Data-driven anomaly detection techniques have been widely used in recent years for condition monitoring of complex mechanical equipment. For example， wind turbines and other equipment often operate under time-varying conditions due to random changes in wind speed， and the changes in conditions further lead to the drift of the distribution of monitoring data and thus the coupling of health status and condition changes. To this end， a new feature decoupling learning framework is proposed to solve the coupling problem between operating conditions and health states under time-varying operating conditions. Specifically， a feature decoupling conditional variation auto encoder （CVAE） network is constructed based on the variational self-encoder to achieve the decoupling of operating conditions（OCs）and health states. Anomaly indicators （ANI） are constructed by dimensionality reduction of the decoupled health-state-related features. Combined with statistical anomaly thresholds to achieve the anomaly detection of bearings under time-varying operating conditions. Finally， the experiments of accelerated fatigue degradation of bearings set at time-varying speed degradation verify the effectiveness of the proposed method， and further demonstrate the superiority of the constructed health indicators in eliminating the interference of time-varying working conditions.