edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 36 that Pl and Ml are the most significant channels to classify the materials using the MLP classifier. Using the number of TP, TN, FP, and FN the overall performance parameters were calculated as: accuracy: 92%; precision: 92%; recall: 92%; and F1-score: 92%. The performance parameters of each material are entered in Table 3. The materials ABS, HDPE, PA, PET, and PVC had the highest classification F1-score of 1.00. The lowest F1-score of 0.53 was achieved by the classification of PC, whereby only half the data was correctly assigned. TensorFlow based MLP training yielded almost the same results within a tolerance of 3%. The overall results of the TensorFlow MLP using the layers I, P, M, Pl, and Ml were: accu- racy: 97%; precision: 97%; recall: 97%; and F1-score: 91%. The results using only Pl and Ml were: accuracy: 95%; precision: 96%; recall: 95%; and F1-score: 95%. The results also showed that Pl and Ml have a major significance in perfor- mance parameters. The combination of Pl and Ml is found to be the two channels for a significant classification. CONCLUSION The classification of 14 material samples containing twelve plastic types, tire rubber, and a cigarette filter was done using a data-based MLP approach with Scikit-Learn and TensorFlow based training. The data for the NNs were obtained using an FD-FLIM camera system, which measures five image layers: intensity, phase shift, modulation, phase dependent fluorescence lifetime, and modulation dependent fluorescence lifetime. The results of both NNs seem promising as only the small amount of 45 measurements from each material was used. The calculated performance parameters of data-based classification using an MLP show that the materials can be classified at an F1-score of 94% if all five layers are used for the classification. Additionally, the classification using the MLP at the layers Pl and Ml shows slightly less accuracy with a F1-score of 92%. But as only two layers are used, the combination of Pl and Ml is found to be the most effective, as the time needed for training and testing decreases. The TensorFlow approach shows similar results of the F1-scores. The F1-score using all five layers for the classification is 95%. If the Pl and Ml layer are used, a F1-score of 94% is reached. Thus, the classification only using Pl and Ml is also possible resulting in a high F1-score. In summary, it was successfully demonstrated that materials can be classified by NNs trained, validated, and tested on FD-FLIM measurements. Further research must be done, including the usage of more data to train, validate, and test the networks. REFERENCES 1. Plastics Europe Deutschland e. V., “Plastics – the Facts 2020, Plastics Demand by Segment,” special show of K 2020, Düsseldorf, 2020 2. Heinrich Böll Stiftung, Bund für Umwelt und Naturschutz, “Plastikatlas 2019, Daten und Fakten über eine Welt voller Kunststoffe,” Plastikatlas, 2019, 3, Auflage, Berlin. 3. S. Pieh, et al.: “Identification and Quantification of Macro- and Microplastics on an Agricultural Farmland,” Scientific Reports, Dec. 2018, 8, p. 17950. Fig. 5 Confusion matrix of the Scikit-Learn based MLP classification for 14 material samples using the mean and standard deviation of the layers Pl and Ml. Table 3 Performance parameters for the Scikit-Learn based MLP classification for 14 material samples using the mean and standard deviation of the layers Pl and Ml Material Accuracy Precision Recall F1 score ABS 1.00 1.00 1.00 1.00 Filter 0.99 1.00 0.89 0.94 HDPE 1.00 1.00 1.00 1.00 PA 1.00 1.00 1.00 1.00 PC 0.94 0.67 0.44 0.53 PE 0.97 0.73 0.89 0.80 PET 1.00 1.00 1.00 1.00 POM 0.99 0.90 1.00 0.95 PP 0.99 1.00 0.89 0.94 PS 0.98 0.89 0.89 0.89 PU 0.98 0.89 0.89 0.89 PVC 1.00 1.00 1.00 1.00 Rubber 0.99 0.90 1.00 0.95 SAN 0.99 0.91 1.00 0.95 MULTILAYER PERCEPTRON DEVELOPMENT TO IDENTIFY PLASTICS (continued from page 34)
RkJQdWJsaXNoZXIy MTMyMzg5NA==