The object of the study is the fault detection process in critical rotating machinery, specifically steam turbines and compressors, operating within a petrochemical production environment. Traditional fault detection methods, though proven and cost-effective, struggle to address modern industrial challenges – such as the increasing complexity of sensor data, class imbalance in failure records, and the need for real-time interpretability. Recent advancements in deep learning offer promising solutions to these limitations. This study proposes an integrated framework that combines Wasserstein Generative Adversarial Network (WGAN) for data balancing and TabNet, an interpretable deep learning model optimized for tabular sensor data. The goal is to enhance the accuracy and interpretability of fault detection under imbalanced, high-dimensional industrial datasets. Using historical data from a petrochemical plant (2015–2024), the WGAN-TabNet model demonstrated superior performance compared to traditional classifiers (Logistic Regression, SVM, XGBoost), achieving an accuracy of 96.01%, precision of 93.25%, recall of 93.14%, F1-score of 93.20%, and AUC score of 93.13%. The interpretability provided by combination of TabNet and SHAP analysis further identified key operational variables influencing failure such as oil temperature and gas flow rate, offering actionable insights for predictive maintenance. The results underscore that integrating deep learning with robust data balancing significantly improves fault detection where traditional methods fall short, supporting practical implementation in modern predictive maintenance systems