Multisensor Bearing Fault Diagnosis Based on the Improved DIVERSIFY Under Multiple Unknown Working Conditions

Existing solutions for bearing fault diagnosis often exhibit a considerable decline in performance when faced with multiple unknown working conditions, a critical out-of-distribution (OOD) generalization challenge. To address this, this paper introduces an improved DIVERSIFY framework with two key enhancements. Firstly, to overcome the critical issue of baseline DIVERSIFY requiring manual tuning for the number of latent domains, we propose a novel identification algorithm underpinned by a unique Kullback-Leibler (KL)-Log metric. This metric’s innovation lies in its ability to transform the continuous KL-divergence value into a discrete, integer-based format, which enables a direct and rule-based determination of the number of domains from the data distribution itself. Secondly, to adapt the DIVERSIFY structure for complex multi-sensor scenarios, a multi-sensor convolution fusion block (MCFB) is designed with a parallel multi-kernel architecture. By simultaneously employing convolutional kernels of varying sizes, the MCFB explicitly captures and fuses both intricate short-term local features and essential long-range temporal dependencies from multi-sensor signals. Comprehensive validation on a series of OOD generalization tasks demonstrates the method’s superiority. The proposed framework achieves average diagnostic accuracies of 97.41%, 95.20%, and 88.63% in challenging scenarios with one, two, and three unknown working conditions, respectively, significantly outperforming the baseline and other state-of-the-art approaches.