A deep-transfer-learning fault diagnosis method for gearboxes based on discriminative feature extraction and improved domain adversarial neural networks

Gearbox vibration data collected under different operating conditions often suffer from inconsistent feature distribution and the challenge of removing noise components, which complicates fault diagnosis in real-world engineering applications. This paper proposes a deep transfer learning-based fault diagnosis method that addresses these challenges by leveraging discriminative feature extraction and improved domain adversarial neural networks. First, labelled and unlabelled vibration signals are organised into datasets using a fixed-length data segmentation method. To mitigate the interference of noise signals in real-world operating conditions, we employ a convolutional block attention module and a discriminant loss function-assisted feature extractor to extract distinguishing features. Next, to tackle the problem of inconsistent feature distribution, the multiple kernel maximum mean discrepancy is used to align the global distributions of the source and target domains, while an adversarial method is applied to align the sub-domain distributions between the two domains. Finally, experiments are conducted on an open-variable condition gearbox fault dataset. The results demonstrate that the proposed method achieves an average recognition accuracy of over 96.12%, confirming its effectiveness and superiority over other diagnostic methods. This algorithm offers significant value in practical engineering, providing a robust approach for fault diagnosis in industrial gearboxes with varying operating conditions and noise, thus improving maintenance efficiency and reducing downtime.