edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 34 standard deviation of respective layers. Initially, the data of mean and standard deviation of each layer is imported to python from the CSV file as pandas data frame. The Label Encoder from Scikit-Learn is used to convert the names of the samples into numerical values. The “Standard Scalar” from Scikit-Learn is used to normalize the numerical data. In both cases the data is split into training (80%) and test data (20%). After splitting, the function “MLPClassifier” is used in Scikit-Learn including the activation function “relu” and the solver “adam.” Additionally, the hidden layer sizes, batch size, learning rate, and number of iterations are defined and early stopping is set to “true” in order to avoid overfitting. In the TensorFlow based approach, each hidden layer is specified with the “relu” activation function except for the outermost layer, which is specified with the “softmax” activation function due to multiple classification. All other parameters are equally defined as using the MLP Classifier from the Scikit-Learn approach. After setting the parameters of the two MLPs, 31 different layer combinations of the extracted mean values and standard deviation are trained. To evaluate the classification results of the MLP, the performance parameters were calculated. RESULTS Figure 4 shows the confusion matrix of the Scikit-Learn MLP Classifier using the combination of the mean and standard deviation from the five layers I, P, M, Pl, and Ml in combination to classify the samples. As a result, the performance of the network can be inferred directly from the confusion matrix by an apparent observation. Hereby, the diagonal elements show the correctly predicted materials. From the confusion matrix it was evident that polymers ABS, filter, HDPE, PET, PU, PVC, rubber, and SAN were all identified correctly. For example, using the material PC, TP =8, FP = 3 (sum of elements in orange color), FN = 1 (sum of elements in red color) and TN = 114 (Sum of all elements – (TP+FP+FN)). Using TP, FP, FN, and TN of each material, the sample specific performance parameters of accuracy, precision, recall, and F1-score were derived (see Table 2). Additionally, the overall performance of the network was calculated as: accuracy: 94%; precision: 94%; recall: 94%; F1-score: 94%. In Fig. 5, the confusion matrix of the Scikit-Learn MLP Classifier approach using only the layers Pl and Ml is shown. Looking at the diagonal assignments of the confusion matrix, it can be obtained that ABS, HDPE, PET, PVC, rubber, and SAN were all assigned correctly. Nevertheless, if the confusion matrix in Fig. 4 is compared to the confusion matrix in Fig. 5, a slight reduction in performance can be observed. But, using only Pl and Ml to classify the materials reduces the amount of data used for training and testing to approximately 40%, which makes the network more efficient. Additionally, the reduction of data shows Table 2 Performance parameters the ScikitLearn based MLP classification for 14 material samples using the mean and standard deviation of the layers I, P, M, Pl, and Ml Material Accuracy Precision Recall F1 score ABS 1.00 1.00 1.00 1.00 Filter 1.00 1.00 1.00 1.00 HDPE 1.00 1.00 1.00 1.00 PA 0.98 0.80 0.89 0.84 PC 0.97 0.73 0.89 0.80 PE 0.98 0.89 0.89 0.89 PET 1.00 1.00 1.00 1.00 POM 0.98 0.88 0.78 0.83 PP 0.98 0.88 0.78 0.83 PS 0.99 1.00 0.89 0.94 PU 1.00 1.00 1.00 1.00 PVC 1.00 1.00 1.00 1.00 Rubber 1.00 1.00 1.00 1.00 SAN 1.00 1.00 1.00 1.00 Fig. 4 Confusion matrix of the Scikit-Learn based MLP classification for 14 material samples using the mean and standard deviation of all layers. (continued on page 36)
RkJQdWJsaXNoZXIy MTMyMzg5NA==