A Dual-View Information Complementary Fault Diagnosis Method for Combined Bearings in Gearboxes of Nuclear Power Circulation Pumps

Abstract To address limitations of single-source signals and single-view models for high-precision diagnosis of gearbox combined bearings in nuclear power circulating pumps, this study proposes a diagnostic method integrating the complementary representations of image- encoded frequency-domain information from multi-source signals and a dual-view model. First, the frequency-domain information of multi-source signals was encoded into feature matrices, which were then mapped to the RGB channels of true-color images, realizing the conversion from one-dimensional signals to color images. Based on this process, this paper proposes a multi-source signal image encoding method that integrates the frequency-domain fusion Symmetric position matrix (FSPM) and frequency-domain fusion Relative position matrix (FRPM). Through dimensional concatenation and fusion of feature matrices from multi-source signals, the two generated types of image datasets exhibit information complementarity. Moreover, the proposed method takes image size as the sole variable, thereby avoiding the impact of multi-parameter signal processing on diagnostic results. Subsequently, based on the two types of generated image datasets, this study designs a Dual-view multi-loop cascaded residual neural network (DMCRNN). The multi-loop cross-scale cascaded residual structure enhances the model’s capability to extract complex fault features, while the dual-view parallel input mechanism ensures the synchronous extraction and fusion of complementary image information. The proposed method achieved an average diagnostic accuracy of 99.68% on a self-built combined bearing test bench. Ablation experiments and noise robustness analysis proved the effectiveness of complementary image information and the robustness of the proposed method, while comparisons with other methods further highlighted its superiority. Finally, the method's generalization capability was further validated on a self-collected gear dataset.