PSI - Issue 37

Tomasz Rogala et al. / Procedia Structural Integrity 37 (2022) 187–194 / Structural Integrity Procedia 00 (2019) 000 – 000

192

6

determine features for two-dimensional objects. The features used include both simple, e.g. surface area, as well as advanced e.g. Hu statistical moments shown in Shivanand (2020).

Fig. 5 Subsequent operation of transformation of a real damage object represents crack. a) original object) b) equivalent ellipsoid, c) boolean product operation d) surface mesh e) initial image after the transformation f) final image after image processing

4. Classification 4.1. Methodology

The general concept of damage classification was proposed to be based on a deep learning technique (Li et al., 2021). A convolutional neural network (CNN) was selected in this study to automatically recognize the type of damage based on selected features of objects observed on tomographic scans. 2D and/or 3D features were normalized and transformed in order to create the 2D input image of the network. The neural model was set to return the label representing the damage type. The major properties and parameters needed to create the structure of a convolutional neural network were as follows. The network was composed of 15 layers. The first one was an image input layer of size 9x9x1 with the zero-centre normalization technique. Next three computational blocks in the network included sublayers such as: convolution layer with strid e [1 1] and padding ‘same’; batch normalization layer, ReLU layer; max pooling layer with stride [2 2] and padding [0 0 0 0]. Fully connected and softmax layers were added in the decision part of the network. The last layer was a classification output layer with a weighted cross entropy loss function to prevent overfitting due to the imbalanced dataset problem. Because of 2D/3D feature normalization and transformation pre-processing procedures proposed in our approach, it was not reasonable to use the transfer learning method. Moreover, the proposed representation of the input pattern caused that data augmentation technique was not possible to apply, as well. The whole dataset (11 242 patterns representing crack damage and 2 036 patterns corresponding to delamination damage) was divided into three subsets in the following proportions: 80% training patterns, 10% validation patterns and 10% test patterns. The additional patterns (3 362 crack damage patterns and 672 delamination damage patterns) were also collected to verify and highlight the final performance of the model. An update rule based on the stochastic gradient descent with momentum was selected to train the neural model (Murphy, 2012). The final result of the training process strongly depends on the values of the behavioural parameters of the training algorithm. Therefore, several variants were examined. The best results were obtained when the mentum coefficient was equal to 0.85, L2 regularization parameter was set to 0.0001. The values of other parameters of the algorithm were set as follows: initial learn rate 3E−4, learn rate drop period = 5, verbose frequency = 8, validation frequency = 10, learn rate schedule was set to ’piecewise’ and shuffling option was set to ’every - epoch’. Deep learning