Global contextual multiscale fusion networks for machine health state identification under noisy and imbalanced conditions

CNN-based intelligent fault diagnosis methodologies have demonstrated excellent performance in machine health condition monitoring and safety assessment. However, the majority of existing CNN models are developed on the basis of undisturbed and balanced distribution of samples, which is inconsistent with real industrial scenarios. To tackle this issue, a global contextual multiscale fusion network (GCMFN) is developed in this study. The main contributions of this study are highlighted and summarized as follows: (1) a multi-dilated fusion layer and a non-local activation module are developed as the building units of GCMFN to guide the model for exploring multiscale features; (2) a global contextual denoising module is applied to amplify important features and eliminate interference features, and (3) an online label smoothing algorithm is utilized to promote the better diagnostic performance of GCMFN under imbalanced scenarios. Three experiments using the benchmark motor dataset, the planetary gearbox dataset, and the industrial pump dataset are implemented to test the applicability of the proposed GCMFN in machine health state identification. The experimental results show that GCMFN is competent and a promising diagnostic tool for various machine reliability monitoring tasks.