Full-field Measurements for Anomaly Detection of Mechanical Systems using Convolutional Neural Networks and LSTM Networks 109 • Generating copies of the videos: Each original video was augmented with seven additional copies using data augmentation techniques, increasing the dataset to a total of 560videos. This helped improve the neural network’s robustness against variations and prevent overfitting. • Incorporating and unifying classes: New classes corresponding to different bumpers were added. After detailed analysis, it was decided to unify some classes to improve representation in the dataset. The classes assigned to the bumpers were: Bumper 1, Bumpers 1 and 2, Bumper 3, Bumper 4, Bumper 5 and Linear (no bumpers). In addition, the hyperparameters selected for the neural network are presented in Table 1. Table 1 Hyperparameters of the proposed approach. Analysis and Results The results presented in this section are divided into different cases, starting with Case 1, where the proposed method will do a simple classification of linear and nonlinear behavior. Case 2 will later present the results for the multiclass classification, with the linear response, and 5 different setups with bumpers as shown in Figure 3. Case 1. Training with Linear and Non-Linear classes. As a first step, a binary simulation was performed with only two classes: Linear and Non-Linear (which groups all bumpers). The accuracy was 97.14%. Case 2. Training with Bumper Classes Trainings were carried out using the classes corresponding to the different bumpers and with all the frequencies. The accuracy was 94.28% The results indicate that: • Effectiveness of Preprocessing: The reduction and balancing of the dataset, along with video preprocessing, contributed to improving the efficiency and precision of the simulations. Removing irrelevant segments allowed the network to focus on the most significant information. (a) (b) Fig. 5 Confusion matrix of linear and nonlinear cases (a) and accuracy of each class (b).
RkJQdWJsaXNoZXIy MTMzNzEzMQ==