An intelligent fault diagnostic framework for cross-domain adaptation transferred with kurtosis-weighted multi-scale dynamic gated convolution

Abstract In cross-domain scenarios, the traditional transfer learning with fault diagnosis algorithms exhibits issues in multi-scale feature extraction and fusion. An intelligent diagnostic framework combining a kurtosis-weighted multi-scale dynamic gated convolution module with transfer learning is proposed. To address multi-scale feature extraction of vibration signals along the channel dimension, the parallel convolutional branches with distinct receptive fields are initially built. Gating units then adaptively control the flow of different feature information to downstream dynamic modules. A kurtosis-weighted mechanism is designed. Unlike most attention-based methods that rely on model selection or parameter stacking, the proposed approach embeds kurtosis as a form of physical prior knowledge into the weight assignment process. Specifically, adaptive channel weights are generated by quantitatively characterizing the transient impact intensity of multi-scale features, enabling the model to effectively focus on the critical regions where fault-related features are concentrated. Lastly, a domain adaptation method on multi-kernel maximum mean discrepancy is used to align the feature distribution between the source and target domains. In the experimental stage, the proposed algorithm has achieved average accuracies on rolling bearing datasets for cross-operating-condition transfer learning of 98.58% for Jiangnan University (JNU) and 98.12% for Case Western Reserve University (CWRU), respectively. In the cross-device transfer with JNU as the source domain and CWRU as the target domain, the algorithm achieves an average accuracy of 96.59%, significantly outperforming the comparison approaches. The results substantiate that the algorithm has effectively accomplished the mechanical fault diagnosis and demonstrated strong reliability and superiority.