Low-Parameter Critic-Based Multivariate WGAN Model for Clogging Detection in Drives

Abstract Efficient detection of anomalies in the cooling system of variable frequency drives (VFDs) is crucial to minimise downtime costs from overheating. Smart condition monitoring tools, especially those using machine/deep learning, have proven effective for failure detection. Recent research has focussed on environmental factors, such as pollution and humidity, affecting VFDs. Clogging is particularly harmful as it can damage power electronics, leading to extended downtimes. This study explores the use of Wasserstein Generative Adversarial Networks (WGANs) for detecting clogging in drives, including inlet/outlet/heatsink clogging and fan blockage. WGANs are adept at recognising complex temporal patterns due to their feedback-driven training. Despite generative AI models being typically large and unsuitable for embedded systems, this work demonstrates the feasibility of a low-parameter WGAN critic-based model for detecting cooling issues in VFDs. Using temperature signals, the model can detect clogging as low as 20%–30% with high performance metrics, achieving up to 90% accuracy and an F1 score above 0.9 for heatsink clogging detection, using a lightweight 26-parameter critic model. This study shows the potential for developing low-parameter WGAN critic-based models for clogging detection in VFDs.