PSI - Issue 84

Prajwal Giri et al. / Procedia Structural Integrity 84 (2026) 1119–1126

1125

Figure 6c presents the confusion matrix of the MLP classifier for the five damage classes (DS-0 to DS-4), achieving a mean classification accuracy of 92.6%. The model demonstrates strong performance across all classes, with particularly high accuracies for DS-0 (99%), DS-2 (92%), and DS-3 (99%). DS-1 and DS-4 also exhibit reliable performance, with classification accuracies of 89% and 84%, respectively. The primary misclassifications occur between DS-1 and the healthy state (DS-0), and between DS-4 and DS-0. These confusions are attributed to the similarity in structural responses of these cases, as DS-1 involves minor sti ff ness degradation near the damage threshold, while DS-4 reflects prestress loss that predominantly a ff ects static behaviour and induces no variation in modal features. Overall, the MLP e ff ectively distinguishes most damage scenarios, with only minor overlap among structurally similar conditions. 7. Comparison between Di ff erent Base Learning Models The classification performance of the four base learners—MLP, KNN, SVM, and RF—together with their ensemble combinations across the five damage classes (DS0–DS4) is summarized in Figure 7a. The ensemble integrates individual predictions using both hard and soft voting strategies; in soft voting, class probabilities are weighted by normalized model accuracies (KNN: 0.229, SVM: 0.245, RF: 0.264, MLP: 0.263) before averaging, whereas hard voting assigns the final class by majority decision.

0 99% 89% 9 % 99% 8 %

0 5% 76% 87% 96% 99%

0 99% 88% 87% 97% 6 %

0 97%

00% 00% 00%

C

68%

9 .6%

80.6%

86. %

9 %

0 99% 97% 99% 98% 87%

0 00% 9 % 97% 98% 7 % 9 %

96%

Fig. 7. Classification and ROC comparison of individual models and the ensemble across damage classes (DS-0–DS-4). The ensemble significantly outperforms the individual classifiers, with the soft voting ensemble achieving the highest mean accuracy of 96.0%, followed by the hard voting ensemble at 92.0%. Among the base learners, RF performs best (93.0%), closely followed by MLP (92.6%), SVM (86.4%), and KNN (80.6%). Class-wise results further confirm the robustness of the ensemble: the soft voting approach maintains consistently high accuracy across all damage states, achieving 96% for DS0 and 99% for both DS2 and DS3, while e ff ectively compensating for large variability among individual models, particularly for KNN in DS0 (45%) and RF in DS4 (68%). Even for the most challenging class, DS4, the ensemble attains a balanced accuracy of 87%. Although hard voting performs well for dominant classes (DS0: 100%, DS3: 98%), it shows reduced performance for DS1 (92%) and DS4 (73%) compared to soft voting. The ROC curves in Figure 7b further confirms these trends. The ROC analysis was performed using a micro averaged multi-class evaluation across all damage scenarios (DS0–DS4), where each damage class was treated in a one-versus rest manner to form a single global ROC curve. The decision threshold is automatically set to each unique predicted score value produced by the classifiers. For the soft voting ensemble, these scores represent the weighted class probabilities of the base learners, while for hard voting they correspond to normalized vote fractions from majority voting. At each threshold level, samples with scores greater than or equal to the threshold are classified as positive and the corresponding True Positive Rate (TPR) and False Positive Rate (FPR) are computed.