Bearing Fault Diagnosis Based on Dual-Branch Multi-Scale CNN-LSTM Feature Fusion

• Published in: 2024 Second International Conference on Cyber-Energy Systems and Intelligent Energy (ICCSIE) pp. 1 - 6
• Main Authors: Zhang, Tong, Liu, Qingke, Li, Kai, Wang, Nan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.05.2024
• Subjects: Accuracy; Bearing fault diagnosis; Convolutional neural network; Convolutional neural networks; Dual-branch; Fault diagnosis; Feature extraction; Long short term memory; Multi-scale; Representation learning; Robustness; Training; Training data; Vibrations
• DOI: 10.1109/ICCSIE61360.2024.10698409

The condition detection and fault diagnosis of bearings are the basis for the smooth operation of rotating machinery. Since the traditional convolutional neural network cannot make full use of the time correlation of the original signal and it is difficult to obtain enough training data under multiple working conditions, this paper proposes a CNN-LSTM fault diagnosis method based on dual-branch multi-scale feature fusion (DMCNN-LSTM). Without any preprocessing and traditional feature extraction, the original vibration signal is directly used as the input of the model, and a two-branch convolutional neural network with different kernel sizes is established. The system can autonomously learn features from the unprocessed data stream. The long short-term memory is then used to identify fault types based on the learned features. The method is verified in the Case Western Reserve University (CWRU) dataset, and the proposed method can achieve an average accuracy of 99.867% and show robustness in parameter changes. Compared with deep neural network (DNN), long short-term memory (LSTM), and Convolutional Neural Network (CNN) algorithms, DMCNN-LSTM has better performance under common fault conditions, which provides a scalable and efficient solution for industrial applications.