Transfer learning-based channel attention enhancement network combined with Gramian angular domain field for fault diagnosis

Abstract

Convolutional neural networks are increasingly used in the field of fault diagnosis, in order to give full play to the performance of the network within a certain number of model layers. While ensuring a high diagnostic accuracy, with strong generalization performance. We proposed a method that is simple, but effective. In this paper, we design a network structure for channel attention enhancement based on transfer learning (TL). The low-level is combined with TL to extract generic features of the target domain, and the high-level use a more refined channel attention module to extract and filter the abstract features of the current task object. The structure can fully exploit the fault information without increasing the network depth. Combined with Gramian angular difference field (GADF) to encode the vibration signal into 2D images as the input of the training model for fault diagnosis of rolling bearings. Source and target domains in TL uniformly use GADF encoded maps, effectively reducing the need for labeled samples. In order to validate the effectiveness of the method proposed in this paper, experiments were conducted using two publicly available bearing fault datasets and one laboratory-collected data, respectively. The results show that the proposed method is suitable for fault diagnosis of bearings in complex operating conditions and is highly generalizable.