Semi-supervised diagnosis method of refrigeration compressor hidden defect based on convolutional transformer autoencoder model

Abstract

Improving product quality and service life by diagnosing defects in compressors through vibration signal analysis of the compressor's shell is challenging due to strong noise interference and limited labeled samples. The paper proposed a semi-supervised diagnosis method based on the convolution Transformer autoencoder (CTAE) for diagnosing hidden defects with small samples and strong noise. Firstly, an optimized variational modal decomposition is utilized to decouple and reduce noise from the raw vibrations, and the decomposition results are concatenated with the vibration signals as input to the model; Secondly, the CTAE is employed to learn the feature distribution of unlabeled samples and to extract and fuse local and global features from the input data; Finally, a labeled samples are used to fine-tune the model and to fuse features from multi-sensor information. The results of using a compressor dataset for validation show that the proposed method has high diagnosis accuracy and robustness with limited labeled samples and different signal-to-noise ratios.