A residual denoising and multiscale attention-based weighted domain adaptation network for tunnel boring machine main bearing fault diagnosis

Abstract

As a critical component of a tunnel boring machine (TBM), the precise condition monitoring and fault analysis of the main bearing is essential to guarantee the safety and efficiency of the TBM cutter drive. Currently, under conditions of strong noise and complex working environments, traditional signal decomposition and machine learning methods struggle to extract weak fault features and achieve high fault classification accuracy. To address these issues, we propose a novel residual denoising and multiscale attention-based weighted domain adaptation network (RDMA-WDAN) for TBM main bearing fault diagnosis. Our approach skillfully designs a deep feature extractor incorporating residual denoising and multiscale attention modules, achieving better domain adaptation despite significant domain interference. The residual denoising component utilizes a convolutional block to extract noise features, removing them via residual connections. Meanwhile, the multiscale attention module uses a 4-branch convolution and 3 pooling strategy-based channel-spatial attention mechanism to extract multiscale features, concentrating on deep fault features. During training, a weighting mechanism is introduced to prioritize domain samples with clear fault features. This optimizes the deep feature extractor to obtain common features, enhancing domain adaptation. A low-speed and heavy-loaded bearing testbed was built, and fault data sets were established to validate the proposed method. Comparative experiments show that in noise domain adaptation tasks, proposed the RDMA–WDAN significantly improves target domain classification accuracy by 42.544%, 23.088%, 43.133%, 16.344%, 5.022%, and 9.233% over dense connection network (DenseNet), squeeze-excitation residual network (SE-ResNet), antinoise multiscale convolutional neural network (ANMSCNN), multiscale attention module-based convolutional neural network (MSAMCNN), domain adaptation network, and hybrid weighted domain adaptation (HWDA). In combined noise and working condition domain adaptation tasks, the RDMA–WDAN improves the accuracy by 45.672%, 23.188%, 43.266%, 16.077%, 5.716%, and 9.678% compared with baseline models.