A Deep Learning Based Fault Diagnosis Method Combining Domain Knowledge and Transfer Learning

Abstract

Deep learning (DL) based fault diagnosis methods have gained considerable attention in the field of machine health monitoring due to their powerful feature learning capabilities. However, embedding domain diagnosis knowledge into the DL framework to obtain enhanced features having better correlation with the exact health conditions of machine elements for improved fault predictions is still an open challenge. In this paper, a fault diagnosis method combining two-dimensional (2D) image representations of squared envelop spectrum (SES) of vibration signals of bearings, a critical machine element, and a pretrained convolutional neural network (CNN) is proposed. SES is one of the most efficient indicator for the assessment of second order cyclostationary symptoms of damages, which are typically observed in bearings. In this paper, we integrate this knowledge in designing a DL framework for efficient fault diagnosis in bearings. The proposed method is tested and evaluated on an experimental bearing vibration dataset collected under different operating and fault conditions. Experimental results demonstrate that the proposed method can achieve a high diagnosis accuracy and present a better generalization ability both in balanced and imbalanced data scenarios