Self-supervised contrastive learning with scale-sensitive receptive fields for machine fault diagnosis under sharp speed variation

Ensuring safety of machine operation is of great importance, particularly during periods of sharp speed variation. However, in condition monitoring data under variable speeds, the fault impulse signals are not periodic and exhibit considerable non-stationarity, owing to the fixed sampling frequency. To address this issue, we propose a self-supervised contrastive feature learning method with time scale-sensitive feature extraction ability. Our method starts with constructing a feature extractor with multiple scale receptive fields, which enables feature extraction and weighted fusion from time scale-varying fault impulse signals using attention mechanism. Next, we design a self-supervised training strategy to train the feature extractor, thereby reducing the distribution differences of fault features at different speed levels by similarity comparison based on Euclidean distance. Finally, a softmax classifer is used for fault identification. Our method is applied to two fault diagnosis cases and proves to be more robust and effective compared to state-of-the-art methods. Under sharp speed variations, our method achieved an accuracy of 96.72% in identifying bearing faults.