CN-TimeXer: a healthy-state data-only, residual-driven framework for unsupervised fault diagnosis of rotating machinery under variable operating conditions

Due to the fact that the operating conditions of the equipment change over time and the scarcity of labeled fault samples, the diagnosis of faults in rotating machinery faces significant challenges. To address variable operating conditions, we introduce Cycle-Normalized Relative Positional Encoding (CNRPE). By integrating CNRPE, rotational speed, load, and time as exogenous variables with endogenous vibration signals via cross attention mechanisms, and optimizing with a time-frequency joint loss, the model accurately captures complex data distributions. To address the lack of labeled fault samples, we propose the CN-TimeXer framework, trained exclusively on normal operating data. Under unknown conditions, the model predicts vibration data; the difference between these predictions and actual measured signals is recorded as residuals. Multi-view features are then extracted from these residuals for unsupervised clustering, and cluster labels are matched one-to-one with actual sample labels to complete the diagnosis. Extensive experiments on three public datasets (HUST, MCC5-THU, and CWRU) demonstrate that CN-TimeXer achieves over 97% accuracy, outperforming state-of-the-art methods by about 6% on average. This demonstrates that the proposed method maintains excellent diagnostic performance even under conditions of scarce fault samples and variable operating conditions.