cracks in surfaces
Approach
LIBS
Experimental Procedure
Spectrum Preprocessing
PLS Analysis
Determine Chemical
Composition
Current Experimental Setup
1. Nd: YAG Laser (Brilliant B)2. Control Pad (flash lamp, Q-switch)3. Sample Chamber4. Applied Spectra Spectrometer5. Control Unit6. Laptop
1. Nd: YAG Laser (Brilliant B)2. Mirror3. Focusing Lens4. Beam Splitter (not currently used)
2 1
3
5
4
6
1
2
3 4
1. Adjustable Sample Stage2. Spectral Emission Redirection Mirror3. Sample Tray4. Focal Point Indicator
3
2
1
4
Current Experimental Setup
Additional Notes
The beam splitter is no longer used. System timing has been adjusted so that
more laser energy is used with a shorter spectrometer delay.
A fresh battery of tests were performed following these adjustments.
Comments on Preliminary Tests
Relative light intensities were generally relatively consistent as long as sufficient pulse energy was provided.
Measured on cooler plasma – it was assumed that only neutral atoms were present.
Testing had suggested that a 50 mJ laser pulse was optimal (nearly 100 mJ are presently used).
The addition of the beam splitter and use of the higher pulse energy reduced the frequency of low-emission shots, but resulted in a less focused beam.
Current Model Testing
Models were calibrated using new data from 5 locations per sample, for each of 10 samples, for each of 10 stone types.
Each spectrum was the sum of the emissions from 100 shots to mitigate shot to shot variation.
Approach
LIBS
Experimental Procedure
Spectrum Preprocessing
PLS Analysis
Determine Chemical
Composition
Analysis
Data Pre-processing: Subtract baseline (done automatically by software) Remove negative intensity values Normalize to total light emission (where applicable) Various other techniques were used (detailed in Results
section) Base model Y-scaling Averaged calibration set Amplitude scaling Spectral Derivatives Split training
Number of PLS components was optimized for each model via a built-in cross validation function.
Approach
LIBS
Experimental Procedure
Spectrum Preprocessing
PLS Analysis
Determine Chemical
Composition
PLS Regression Analysis Partial Least Squares Regression Analysis (PLS) analysis has been
used to develop models to predict concentrations of compounds within stone samples.
PLS Analysis can be used to generate predictive models based on single values corresponding to an entire spectrum
Predictions are made in a manner similar to Multiple Linear Regression, but coefficients are determined by maximizing covariance between X data and known Y values rather than by minimizing square error.
PLS can also be used to differentiate different types of samples, but only for linear relationships (PCA is generally used for classification)
Chesner 2012
Approach
LIBS
Experimental Procedure
PLS Analysis
Spectrum Preprocessing
Determine Chemical
Composition
Previous Testing
Initial Testing: Metal samples to confirm that LIBS could be used
to qualitatively identify elements in samples Mica and limestone samples to observe output
spectra of two of the target minerals Various coins of known composition to develop a
predictive model for a simplified case Preliminary Testing and Models:
Obtaining output spectra from aggregate samples Calibrating and testing initial predictive models
Preliminary Predictive Models Early models were very inaccurate, and
were calibrated using 27 PLS components, which was later determined to be excessive.
Models may have been weak due to a lack of variation in calibration data, flaws in system timing, sub-optimal data pre-processing, or a combination of factors.
Notes on Accepted Values Testing
DOT XRF results were expected to be, and were found to be more reproducible due to tests being performed on powdered samples.
DOT results continued to be used as accepted values.
Prior to the most recent testing, the experimental setup was adjusted, cleaned, and realigned to produce more reliable spectrum data.
Includes Discussion of Data Pre-Processing Techniques
Current Test Results
Base Model
Negative values were removed, but center clipping was not otherwise applied.
Spectra were normalized to a metric of total light emission prior to calibration and testing.
No other adjustments were made. All other models were compared to this
baseline model.
Y-Scaling, ratio:1
Identical to Base Model, but Y variables (known concentrations) were scaled by dividing the values for each compound by the maximum in the calibration set.
The reverse adjustment is applied to predicted results to produce a prediction.
This forces PLS Regression to consider all compounds with equal priority.
Y-Scaling, 0:1
Identical to the other Y-Scaling method, however the minimum value of each compound is first subtracted from each, before each value is divided by the range for the compound’s accepted values.
