Bearing fault diagnosis method based on multi-scale domain adaptative network across operating conditions

Abstract The intelligent rolling bearings fault diagnosis methods adopting a single vibration signal as the model input present low diagnostic precision, poor noise robustness, and difficulty in applying to variable operating conditions, so a multi-scale domain adaptation network (MSDAN) was put forward for variable load fault diagnosis of rolling bearings. This method combined multi-scale feature extraction with a lightweight convolutional neural network to extract complementary fault features from coarse-grained vibration signals at multiple time scales. Then, correlation alignment (CORAL) distance and domain identification adversarial learning were applied to extract domain invariant features to establish an end-to-end unsupervised fault diagnosis system for rolling bearings. The MSDAN model was evaluated using variable load-bearing datasets of two experimental setups and compared with other methods. The results show that MSDAN has better diagnostic accuracy and cross-domain adaptability than other domain adaptation fault diagnosis methods. In addition, our multi-scale method has more robust stability and generalization ability than any single-channel feature extraction method.