edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 32 Neural networks (NNs) can be classified based on their purposes into MLPs, convolutional neural networks (CNNs), and recurrent neural networks (RNNs). The MLP is a feedforward neural network in which data is collected by an input layer. During the forward propagation, data is transferred to the hidden layers where the learning process happens by calculating the weights and activation functions used for each layer. The resulting error rate is determined in an output layer. During the back propagation, a readjustment of the weights is done corresponding to the activation function and loss functions until the error is reduced to a minimum.[9] MLPs can be programmed using Scikit-Learn or TensorFlow, which are machine learning libraries for Python programming.[10] MLP Classifier is a module of Scikit-Learn and has ready-touse algorithms. It is specifically targeted to classification and works mainly on tabular data.[11] TensorFlow is a flexible machine learning library that can be easily employed to combine probabilistic programming and deep learning modules.[12] The performance of a model is evaluated by the accuracy, precision, recall, & F1-score. True positive (TP), true negative (TN), false positive (FP) and false negative (FN) classification results are used to calculate the performance parameters. A true positive (TP) result describes that the predicted class is correct. True negative (TN) results show that the value of the actual class and the predicted class are false, which leads to a correct assigned classification. A false positive (FP) classification means that the actual class is false and the predicted class is correctly assigned. The last result a classification can give is false negative (FN), which means that the actual class is correct, but the predicted class is false. Using the values of TP, TN, FP, and FN, the accuracy, which is the ratio of correctly predicted classes to the number of total observations, can be calculated: Accuracy = (TP+TN)/(TP+TN+FP+FN) (Eq 1) The precision is the ratio of correctly predicted to the total predicted positive observations: Precision = TP/(TP+FP) (Eq 2) The ratio of correctly predicted positive observations to all observations is the recall: Recall = TP/(TP+FN) (Eq 3) The F1-score gives the weighted average of precision and recall: F1_score = 2*Recall*Precision/(Recall + Precision) (Eq 4) Furthermore, a confusion matrix can be created, giving a graphical overview of the correctly assigned values.[13] EXPERIMENTAL METHODS Fourteen samples were examined including red high-density polyethylene (HDPE) commonly used in plastic bottles and toys, Styrene acrylonitrile resin (SAN) employed in automotive parts and kitchenware, Acrylonitrile butadiene styrene (ABS) utilized in pipes and automotive body covers, polyamide (PA) used in areas of clothing and carpets, polycarbonate (PC), polyethylene (PE) mainly used in packaging films and garbage bags, polyethylene terephthalate (PET) employed in food packaging and beverages, polyoxymethylene (POM), polypropylene (PP) widely utilized in ropes and carpets, polystyrene (PS) used in refrigerators and air conditioners, polyurethane (PU) used in furniture and thermal insulations, polyvinyl chloride (PVC) mainly employed in pipes and fittings; along with rubber from tire (rubber) and cigarette filters (filter). Tire rubber was included due to its abundance and contribution of microplastics in the environment.[14] Cigarette filters which contain cellulose acetate were chosen as it poses a serious threat to animals and the environment.[15] The polymer samples were mostly granules without any additives or impurities. Additionally, a red HDPE from a Euronorm E2 box was used. An example for the sample ABS and SAN is given in Fig. 1. EXPERIMENTAL SETUP AND PROCEDURE The experimental setup used in references 6 and 7 was used for the investigations (Fig. 2). The setup con- sists of a laser diode, optical filters, a microscope, and a FD-FLIM camera. The LaserNest laser diode from Omicron Laserage GmbH has an excitation wavelength of 405 nm and an optical output power of 300 mW. The excitation light is guided by a liquid light guide into a PSM 1000 microscope from MOTIC. The FD-FLIM camera, a pco.film from Excelitas PCO GmbH, is connected to the microscope via a C-Mount. The Fig. 1 SAN (left) and ABS (right) plastic granules served as samples.
RkJQdWJsaXNoZXIy MTMyMzg5NA==