Research on Multi-Source Cooperative Transfer Learning-Based Health Diagnosis for Rotating Machinery Bearings

Abstract

Rotating machinery bearings are the core load-bearing components of major industrial equipment such as nuclear power circulating water pumps and wind turbines. Their operating conditions directly determine the safety and stability level as well as operation and maintenance economy of units, and high-precision health diagnosis is a core research demand in the field of industrial intelligent operation and maintenance. Most existing diagnostic methods rely on single vibration signals and have limitations including insufficient recognition of early friction deterioration, weak model generalization under variable working conditions, and outputs dominated by discrete fault classification without continuous health degree support, which cannot adapt to the full-life-cycle operation and maintenance decision-making demands of complex industrial scenarios. To solve the above problems, this paper proposes an Acoustic-Vibration-Temperature Attention Transfer diagnosis method (AVT-AT). The method uses a fused convolution module with parallel standard convolutions and dilated convolutions to extract multi-scale deep features, and adaptively fuses vibration, sound pressure and temperature signals via a two-layer attention mechanism with intra-modal screening and inter-modal weighting. Fisher Score and Maximum Mean Discrepancy are combined to screen sensitive fused features, domain alignment loss is introduced to construct a transfer learning framework, and an equipment health index is built based on feature class center distance mapping to realize continuous quantification of health degree and accurate fault localization. Experimental results demonstrate that the fault diagnosis accuracy of the proposed method reaches 96.7%. Its diagnostic accuracy under variable working conditions is improved by more than 6% compared with traditional transfer learning methods, and it can automatically identify rolling element bearing damages with vibration over-limit and abnormal noise. The proposed method breaks the information blind zone of single-sensor diagnosis, provides feasible technical support for multi-source collaborative intelligent operation and maintenance of rotating machinery bearings, and has favorable engineering application value.