A Time Series Diagnosis Method for Industrial Equipment Faults Combining Complex-Valued Spectral Attention and Deformable Convolution

To address the problem of resonance and noise coupling in sensor signals under complex working conditions, this study proposes a time series diagnostic method that combines complex-valued spectral attention with deformable convolution. First, the method captures multi-scale interactions between amplitude and phase in the complex frequency domain. Then, deformable convolution is used to adaptively extract spatial features of abnormal patterns. The method is validated on a rolling mill bearing dataset from a steel plant. Results show that the early fault detection recall reaches 95.2%, which is significantly better than the current mainstream Spectral-LSTM method.