Intelligent Bearing Fault Diagnosis Using Multi-Head Attention-Based CNN

Abstract

Aiming at automatic feature extraction and fault recognition of rolling bearings, a new data-driven intelligent fault diagnosis approach using multi-head attention and convolutional neural network (CNN) is proposed. Firstly, a simple signal-to-image spatial transform method is utilized to generate 2D gray images within different health states from raw bearing vibration signals, and generated images are divided randomly into training data and testing data. Secondly, we adopt the multi-head attention mechanism to optimize the CNN structure and develop a new convolutional network model for intelligent bearing fault diagnosis. Next, the training data is used to train network parameters of the designed CNN model to accurately realize bearing fault recognition. Finally, the diagnostic performance of the proposed bearing diagnosis framework is verified on different sub-datasets under working loads of 0~3 hp, and comparison experiments are also conducted with other CNN models and bearing fault diagnosis methods. Experiment results show that the proposed diagnostic method can achieve effective bearing fault diagnosis with relatively fewer CNN model parameters, and has some generalization ability. The recognition rate of bearing states under working loads of 0~3 hp all reach over 99%.