The reverse adjustments are again applied to predicted values.
This is simply an alternative method of the previous principle.
Y-Scaling Results
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
Carbonate Dolomite
XRFBase Model UpperBase Model Lowerratio:1 Upperratio:1 Lower0:1 Upper0:1 Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
80
Woodboro Carbonate
XRFBase Model UpperBase Model Lowerratio:1 Upperratio:1 Lower0:1 Upper0:1 Lower
% C
om
posit
ion
Y-Scaling Results
SiO2 Al2O3 Fe2O3 CaO MgO
-20
-10
0
10
20
30
40
50
60
70
Gneiss
XRFBase Model UpperBase Model Lowerratio:1 Upperratio:1 Lower0:1 Upper0:1 Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO
-20
-10
0
10
20
30
40
50
60
Argillite
XRFBase Model UpperBase Model Lowerratio:1 Upperratio:1 Lower0:1 Upper0:1 Lower
% C
om
posit
ion
Averaged Calibration Set
Identical to the Base Model, however the spectra for each type of stone are first averaged together, resulting in a single resultant spectrum for each stone type.
Testing uses individual spectra or averaged testing data.
Averaged Calibration Results
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
Carbonate Dolomite
XRFBase Model UpperBase Model LowerAveraged UpperAveraged LowerAveraged Test
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
80
Woodboro Carbonate
XRFBase Model UpperBase Model LowerAveraged UpperAveraged LowerAveraged Test
% C
om
posit
ion
Averaged Calibration Results
SiO2 Al2O3 Fe2O3 CaO MgO-10
0
10
20
30
40
50
60
70
Gneiss
XRFBase Model UpperBase Model LowerAveraged UpperAveraged LowerAveraged Test
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO-10
0
10
20
30
40
50
60
70
Argillite
XRFBase Model UpperBase Model LowerAveraged UpperAveraged LowerAveraged Test
% C
om
posit
ion
Amplitude Scaling
As the Base Model, but spectra are not normalized, and instead each spectrum for a certain type of stone is scaled in amplitude relative to the average light emission for that stone type.
This was done in an attempt to normalize spectra without removing information on varying overall amplitude.
Amplitude Scaling Results
SiO2 Al2O3 Fe2O3 CaO MgO
-100
-50
0
50
100
150
Carbonate Dolomite
XRFBase Model UpperBase Model LowerAmp Scale UpperAmp Scale LowerScaled Test UpperScaled Test Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO
-100
-50
0
50
100
150
200
Woodboro Carbonate
XRFBase Model UpperBase Model LowerAmp Scale UpperAmp Scale LowerScaled Test UpperScaled Test Lower
% C
om
posit
ion
Amplitude Scaling Results
SiO2 Al2O3 Fe2O3 CaO MgO-10
0
10
20
30
40
50
60
70
Gneiss
XRFBase Model UpperBase Model LowerAmp Scale UpperAmp Scale LowerScaled Test UpperScaled Test Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
Argillite
XRFBase Model UpperBase Model LowerAmp Scale UpperAmp Scale LowerScaled Test UpperScaled Test Lower
% C
om
posit
ion
Spectral Derivatives
An approximation of the derivative of each unadjusted spectrum is used for calibration and testing in place of the spectra themselves.
This was done in an attempt to consider the slope trends in the spectra rather than amplitudes.
Spectral Derivative Results
SiO2 Al2O3 Fe2O3 CaO MgO-10
0
10
20
30
40
50
60
70
80
90
Carbonate Dolomite
XRFBase Model UpperBase Model LowerDerivative UpperDerivative Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
80
Woodboro Carbonate
XRFBase Model UpperBase Model LowerDerivative UpperDerivative Lower
% C
om
posit
ion
Spectral Derivative Results
SiO2 Al2O3 Fe2O3 CaO MgO
-40
-20
0
20
40
60
80
Gneiss
XRFBase Model UpperBase Model LowerDerivative UpperDerivative Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
Argillite
XRFBase Model UpperBase Model LowerDerivative UpperDerivative Lower
% C
om
posit
ion
**Split Training Sets**
Training Data was divided into carbonate and non-carbonate rocks.
Separate models were generated for each subset to narrow the range of expected values within each model.
