A Method of Rolling Bearing Fault Diagnose Based on Double Sparse Dictionary and Deep Belief Network

Feature extraction is the key technology in the data-driven intelligent fault diagnosis methods of rolling bearing. However, the acquired features by the traditional methods, which mainly based on time-frequency domain, sometimes cannot well represent the characteristics of the signal and are difficult to accurately identify because of their complexity and subjectivity. Aiming at this problem, the sparse representation theory is used to the field of fault diagnosis because the different types of rolling bearing fault signals only has the highest matching degree with the dictionary atoms trained by the same type of fault signals. A novel method based on the double sparse dictionary model joint with Deep Belief Network (DBN) is proposed for the fault diagnosis of rolling bearing. Firstly, each type of fault signal is trained according to the double sparse dictionary learning algorithm and the corresponding double sparse subdictionaries is obtained. In order to reduce the feature dimension of sparse representation coefficients, the low contribution atoms of all subdictionaries are removed and the rest are recombined into a comprehensive double sparse dictionary. Then, the Orthogonal Matching Pursuit (OMP) algorithm is adopted to obtain the corresponding sparse feature coefficients of each fault signal on the comprehensive double sparse dictionary. Finally, the coefficient is used as the input of DBN to train and judge the faults of the rolling bearing. The experimental results show that the proposed method has higher diagnosis accuracy and stability compared with the traditional intelligent fault diagnosis methods, and the training and testing time of DBN is greatly reduced.