Self-Supervised Drift-Resilient Classification for Time Series Industrial Anomaly Detection

In modern industrial environments, early detection of anomalies is essential to prevent unplanned downtime and maintain operational efficiency. Traditional rule-based and supervised methods often struggle with data drift, limited labeled data, and the inability to capture the complex interdependencies inherent in interconnected industrial systems. To address these challenges, we propose a drift-resistant self-supervised anomaly detection framework specifically designed for industrial applications. Unlike conventional fault classification approaches that identify predefined defect types, our model focuses on detecting subtle, evolving anomalies in time-series vibration data. The framework integrates advanced statistical and spectral feature extraction with a dynamic, rolling-window-based data grouping strategy, enabling the model to adapt robustly to temporal variations. Evaluation is based solely on accuracy and the experimental results demonstrate that our approach achieves up to 75% faster anomaly alerts compared to conventional ISO 10816 standards. Validation in NASA bearing datasets, as well as real-world vibration data from industrial fans, motors, and gearboxes confirms the model scalability and practical applicability. This work provides a cost-effective and reliable solution for continuous condition monitoring, thereby laying a strong foundation for enhanced predictive maintenance in complex industrial settings.