Applications of hyperspectral
imaging and machine learning
methods for real-time
classification of waste stream
components
Martin Sevcik1 Jan Skvaril1
Elena Tomas Aparicio2
1Future Energy Center, Mälardalen University, Västerås, Sweden
2Mälarenergi AB, Västerås, Sweden
17th September 2019
19th International Council of Near Spectroscopy Conference, Gold Coast, Australia
Would you like to discover the inner soul of the materials and processes through exciting real-time analytical techniques?
JOIN OUR NEW ONLINE COURSE AT MÄLARDALEN
UNIVERSITY!
Applied Spectroscopy for Future Energy and Environmental Systems https://www.mdh.se/utbildning/livslangtla rande/futuree/applied-spectroscopy-for- future-energy-and-environmental-systems-1.118583?l=en_UK Information (www)
Introduction
● Combined heat and power production
● Municipal solid waste → Refuse-derived fuel (RDF) ● Fuel is highly variable: → Process instabilities
● Increase formation emissions
● Lower efficiency CONTROL SYSTEM Steam to the turbine Flue gas
for further treatment
Ash
used bed material
Steam boiler
Circulating fluidized bed
Refuse derived fuel
Composition Moisture content, Ash content, Heating value, Other properties Paper and carbord 28% Others 9% Textiles 9% Cellulose 6%
Introduction
CONTROL SYSTEM Steam to the turbine Flue gasfor further treatment
Ash
used bed material
Steam boiler Circulating fluidized bed Single-point NIR sensor Real-time measurement of fuel properties
Refuse derived fuel
Composition Moisture content, Ash content, Heating value, Other properties
● Previously installed single-point NIR sensor → Provide good conditions for feed-forward control
CONTROL SYSTEM
Steam
to the turbine
Flue gas
for further treatment
Ash
used bed material
Steam boiler Circulating fluidized bed Single-point NIR sensor Automatic pre-sorting? Hyperspectral imaging NIR sensor? Real-time identification of components and measurement of fuel properties
Refuse derived fuel
Composition Moisture content, Ash content, Heating value, Other properties ← Further upstream
Introduction
Spectroscopy workflow
Samples (training and internal validation) Samples (external validation set) NIR-HSI acquisition Hypercube unfolding and spectra extraction Data pre-treatment (various methods) Machine learning classification algoritms PLS-DA, SVM, RBNN 1000 pixels training set 1000 pixels internal validation set Model Identification of waste material components by reference method NIR-HSI acquisition Hypercube unfolding and spectra extraction Data pre-treatment (various methods) Prediction on external validation data set(whole image) False colour images Model performance evaluation Internal validation set SENSITIVITY 1 - SPECIFICITY Overall accuracy Model performance evaluation external validation set SENSITIVITY 1 - SPECIFICITY Overall accuracy Prediction on internal validation data set Computational time Selected pre-treatment Data pre-treatment (various methods) Training set - 1000 pixels
per material category
Internal validation set – 1000 pixels per material category
Sampling and sample praparation
Category Number of
sample objects Examples of the samples
1. Paper 22
Carboard boxes, paper tissues, napkins, printing paper, notebooks, recycled paper, newspapers, magazines
2. Incombustibles 12 Ceramics, cups, stones, glass bottles, metals 3. Food 7 Dried fruits, pasta, rice
4. PE 12 HDPE – shampoo bottle, bottle capsLDPE – plastic bags, food packages 5. PET 4 Plastic bottles, food containers 6. PP 5 Food containers, straws 7. PS 5 Expanded foam, plastic cutlery
8. PVC 2 Tubes
9. Textile 4 Old clothes, curtains
10. Wood 6 Wood chips
11. Background - Carbon black styrene-butadiene rubber
● Sampling from the RDF and other sources…
PET HDPE Textile PS PP PVC Paper LDPE
Glass Paper hygienic
Ceramics Cardboard
Wood Print recycled
Print white Food
● Samples identified by FT-IR Spectrometer in ATR mode, followed by library search:
● Common materials
● Polymers, Polymer Additives & Plasticizers
● NIR-HSI Spectral data push-broom camera
● 900 - 1700 nm, InGaAs detector,
● 224 spectral bands, 640 spatial pixels
● Framerate 300 fps, 5 ms exposure time
● Approx. 1200 frames per image
● Two linear halogen light sources
● Laboratory scanning table 20 cm x 40 cm
● 90 mm/s, translational movement Scanner table translational movement Sensing area width approx. 18 cm Hyperspectral imaging camera (900 – 1700 nm) Illumination source 2 Illumination source 1 NIR-HSI DATA ACQUISTION Samples
Data acquisition
ATR Mid IR spectraData processing and chemometrics
● Decomposition of the hyperspectral cube, calculations, pre-processing and modelling were done using scripts programmed in MATLAB
● Removing bands based on signal-to-noise ratio → 192 bands
● The training data set and internal validation set included different randomly selected data points (i.e. pixels), 1000 per material category
● Relative absorbance - converted to relative absorbance based on raw reflectance image from the sample, dark reference image and white reference image.
