A Width Multi-Scale Adversarial Domain Adaptation Residual Network with AConvolutional Block Attention Module

Abstract

Although the fault diagnosis methods based on deep learning have attracted widespread attention in the academic field in recent years, such methods still face many challenges, including complex and variable working conditions, insufficient ability to extract key features, and large differences in sample data. To address these problems, a width multi-scale adversarial domain adaptation residual network with a convolutional block attention module (WMSRCIDANN) is proposed in this paper, which consists of a feature extraction network, a domain discriminant network, and a label classification network. In the feature extraction network, an improved width multi-scale residual network combined with a convolutional block attention module (WMSRC) is used as the feature extractor to achieve a weighted fusion of multi-depth features.In the domain discriminative network, the fully-connected network is replaced by a four-layer convolutional structure, which can further reduce the difference in feature distribution and improve the cross-domain invariance of deep features. In the label classification network, the classifier uses the extracted domain-invariant features to perform cross-domain fault identification. The experimental results show that WMSRCIDANN is effective in cross-domain bearing fault diagnosis.

An Unsupervised Rotating Machinery Fault Diagnosis Method Based on Multi-Scale Feature Residual Neural Network

“…Since fault signals exhibit different characteristics at different time scales and frequencies, using fixed-scale convolution operations alone cannot fully capture these features. Convolution layers with different scales have different frequency resolutions, which is why using multiple scales of convolution layers can help comprehensively learn fault features and improve fault detection accuracy [22]. By extracting more comprehensive feature information, the proposed method in this paper can more accurately characterize fault features.…”

Innovative integration of multi-scale residual networks and MK-MMD for enhanced feature representation in fault diagnosis

“…The multi-scale residual neural network (MSRNet) adopts a multi-branch structure, where each branch uses filters of different scales to capture features at different scales [36]. Residual connections are employed within each branch to avoid the issue of gradient vanishing, and the outputs of all branches are combined through weighted average pooling.…”