Bearing fault diagnosis based on Gramian angular field and DenseNet

Abstract

Rolling bearings are the core components of mechanical and electrical systems. A practical fault diagnosis scheme is the key to ensure operational safety. There are excessive characteristic parameters with remarkable randomness and severe signal coupling in the rolling bearing operation, which makes the fault diagnosis to be challenging. To deal with this problem, the Gramian angular field (GAF) and DenseNet are combined to perform feature extraction and fault diagnosis. The GAF can convert 1-dimensional time series into an image, which can guarantee the completeness of feature information without temporal dependence. The GAF images are then trained by using the DenseNet to generate a data set network. In this process, the transfer learning (TL), which can solve the problem of insufficient samples, is integrated to the DenseNet to enhance its extensibility. The comparative simulations are carried out to illustrate the effectiveness of the proposed method.