Application of Convolutional Neural Network and Data Preprocessing by Mutual Dimensionless and Similar Gram Matrix in Fault Diagnosis

Abstract

Bearing fault diagnosis is of great significance to the reliability and stability of modern petrochemical systems. The existing dimensionless index-based bearing fault diagnosis methods suffer from several shortcomings, which are associated with excessive dependence on expert knowledge, insufficient sensitivity in fault feature extraction, and low diagnostic accuracy of nonstationary nonlinear dynamic signals. In this article, a data preprocessing by mutual dimensionless and similar Gram matrix in fault diagnosis is proposed. In this preprocessing, the vibration signal of bearing fault is treated by the mutual dimensionless theory and similar Gram matrix, which is further integrated with the convolutional neural network. The proposed method is tested on two datasets, including a multistage centrifugal fan dataset from our laboratory and a motor bearing dataset from the Case Western Reserve University, achieving an average prediction accuracy of 89.65% and 97.21%, respectively, while reducing the training time significantly. Moreover, the proposed method is compared with other deep learning and traditional methods, including recurrent neural network, support vector machines, and multigenetic algorithm. The experimental results demonstrated that this method can effectively identify fault types by combining various intelligent methods, and compared to traditional dimensionless fault diagnosis methods, the average diagnosis accuracy is significantly improved. The results are also compared with results reported in the literature, indicating that the proposed method can improve fault diagnosis accuracy in an effective and stable way.