PSI - Issue 64

Nikhil Holsamudrkar et al. / Procedia Structural Integrity 64 (2024) 580–587 Holsamudrkar Nikhil et al./ Structural Integrity Procedia 00 (2019) 000 – 000

583

4

3. CNN Model Development and Training The CNN model by Simonyan et al. (2014) was adopted in the study, popularly known as VGG16, as shown in Fig. 4. The data is split into classic 80:20 ratios as train and validation. The test dataset is kept separately and augmented 3 times by adding white Gaussian noise of 5%, 10%, and 15% to simulate the field setting. The data is randomly shuffled before being fed into the model. The model parameters are given in Table 2.

Fig. 2. Filtered signals and corresponding DWT for the (a) Fabric rupture, (b) Fabric-matrix debonding, (c) Steel yielding, (d) Concrete Cracking, (e) Mortar matrix cracking

This model is a convolutional neural network (CNN) tailored for image classification tasks, featuring a 3x3 filter size for fine feature extraction, 2x2 max pooling for downsampling, and a (2, 2) stride to reduce spatial dimensions. With a batch size of 30 and 15 epochs, it undergoes iterative training while employing a dropout rate of 0.5 to mitigate overfitting. The Rectified Linear Unit (ReLU) activation function introduces non-linearity, while the Adam optimizer dynamically adjusts learning rates. Categorical cross-entropy is the loss function, and accuracy is the evaluation metric. The model is cross-validated and utilizes softmax in the final layer for multi-class classification. Further, it employs specific Adam optimizer parameters (Beta1 = 0.9, Beta2 = 0.99, Epsilon = 1e-7) and sets a learning rate 0.0001 to fine-tune training dynamics for optimal performance. Results for training and validation accuracy are shown in Figure 5. The wobbly nature of validation results is due to the limitation of a small dataset. Although increasing epochs increases validation accuracy, it was observed that corresponding loss tends to increase, indicated by the formation of a valley in the loss curve. Therefore, overfitting the CNN model is avoided by interrupting the model training at 15 epochs.