Prior distribution-guided disentanglement autoencoder for abnormal detection under variable operating conditions

Abstract

This paper proposes an anomaly detection method based on disentanglement learning to address the challenge of operating condition (OC) interference in the health monitoring of a machine. The method first utilizes ternary mutual information theory to analyze and deconstruct the coupling relationships among OCs, health state, and monitoring signals. A prior-guided disentanglement loss function is derived based on the variational estimation of mutual information. A disentanglement autoencoder with a dual-encoder structure is designed to learn a feature representation space that is invariant to OCs. By calculating the Wasserstein distance between the current state and the baseline normal state within this space, a final health indicator is constructed. Furthermore, kernel density estimation with a 99.7% confidence level is employed to determine the warning threshold adaptively. Experiment results on two bearing accelerated degradation datasets demonstrate that the proposed approach suppresses the interference from OC variation in health state assessment. Compared to the existing methods, the proposed prior distribution-guided disentanglement autoencoder achieves more timely and precise abnormal detection. The proposed method effectively implements artificial intelligence techniques to enhance the intelligence and reliability of anomaly detection in mechanical systems.