A Method for Bearing Fault Diagnosis Based on a CNN-Transformer Deep Neural Network

Abstract

To address the limitations of traditional methods for compound fault diagnosis of rolling bearings under variable-speed conditions-namely, their reliance on manual feature extraction and the difficulty of single deep learning models in simultaneously capturing local impact details and global compound coupling relationships-this paper proposes an end-to-end rolling bearing fault diagnosis method based on a CNN-Transformer architecture. Taking three-channel vibration signals as input and requiring no manual feature design, the method first performs instance normalization on each sample independently to eliminate amplitude shifts caused by rotational speed; subsequently, a multi-scale one-dimensional convolutional neural network extracts local impact features, whilst a Transformer encoder utilizes self-attention mechanisms to model the cross-position compound relationships of fault features globally; Finally, key information is adaptively aggregated via learnable query-attention pooling, followed by fault classification through a fully connected layer and a softmax function. Validation was conducted on a dataset comprising eight fault categories and five rotational speeds, collected from a compound fault injection test bench. The results demonstrate that, in a noisy environment with a signal-to-noise ratio of 4 dB, the proposed method achieves a diagnostic accuracy of 86.9%; in tests involving variable-speed operating conditions not covered by the training set, the accuracy significantly outperformed the comparison models, demonstrating excellent generalization stability. This method provides a high-precision, highly robust solution for the diagnosis of compound rolling bearing faults in industrial settings characterized by variable speeds and high-noise environments.