A novel Order-domain rotating machinery anomaly detection towards non-stationary operating conditions

Rotating machinery in real industrial systems operates under non-stationary conditions with continuous speed fluctuations, while fault data are extremely scarce due to the high cost and risk of data acquisition. This study proposes a novel Order-domain Graph Residual Autoencoder for anomaly detection under variable-speed conditions with zero fault samples. The method first performs feature extraction based on local time–frequency domain order mapping through the short-time Fourier transform, decoupling fault features from rotational speed. An order-domain graph structure is then constructed by balancing global temporal information aggregation with local speed-invariant characteristics, enabling the model to autonomously learn optimal inter-window feature aggregation patterns. Finally, the graph residual autoencoder with an adaptive gating mechanism learns low-dimensional manifold representations of normal states through dynamic residual connection adjustment and adaptive noise injection, achieving fault detection under both known and unknown speed conditions. Experimental validation is performed on a public dataset and an evolving simulator dataset characterized by fluctuating operating conditions. Superior performance compared to other commonly used anomaly detection methods is achieved without any fault information, along with satisfactory interpretability. The proposed framework provides a promising solution for intelligent health monitoring of rotating machinery in complex variable-speed industrial scenarios.