In a finished product, a broad-base model would be used as a preliminary estimate before a more specialized model would be used for actual prediction.
Split Training Set Results - Carbonate
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
Carbonate Dolomite
XRFBase Model UpperBase Model LowerSplit Training UpperSplit Training Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
70
80
Woodboro Carbonate
XRFBase Model UpperBase Model LowerSplit Training UpperSplit Training Lower
% C
om
posit
ion
Split Training Set Results Non-Carbonate
SiO2 Al2O3 Fe2O3 CaO MgO-10
0
10
20
30
40
50
60
70
Gneiss
XRFBase Model UpperBase Model LowerSplit Training UpperSplit Training Lower
% C
om
posit
ion
SiO2 Al2O3 Fe2O3 CaO MgO0
10
20
30
40
50
60
Argillite
XRFBase Model UpperBase Model LowerSplit Training UpperSplit Training Lower
% C
om
posit
ion
Portable LIBS laser and laser for morphology characterization
Equipment Acquisition
Portable Units
A Quantel ULTRA U1064E100R020LN compact laser was purchased following comparison of comparable systems.
This laser will be installed and results compared to previous tests to ensure accuracy with the portable system.
Morphology Characterization Two techniques considered:
Pulsed Digital Holography
Optical Coherence Tomography
Pulsed Digital Holography
Using a pulsed laser as the source of coherent light.
Three-dimensional size and shape information is encoded in the interference pattern produced.
Computer encodes images from interference pattern.
Strength: simplicity and robustness of the hardware implementation.
Limitation: image reconstruction is computationally intensive.
Experimental setup for Pulsed Digital Holography
Optical Coherence Tomography
Based on interferometry principle Interferometer output is analyzed with a
grating spectrometer. Strength:
high-resolution depth information for the scattering surface.
commercial system is available. Team is currently discussing with Thorlabs
regarding the technical capabilities of several OCT systems available for purchase.
Conceptual Depiction of OCT method
Conclusions and Future Work
Previous research has shown that LIBS is feasible as a means to quantify chemical composition of aggregate.
By dividing the training set into carbonate and non-carbonate stones, prediction accuracy and reliability improved considerably.
Future testing will include further refining of predictive models; building on the split training set strategy.
Future testing will also expand the training set and optimize the testing set size, before moving on to testing unknown samples and field tests.
The model can be calibrated to predict other sample traits. Research has begun into measuring particle morphology,
equipment will be selected, and a standard procedure developed.
References
Chesner, Warren and McMillan, Nancy. (2012). “Automated Laser Spectrographic Pattern Matching For Aggregate Identification.” Highway IDEA Program.
Cremers, D. A. and Radziemski, L. J. (2006). “Handbook of Laser-Induced BreakdownSpectroscopy.” John Wiley & Sons, Ltd.
Mansoori, A. et. al. (2011) “Quantitative analysis of cement powder by laser induced breakdown spectroscopy.” Optics and Lasers in Engineering. Vol. 49, Issue 3, 318-323.
Nelson, Stephen. (2013). “Mineral Chemistry.” Tulane University. <http://www.tulane.edu/~sanelson/eens211/mineral_chemistry.htm> Oct. 25, 2013.
Pasquini, Celio et. al. (2007). “Laser Induced Breakdown Spectroscopy.” J. Braz. Chem. Soc., Vol. 18, No. 3, 463-512.
Taffe, A. et. al. (2009) “Development of a portable LIBS-device for quality assurance in concrete repair.” Concrete Repair, Rehabilitation and Retrofitting II. Taylor and Francis Group, London.
Tucker, J.M. et al. (2010) “Optimization of laser-induced breakdown spectroscopy for rapid geochemical analysis.” Chemical Geology. Vol. 277, Issues 1-2. 137-148.
Wessel, W. et. al. (2010) “Use of femtosecond laser-induced breakdown spectroscopy for micro-crack analysis on the surface.” Engineering Fracture Mechanics. Vol. 77, 1874-1883.
Xia, H. and M.C.M. Bakker. “Online Sensor System Based on Laser Induced Breakdown Spectroscopy in Quality Inspection of Demolition Concrete.” Delft University of Technology, the Netherlands.
Acknowledgments
Jr/Sr clinic students Eric Seckinger Saima Mahmud Christine Neppel Joshua Edwards
Questions