Abstract
Imbalanced datasets in Industrial Internet of Things (IIoT) environments pose a serious challenge for reliable pattern classification. Critical instances of minority classes (such as anomalies or system faults) are often vastly outnumbered by routine data, making them difficult to detect. Traditional resampling and machine learning methods struggle with such skewed data, usually failing to identify these rare but significant events. To address this, we introduce a two-stage generative oversampling framework called Enhanced Optimization of Wasserstein Generative Adversarial Network (EO-WGAN). This enhanced WGAN-based Oversampling approach combines the strengths of the Synthetic Minority Oversampling Technique (SMOTE) and Wasserstein Generative Adversarial Networks (WGAN). First, SMOTE interpolates new minority-class examples to roughly balance the dataset. Next, a WGAN is trained on this augmented data to refine and generate high-fidelity minority samples that preserve the complex non-linear feature distributions characteristic of IIoT data. Unlike prior SMOTE and GAN methods, our framework leverages the Wasserstein loss for more stable training. It incorporates an optimized sampling strategy to ensure that the synthetic data meaningfully extends the classifier's decision boundaries. Integrating an advanced oversampling technique with a critic-guided generative model significantly improves minority-class recognition, eliminating the need for extensive feature engineering or domain-specific tuning. We validate EO-WGAN on an IIoT cybersecurity dataset (UNSW-NB15) and several other imbalanced benchmarks. The proposed method consistently outperforms state-of-the-art oversampling techniques, achieving up to 95.2% accuracy (with precision and recall of 94.6% and 95.4%, respectively) in our experiments. EO-WGAN offers a scalable and cost-effective solution for anomaly detection and predictive maintenance in Industrial Internet of Things (IIoT), and its generality makes it applicable to other domains that face severe class imbalance. The results demonstrate that our approach significantly enhances the detection of minority-class events, resulting in more reliable industrial analytics and informed operational decision-making.