Out of Bounds Anomaly Scores (OBAS) in Smart Failure Detection in Variable Frequency Drives

Smart condition-based maintenance is a critical aspect of Industry 4.0, ensuring the reliable and long-term operation of devices powered by Variable Frequency Drives (VFDs). Detecting early signs of failures in VFDs involves analyzing input data from sensors, including temperature signals, among others. However, due to the complexity of VFD operation, there persists a need for unified approaches to effectively detect anomalies. In this paper, we propose a novel anomaly detection pipeline that integrates a diverse set of models and introduces an innovative anomaly metric called the "Out Of Bounds Anomaly Score" (OBAS). The OBAS metric simplifies model output interpretation and lays the groundwork for an easy-to-use ensemble learning framework. Our approach leveraged temperature signals obtained from both real-world and simulated VFD operation scenarios, specifically wind turbine and virtual drive, respectively. We explored four key methodologies: Hidden Markov Models, Deep Autoencoders, and Principal Component Analysis. These methodologies were employed to detect periodic temperature increases, temperature shifts, and temperature profile changes after retrofit. We compared these models against classical anomaly detection tools such as exponential smoothing, moving average, and Hampel Filter. Our study showed a comprehensive pipeline for anomaly detection, seamlessly integrating different machine/deep learning models with OBAS, which facilitated comparisons of their performance. Furthermore, OBAS fostered construction of ensembles of models of different architectures and outputs. This helped to stabilize anomaly predictions yielding 2-10% improvement in accuracy. This research contributes to advancing anomaly detection in VFDs, bridging the gap between existing techniques and the evolving demands of Industry 4.0.