Generative adversarial networks for overlapped and imbalanced problems in impact damage classification

Abstract

This study develops an oversampling method designed for rebalancing datasets in impact damage classification of reinforced concrete walls. The proposed method addresses class imbalance and overlap issues using the Wasserstein Generative Adversarial Network with Gradient Penalty (WGANGP) approach, aiming to enhance classification accuracy. Initially, sample weights are calculated for overlapped and non-overlapped samples based on their proximity to the decision boundary via an initial classification process using the Support Vector Machine algorithm. Then, these weighted samples are combined with the original dataset to form a training dataset for the WGANGP model. To effectively learn from limited data samples and stabilize the training process, the Model Agnostic Meta-learning (MAML) method is incorporated into the WGANGP algorithm. New samples are generated to balance the impact damage dataset using the generator network of the WGANGP model. The proposed method is compared to advanced oversampling methods handling overlapping, noisy, near-borderline samples between the classes. Four well-known classifiers are trained on the impact damage dataset, including Multilayer Perceptron (MLP), Random Forest (RF), Extreme Gradient Boosting (XGB), and Support Vector Machine (SVM). Implementation results demonstrate the efficacy of the proposed oversampling technique on the impact damage dataset, outperforming others in terms of F1 score.