Anomaly detection of machinery under time-varying operating conditions based on state-space and neural network modeling

Abstract

Anomaly detection is critical for maintaining the health and stability of machinery. However, machines such as wind turbines often operate under time-varying conditions (TVCs), where changes in operating conditions (OCs) introduce disturbances to sensor signals, presenting significant challenges for traditional anomaly detection methods. To address this issue, this paper proposes a novel anomaly detection approach based on state-space and neural network modeling. First, from the perspective of system dynamic response, the machine operating under TVCs is treated as a dynamic response system, with OCs and health states governing the system’s dynamic response. A state-space model is then employed to explicitly model the health state, OCs, and response signals during the normal operation of machinery. Additionally, the nonlinear fitting capability of neural networks is used to parameterize the relationships between these factors. By incorporating OCs and health states into the model, the time-varying response induced by the two factors is effectively modeled as a time-invariant process. Furthermore, an alternating parameter update strategy, utilizing the extended Kalman filter, is developed to estimate both the health state and neural network parameters. Finally, a detection indicator is constructed based on the real-time neural network parameters to achieve machinery anomaly detection. The effectiveness and superiority of the proposed method are validated through simulation experiments and accelerated fatigue degradation experiments on rolling bearings under different time-varying operating conditions.