Two-stage dynamic time warping for intelligent fault detection of rotating machinery under variable speed

Abstract Effective fault detection in rotating machinery operating at variable speeds often necessitates numerous fault samples for deep learning model training. In practice, acquiring sufficient fault samples is frequently challenging due to cost limitations. To address this issue, this paper introduces a two-stage dynamic time warping (DTW) method for intelligent fault detection in rotating machinery at variable speeds. In the first stage, the original fault signal is filtered based on the collected speed data to create a vibration signal for the shaft under variable speeds. A reference signal for fixed speed is then estimated, allowing the DTW algorithm to determine optimal paths for regularization. This process transforms the non-stationary signal, mitigating the effects of speed variations, and normalizes it to produce a consistent signal. The normalized signal undergoes spectral analysis to generate a normalized order spectrum. In the second stage, distances to each standard sample in the order domain are calculated using the DTW algorithm, with the shortest distance indicating the corresponding fault type. Finally, experimental data from bearings and gears is utilized, comparing this method with others to demonstrate its effectiveness and advantages.