ConditionEnsembleNet: an interpretable deep learning model for detecting anomalies in rotating machinery under various operating conditions

Abstract Deep learning models have demonstrated effectiveness in measurement-based rotating machinery anomaly detection, but challenges in interpretability and generalization across diverse operating conditions remain. To address these, a novel deep learning framework named ConditionEnsembleNet is proposed for robust anomaly detection under various operating conditions with interpretable insights. More specifically, the transparency of ConditionEnsembleNet is boosted by integrating wavelet kernels as the signal-processing module and Gaussian mixture models as the prior in probability and statistics within the network architecture. Additionally, we enhance the generalization capacity of ConditionEnsembleNet under different operating conditions by assigning distinct weights to density estimation tasks for each specific operating condition. Experimental results on real-world measurement datasets demonstrate that ConditionEnsembleNet achieves superior anomaly detection accuracy across diverse measurement scenarios compared to state-of-the-art methods, while providing human-understandable explanations of feature relevance and decision-making. This combination of generalization and transparency makes it a robust solution for industrial measurement systems requiring high-precision anomaly detection and diagnostic reliability in rotating machinery.