PSI - Issue 84

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

1124

additionally generated and augmented with noise levels DS-0: 3 . 65%, DS-1: 4 . 5%, DS-2: 4 . 5%, DS-3: 3 . 5%, and DS 4: 2 . 0%. A fixed random seed was used to ensure reproducibility.

0

y

x

5

5

-

5

5%

- 0%

-

-

0.55

0

0

5.75

0

-

0.85

5

7.

5

5

0.9

-

5.75

-

0.55

-

5

- A

5

5%

0%

-

w x

-

0

0

5

0.8

0.85

5

Fig. 5 . Simulated damage scenarios (DS-0 to DS-4), ranging from the healthy state to different levels of stiffness degradation and prestress loss.

6. MLP-based Damage Classification This section presents the MLP-based damage classification model and its network architecture, which serves as one of the base learners in the proposed ensemble framework. The proposed MLP consists of an input layer with 29 features, four hidden layers, and a five-class SoftMax output layer. The network is organized into four fully connected blocks that progressively extract compact feature representations. The first three blocks comprise dense layers with ReLU activation and L2 regularization ( λ = 2 × 10 − 4 ), followed by batch normalization and dropout for improved regularization. Specifically, these blocks contain 200, 100, and 50 units with dropout rates of 20%, 15%, and 10%, respectively. The final block produces a 25-dimensional latent embedding using a dense ReLU layer with L2 regularization and layer normalization. All input features are standardized using z-score normalization prior to training. The model is trained for 150 epochs with a batch size of 32 using the Adam optimizer with a learning rate of 1 × 10 − 4 . Early stopping with a patience of 20 epochs and a ReduceLROnPlateau learning-rate scheduler (reduction factor 0.5, patience 15) are employed to enhance convergence and mitigate overfitting. The average total training time is 28.26 seconds, corresponding to approximately 0.24 seconds per epoch. Figure 6 (a-b) illustrates the training history. Both training and validation losses decrease rapidly during the initial epochs and stabilize after approximately 100 epochs, while the corresponding accuracies converge within the range of 0.90 to 0.95, indicating stable training, e ff ective convergence, and good generalization without noticeable overfitting.

,A y9.6%

Fig. 6. Training and validation curves showing (a) loss convergence and (b) accuracy stabilization, indicating effective learning and generalization; (c) MLP-based confusion matrix summarizing classification performance across all structural states.