Abstract: A fault diagnosis method based on the combination of Gramian angle sum fields (GASF) image encoding and transfer compression neural network (SqueezeNet) is proposed to address the low diagnostic accuracy caused by insufficient fault samples under actual working conditions, as well as the low efficiency and high storage issues of large network models under limited computing power. Firstly, the GASF encoding method is used to transform the one-dimensional bearing signal into a two-dimensional image with temporal correlation, so that the data features contain rich and complete fault information even under small sample conditions. Next, the parameters before the global pooling layer in the pre-trained SqueezeNet on the ImageNet dataset are transferred to the new SqueezeNet, and Sgdm is used as an optimization algorithm to reduce training time and storage. Then, the transferred SqueezeNet is used to extract and classify features from the GASF images, ultimately forming a GSN fault diagnosis model. In order to verify the feasibility and superiority of the proposed method, datasets of different scales were selected for testing, and transfer learning comparisons were conducted with other advanced network models under the same conditions, as well as a comparative analysis with advanced algorithms. The experimental results show that the proposed method has a high fault recognition accuracy of 99.15%, while exhibiting and has stronger small-sample recognition ability and improved more efficient lightweight ability.

Key words: image encoding, migration learning, compressed neural network, rolling bearing, fault diagnosis