● Spectral data pre-processing
● Mean centering (MC),
● Standard normal variate (SNV),
● Savitzky-Golay smoothening (SGS, 2nd order polynomial, 11 smoothing points)
Chemometrics and model evaluation
● Classification algorithms used:
● Partial least-squares discriminant analysis (PLS-DA),
● Support vector machine (SVM),
● Radial-basis neural network (RBNN),
● Model performance evaluation:
● Sensitivity (true positive rate)
● 1- specificity (false positive rate)
● Parameters are based on whether prediction is a true positive (TP), true negative (TN), false positive (FP), or false negative (FN).
● Calculated for each class separately. Overall classification accuracy
of the model is then expressed as a mean value of sensitivities for all classes.
Predicted class A B T ru e cl ass A TP (true positive) FN (false negative) B FP (false positive) TN (true negative)
Results and discussion
● FT-IR analyses reveals that samples in each group contained various amounts of plasticizers and
other additives making the identification challenging.
● Data pre-processing has significant impact on resulting overall accuracy of classification models.
Pre-processing method PLS-DA SVM RBNN
No pre-processing 64% - 90.6%
SGS 64.3% - 90.4%
SGS + mean centering 74.7% 92.4% 89.9%
SGS + SNV 88% 92.6% 89.6%
SGS + 1stderivative 85.1% 93.8% 90.7%
SGS + 1stderivative + mean centering 85.8% 94% 90.3%
SGS + 1stderivative + SNV 83.2% 91.5% 88.3%
SGS + 2ndderivative 84.4% 91.1% 88.8%
SGS + 2ndderivative + mean centering 84.9% 91.5% 88.6%
SGS + 2ndderivative + SNV 82.9% 87.2% 81.8%
4 best pre-treatments used in prediction on external validation set
Results and discussion
Class
PLS-DA SVM RBNN
Sensitivity 1 - Specificity Most confused with Sensitivity 1 - Specificity Most confused with Sensitivity 1 - Specificity Most confused with Backgr. 88.3% 4.0% incombustibles 97.8% 2.2% - 98.9% 4.0%
-Paper 74.3% 0.3% textile 97.5% 0.5% - 95.0% 1.3%
-Incomb. 60.9% 2.5% background 78.7% 3.0% background 64.4% 1.8% background
Food 98.4% 1.6% - 98.7% 0.0% - 96.4% 0.0%
-PE 89.5% 0.2% - 86.0% 0.2% incombustibles 85.3% 0.1%
-PET 97.3% 0.1% - 94.5% 0.0% - 91.6% 0.0% background
PP 90.4% 0.5% incombustibles 95.1% 0.1% - 92.2% 0.2%
-PS 73.7% 0.4% incombustibles 85.0% 1.0% incombustibles 86.0% 2.6% background
PVC 99.2% 0.1% - 97.9% 0.0% - 95.2% 0.0%
-Textile 97.8% 2.1% - 97.8% 0.1% - 94.7% 0.3%
-Wood 98.2% 1.3% - 98.2% 0.1% - 97.9% 0.0%
-● Detailed classification results for best preprocessing for each of the classification methods -internal validation set
Wood
Results and discussion
● Prediction of classes by PLS-DA model (external validation set)
Class
PLS-DA
Sensitivity 1 - Specificity Most confused with Backgr. 86.0% 7.1% -Paper 78.7% 4.2% textile Incomb. 34.8% 1.6% paper Food 98.0% 0.2% -PE 75.6% 0.3% background PET 66.0% 0.1% background PP 88.5% 0.3% background PS 91.9% 3.5% -PVC 74.9% 0.1% background Textile 97.7% 5.3%
-Pre-processing method Accuracy Classification time
SGS + SNV 77.3% 1.9 s SGS + 1stderivative 81% 2.1 s SGS + 1stderivative + MC 80.9% 1.9 s SGS + 1stderivative + SNV 75.7% 1.9 s Background Paper Incombustibles Food PE PET PP PS PVC Textile PET HDPE Textile PS PP PVC Paper LDPE
