Special Issue on ISNNM 2022 - Integrated Computer-Aided Process Engineering
Classification of Surface Fracture in Plastics Using Convolutional Neural Networks
2023 Volume 64 Issue 9 Pages 2191-2195
Details
2023 Volume 64 Issue 9 Pages 2191-2195
In the present study, we investigate the use of convolutional neural network (CNN) models for classifying the characteristics of surface fractures in plastics, which are affected by environmental stress cracking agents. Nineteen CNN models with different architectures are adopted with 4,012 crack images, and they are evaluated based on the classification accuracy. Four models with a relatively higher accuracy are selected and compared with each performance metric obtained from a confusion matrix. The model with the Inception-ResNet-v2 architecture showed the highest performance metrics value of over 0.96. Although the model with the ResNet-18 architecture showed slightly lower levels of performance metrics, its training time was more than 10 times faster.
Fig. 2 Classification accuracy of 19 models versus training time.
Recently, to address the environmental issues caused by CO2 emissions, automobile manufacturers have been attempting to reduce the weight of their vehicles and increase fuel efficiency. To this end, there has been an increased use of lightweight materials, such as high-strength engineering plastics. The advantages of using plastic materials include improved fuel efficiency, free design, reduced weight, and reduced manufacturing costs. Plastic can absorb four to five times more energy upon impact than conventional materials, which helps to increase stability while being lighter than conventional bumpers. In addition, a thin plastic layer such as polyvinyl butyral (PVB) or ethylene vinyl acetate copolymer (EVAC) sandwiched between the two glass layers makes the car glass harder while simultaneously blocking infrared rays.
Generally, most materials, including plastic materials, are destroyed by the formation and propagation of microcracks under various conditions. Among these, environmental stress cracking (ESC) occurs when a material is exposed to a chemical environment and is destroyed by stress.1) This is a process similar to stress corrosion cracking (SCC) in metals, in which crack formation and propagation occurs. A common ESC mechanism is known as an open vacancy, which occurs because of chemical penetration or the unwinding of polymer chains owing to the creation of sites. Henry2) reported on various variables that make glassy plastics vulnerable to ESC. Breen and Van Dijk3) investigated the ESC effect using air and natural gas with varying benzene content in polyvinylchloride (PVC) and PVC blended with 10% chlorinated polyethylene. Ayyer et al.4) studied the effects of ESC agents on the fatigue and creep properties of polyethylene. Chen et al.5) investigated the variation in the mechanical properties according to the change in ultra-high molecular weight polyethylene content by performing ESC resistance experiments. Gobetti and Ramorino6) presented an experimental method to predict the ESC resistance of recycled polyethylene. Contino et al.7) studied the ESC of polyethylene under plane stress conditions using the J integral method.
While ESC-related studies in existing plastic materials focus primarily on changes in the microstructure and mechanical properties, there is limited research on macrostructure differences from an industrial perspective. Recently, studies related to various crack observations using convolutional neural networks (CNNs) have been reported.8–13) However, most of the studies are related to concrete cracks, and to the best of our knowledge, no complete CNN studies that focus on crack characteristics in plastics have been reported. In this study, we present a CNN model that can distinguish macroscopic fracture surfaces of plastics depending on whether ESC is applied or not. By applying various CNN models, we aim to find a model that can most efficiently classify the characteristics of plastic fracture surfaces using ESC.
In this study, 1870 plastic fracture images caused by mechanical cracks and 2142 plastic fracture images caused by ESC were obtained from the experimental samples. The size of the fracture image is 256 × 256 pixels. The original RGB color was preprocessed into a grayscale image. Image data related to mechanical cracking (hereinafter referred to as ‘MEC’) and image data related to ESC (hereinafter referred to as ‘ESC’) were expanded to 14960 and 17136 datasets, respectively, by the augmentation of rotation and inversion. A detailed method is referred to the Ref. 14).
Figure 1 shows representative images of the MEC data set and the ESC data set. Mechanical cracks caused by a physical external force exhibit sharp crack characteristics, while crack images caused by ESC exhibit mixed shapes distinct from mechanical crack characteristics. Differences in these characteristics are discernible with the naked eye, but it is sometimes difficult for humans to accurately distinguish them. For example, it could be difficult for non-expert to distinguish a kind of fracture image of plastics. In this case, machine learning-based models are useful.
Representative images of (a) MEC and (b) ESC dataset.
In this study, a model was built by importing 19 CNN model structures introduced in MATLAB Deep Network Designer (Deep Learning Toolbox 14.5, R2022b).15) A model was built by directly training 32,096 MEC+ESC dataset images without using a pre-trained neural network framework. The learning optimizer used sets ADAM to an initial learning rate of 0.001. The maximum epoch for model comparison was set to five. Training simulations were performed on an AMD Ryzen Threadripper PRO 5945WX 4.10 GHz CPU processor and in an NVIDIA GeForce RTX 3080 Ti 12GB GPU environment. Of the entire data set, image data were randomly used as follows: 70% for training, 10% for verification, and 20% for testing.
