PSI - Issue 77

Muhammad Jahanzeb Zia et al. / Procedia Structural Integrity 77 (2026) 111–118 Muhammad Jahanzeb Zia et al. / Structural Integrity Procedia 00 (2026) 000–000

117

7

3.2. Support Vector Machine (SVM) model The classification of raw AE signals is inherently challenging; therefore, signal normalization was performed to enhance the training efficiency of the Machine learning (ML) model. SVM classifiers are particularly advantageous when dealing with relatively small datasets. In this study, 67 raw AE signal datasets collected from different locations and propagation directions were used to train the classifier. A Radial Basis Function (RBF) kernel with a separation constant of 10 was employed for training and prediction of test samples. RBF was used to map input data in a wide space to classify data; it is commonly used to handle linear as well as non-linear data. The predictive capabilities of the trained model were subsequently investigated on 14 unknown test samples.

(a)

(b)

Fig. 6. Confusion matrices for data used to train and test on same (a) and on different data sets (b).

In Fig. 6, the confusion matrices for the datasets used for training and testing are shown, where Class 0 corresponds to matrix cracking and Class 1 represents fibre breakage. The SVM achieved an F1-score of 1.00 when evaluated on the training dataset, indicating perfect model fitting as all 51 sets for class 0 and 16 data sets for class 1 were flawlessly classified (Fig. 6a). However, during testing on the unseen dataset 8 inputs for class 0 and 3 data sets for class 1 were classified perfectly, whereas three signals from Class 1 were misclassified, as shown in Fig. 6b. For this case, the F1 score for Class 1 was reduced to 0.67 and 0.84 for class 0, which can be attributed to the simultaneous occurrence of multiple damage mechanisms. The precision levels for matrix cracking and fibre breakage in the test data were 0.73 and 1.00, respectively, reflecting strong consistency between the training and testing performance. 4. Conclusions In this study, a finite element modelling approach was employed to generate acoustic emission signals corresponding to different damage modes in E-glass composites. This method provides a cost-effective alternative to experimental testing for creating datasets to train ML models. Matrix cracking and fibre breakage mechanisms were simulated using the ABAQUS/Explicit solver, which resulted in AE waves during the analysis. The frequency content of these signals was compared with experimental results reported by Han et al. (2020), showing close agreement and thereby validating the numerical model. The Support Vector Machine classifier was then subsequently applied and evaluated on both the training and test datasets, confirming the reliability of the model. For previously unknown AE signals, the precision values for matrix cracking were high, whereas some fibre breakage signals were misclassified as matrix cracking. These findings demonstrate that the SVM model, trained on raw AE signals generated through numerical simulations, can achieve reliable performance in classifying composite damage modes. References

Azad, M.M., Kim, S., Cheon, Y.B., Kim, H.S., 2024. Intelligent structural health monitoring of composite structures using machine learning, deep learning, and transfer learning: a review. Advanced Composite Materials 33, 162–188.