Rolling Bearing Fault Diagnosis Based on Multi-Domain Feature Fusion and CBAM-Enhanced Residual Dense Network

Abstract

Based on rolling bearing vibration signals, this paper proposes an intelligent fault diagnosis model that integrates multi-domain features with attention mechanisms. First, by developing a multi-domain feature fusion strategy, time-domain signals, envelope spectra, and Discrete Cosine S-Transform (DCST) features are combined to form multi-channel inputs, effectively enhancing the information-carrying capacity of the original signals. Second, a deep one-dimensional network architecture combining residual and dense connections is designed to mitigate network degradation while enabling full reuse and efficient transmission of multi-layer features. Building on this foundation, the Convolutional Block Attention Mechanism (CBAM) is introduced to adaptively weight features across both channel and spatial dimensions, thereby enhancing the expression of key failure features and improving the model’s ability to identify weak features under complex operating conditions. Experimental results demonstrate that this method exhibits excellent accuracy and stability in multi-class bearing fault recognition tasks, achieving an average recognition rate of 99.19%, which significantly outperforms traditional convolutional networks and classical deep learning models. The proposed method possesses strong generalization capabilities and engineering application potential, offering an efficient and reliable solution for the fault diagnosis of rotating machinery.