Figure 2 is a chart showing the time required to train image data and the overall prediction accuracy for 19 classification models. Table 1 summarizes the specific structure, calculation results, and calculation time of each model. The AlexNet model with the least eight layers required approximately 15 min to train and showed a classification accuracy of 0.8573. The SqueezeNet model recorded the shortest learning time of 9 min because the number of parameters used in the model was the smallest at 1.24 million. The classification accuracy is 0.8681, which is not high. However, the NasNetLarge model recorded the longest training time of 888 min. The classification accuracy was also 0.9351, showing an above-average accuracy (for reference, the average accuracy was 0.9203). The best classification model is the Inception-ResNet-v2 model with a classification accuracy of 0.9681. It is a relatively heavy model with 164 layers and required 198 min to train. The model with the next highest accuracy was the ResNet-18 model, which had 18 layers and required a short training time of 19 min, but recorded a high classification accuracy of 0.9645. The DarkNet-19 and Xception models showed the next highest classification accuracies.
Classification accuracy of 19 models versus training time.
Figure 3 shows the confusion matrix for the binary classification of MEC and ESC data sets. True positive (TP) refers to the number of times that the classification model correctly classified an ESC image as an ESC image. True negative (TN) refers to the number of times that the classification model correctly classified MEC images as MEC images. False positive (FP) refers to the number of times that the classification model misjudged the MEC image as an ESC image. A false negative (FN) refers to the number of times that the classification model incorrectly assessed an ESC image as an MEC image. These TP, FP, FN, and TN values are calculated through learning and testing in each model.
Confusion matrix for binary classification of MEC and ESC data images.
A detailed performance comparison of the classification models is verified using performance metrics such as the accuracy, precision, sensitivity, specificity, and F1 score. The mathematical formulae for these performance measures are as follows.16)
| \begin{equation} \text{Accuracy} = \frac{\mathit{TP} + \mathit{TN}}{\mathit{TP} + \mathit{TN} + \mathit{FP} + \mathit{FN}} \end{equation} | (1) |
| \begin{equation} \text{Precision} = \frac{\mathit{TP}}{\mathit{TP} + \mathit{FP}} \end{equation} | (2) |
| \begin{equation} \text{Sensitivity} = \frac{\mathit{TP}}{\mathit{TP} + \mathit{FN}} \end{equation} | (3) |
| \begin{equation} \text{Specificity} = \frac{\mathit{TP}}{\mathit{TN} + \mathit{FP}} \end{equation} | (4) |
| \begin{equation} \text{F1Score} = \frac{2\mathit{TP}}{2\mathit{TP} + \mathit{FN} + \mathit{FP}} \end{equation} | (5) |
Performance metrics in eqs. (1) to (5) for the four models with the highest classification accuracy among the 19 classification models in Fig. 2: Inception-ResNet-v2 model, ResNet-18 model, DarkNet-19 model, and Xception model. The classification accuracy was compared in detail. Figure 4 compares the performance metrics of the four selection models with respect to training data, test data, and total data. The closer the performance metric is to 1.0, the better the classification performance. It is common to see relatively lower performance metrics using test data that were not used in model development compared with performance metrics using training data that are employed to create the model. However, as shown in Fig. 4(b), the performance metric results obtained using the test data show a high level of prediction, almost corresponding with the performance metric results obtained using the training data in Fig. 4(a).
Performance metrics for the four selected models with different data sets: (a) training data, (b) test data, and (c) total data.
When comparing each model, the Inception-ResNet-v2 model showed the best performance overall. The ResNet-18 model also showed a similar performance but showed somewhat insufficient results in the test results. The DarkNet-19 model showed a lower classification performance than the ResNet-18 model but showed a higher performance metric for the test data. In the case of the Xception model, the precision and specificity results showed relatively low values compared to the results of other models.
Among the 19 classification models compared in this study, it was confirmed that the Inception-ResNet-v2 model architecture was the most effective as a model for classifying MEC and ESC data. However, when considering the learning time, the ResNet-18 model structure, which shows almost similar performance metrics, is also recommended for new model development; this is despite the learning time of 1/10th level. According to a recent report by Jeon et al.,17) as the deep learning neural network structure becomes more complex, the model accuracy increases, but overfit is reached at some point. Therefore, it is proposed that optimizing the hyperparameter is a more effective method even for a model with a simple structure.
This study proposes a CNN-based classification model that can classify macroscopic fracture surfaces of plastics depending on whether or not ESC is applied. A model was built by applying the neural network structure of 19 previously reported CNN models, and 32,096 data set images were directly trained. Four models with a high classification accuracy were selected and their performance metrics were compared from the confusion matrix. The model with the Inception-ResNet-v2 structure showed the best performance metric of over 0.96 in all items. The model with the ResNet-18 structure showed a slightly lower performance metric level than the model with the Inception-ResNet-v2 structure, but the learning time was 1/10th that of the model with the Inception-ResNet-v2 structure; learning was achieved very quickly. This is because the ResNet-18 structure has a much simpler linear structure with 18 layers and is expected to be more advantageous for structure modification in the future.
This research was supported by Hyundai Mobis.
