An Infrared Thermal Image Few-Shot Learning Method Based on CAPNet and Its Application to Induction Motor Fault Diagnosis

Abstract

The induction motor is widely used for providing the running power of rotating machinery. Its fault diagnosis is significant to ensure the operation safety of rotary machinery. Infrared thermal image analysis based on deep learning has attracted the attention of many researchers due to its advantages in non-destructive and space locations. However, obtaining sufficient high-quality thermal image samples in practical applications is relatively difficult. Developing few-shot learning models is significant for extending the engineering application of thermal image analysis. The existing models of few-shot learning often neglect the spatial information extraction of images at the feature extraction module, which limits the performance of models. In this paper, a new prototypical network with coordinate attention (CAPNet) for fault diagnosis of three-phase induction motor is proposed. The coordination attention feature extraction module (CAFEM) is designed by the coordinate attention block and the convolution blocks to obtain the spatial relativities between feature maps. The designed CAFEM can inhibit harsh requirements on the amount of data, which greatly enhances the CAPNet’s ability to mine information. Moreover, the features exacted of fault are identified by the metric classifier. The performance of the CAPNet method is validated by experimental analysis of thermal images of a three-phase induction motor. The superiority of the proposed method is demonstrated in comparison with six other state-of-the-art algorithms in few-shot learning.