An unsupervised invariant feature domain adversarial network with attention mechanism for cross-domain fault diagnosis of rolling bearings

In real industrial scenarios, distribution discrepancies in vibration signals across varying working conditions can weaken the generalisation capacity of fault diagnosis models, which negatively impacts diagnostic accuracy. To address this limitation, this paper proposes a multi-scale kernel invariant feature domain adversarial network (Msk-IFDAN). The method consists of three core components: an improved feature extractor, a classifier and a conditional domain discriminator. The feature extractor integrates a multi-scale kernel efficient channel attention (Msk-ECA) mechanism to achieve multi-scale channel interaction feature extraction and reduce feature redundancy. Meanwhile, invariant feature learning (IFL) is applied to learn domain-invariant features, enhancing the adaptability of the model to improve the generalisation capacity of cross-domain diagnosis of this method. The unified domain adaptation module of the design achieves precise alignment of feature distributions by reducing the distribution discrepancies through adversarial training strategies. Comprehensive comparative experiments on the Jiangnan University (JNU) dataset and Northeast Forestry University (NEFU) datasets demonstrate the effectiveness of the proposed method, with diagnostic accuracy reaching over 99%. The method not only enables accurate cross-domain fault diagnosis of rolling bearings but also exhibits good stability, effectively solving the challenges of cross-domain fault diagnosis for rolling bearings in industrial scenarios.