Glass Paper hygienic
Ceramics Cardboard
Wood Print recycled
Print white Food Metal
Results and discussion
● Prediction of classes by SVM model (external validation set)
Class
SVM
Sensitivity 1 - Specificity Most confused with Backgr. 28.7% 0.8% PS Paper 98.6% 2.5% -Incomb. 78.9% 8.8% PS Food 99.0% 0.2% -PE 86.4% 1.6% incombustibles PET 80.1% 0.2% PS PP 93.7% 0.6% -PS 94.5% 17.6% -PVC 86.1% 0.2% incombustibles PET HDPE Textile PS PP PVC Paper LDPE
Glass Paper hygienic
Ceramics Cardboard
Wood Print recycled
Print white Food Metal Background Paper Incombustibles Wood PET PP PS
Pre-processing method Accuracy Classification time
SGS + SNV 83.7% 228 s
SGS + 1stderivative 84.9% 219.8 s
SGS + 1stderivative + MC 85.1% 202.7 s
15
Results and discussion
PET HDPE Textile PS PP PVC Paper LDPE
Glass Paper hygienic
Ceramics Cardboard
Wood Print recycled
Print white Food Metal
● Prediction of classes by RBNN model (external validation set)
Class
RBNN
Sensitivity 1 - Specificity Most confused with Backgr. 79.0% 4.5% PS Paper 95.0% 3.0% -Incomb. 58.3% 3.0% PS Food 96.4% 0.1% -PE 82.5% 0.3% incombustibles PET 77.9% 0.3% background PP 89.1% 0.2% background PS 97.1% 7.8% -PVC 77.3% 0.1% background Textile 82.4% 0.4% paper Background Paper Incombustibles Food Wood PE PET PP PS PVC Textile
Pre-processing method Accuracy Classification time
SGS + SNV 75.4% 14.9 s
SGS + 1stderivative 83.5% 17.7 s
SGS + 1stderivative + MC 84.6% 18.4 s
0,10 0,08 0,06 0,04 0,02 0 -0,02 -0,04 -0,06 -0,08 0,12 PE PET PP PS PVC R egre ss ion coeffi ci ent s (-) 0,04 0,02 0 -0,02 -0,04 Background Paper Incombustibles Food Textile Wood
Most influential bands for classification
● Regression coefficients in PLS-DA model
● Pre-processing: SGS + SNV
● Most influential range for synthetic organic
polymers: ● 1100 -1250 nm
● 1350 -1450 nm.
● Most influential regions for classification of
food, textile, paper and wood: ● 1450 -1490 nm. R egres sio n coeffi cie nt s (-)
Conclusions
● NIR-HSI shows good potential for development of a classification model that can recognize most common components of MSW→RDF.
● Models based machine learning methods can successfully classify NIR-HSI pixels belonging to textile, wood, food, paper, and most common types of plastics with reasonable accuracy.
● Classification of materials such as metals, glass or ceramics i.e. not NIR-active materials is more challenging.
● SVM based model - accuracies of up to 94% for internal val., 85% for external val. and was the best in
identifying incombustible materials, but highly demanded on computational time.
● RBNN based model - accuracy 91% for internal val. and 85% for external val., 10 times faster than SVM ● PLS-DA based model - accuracy 88% for internal val. and 81% for external val., 100 times faster than SVM. ● Applicability of the developed models can be further enhanced by including new samples of varying composition
Thank you
17th September 2019
19th International Council of Near
Spectroscopy Conference, Gold Coast, Australia
Jan Skvaril
Email: jan.skvaril@mdh.se
Mobile: +46-73-6620977 Future Energy Center
Mälardalens University, Västerås