A fault diagnosis method based on dynamic convolution and polarized self-attention feature fusion networks

Fault diagnosis techniques, which are crucial in the field of industrial intelligent manufacturing, are capable of equipment performance maintenance and productivity improvement. In fault diagnosis, multi-type sensors are commonly used for monitoring because a single data source fails to provide sufficient information to support the comprehensive analysis and accurate diagnosis. Hidden information between modes can be mined using data fusion techniques, enabling more effective decision-making and condition analysis. However, the data measured by multiple sensors are subject to issues such as varying types, an imbalanced ratio of positive to negative samples, and significant differences in data structure, making multi-source data fusion and inter-feature information acquisition challenging. To address these problems, we propose a fault diagnosis method based on dynamic convolution and polarized self-attention (DC-PSA) feature fusion networks. Given that unimodal features are not utilized comprehensively enough, we propose a dynamic convolution-based feature self-convergence model. The ability of the model is improved by attentively aggregating multiple convolution kernels, which are combined in a form dynamically adjusted according to different inputs to fully utilize the features. To enable effective feature-level integration across modalities, we establish a cross-attention-based multimodal fusion model, where each modal branch learns multiscale spatial information independently and forms cross-channel interactions in a localized manner, which can realize the information interactions between local and global channel attention. Empirical results on the Paderborn benchmark dataset validate that the proposed method captures the complementary characteristics across signal types more effectively than existing methods, leading to a notable boost in diagnostic accuracy following the fusion process. The accuracy of the proposed model reached 98.6%, representing an improvement of 8.74% compared to the baseline model.