An improved information fusion-based transfer learning approach for bearing fault diagnosis

Due to the presence of negative transfer (NT) phenomena, transfer learning (TL) algorithms often encounter challenges in achieving optimal results for bearing fault diagnosis (FD) applications. To address NTP in TL and fully leverage multi-source state data, this study proposes an enhanced information fusion-based TL approach for bearing FD. First, we provide a brief introduction to the fundamental theory and algorithm developed in this paper. Secondly, we conduct a detailed analysis of the FD algorithm based on information fusion enhancement. Finally, we perform experimental simulations using publicly available fault datasets and compare and analyze the FD algorithms of CNN, traditional TL, and the proposed method. It demonstrates that the highest accuracies attained through CNN, traditional TL, and our proposed FD algorithms are $79.64 \%, \mathbf{9 0. 0 0 \%}$, and $\mathbf{1 0 0. 0 0 \%}$, respectively. These findings indicate that they can effectively integrate multi-source fault information, mitigate NT occurrences, as well as enhance model accuracy, stability, and generalization ability for FD.