A fault diagnosis method using improved prototypical network and weighting similarity-Manhattan distance with insufficient noisy data

Currently, few samples and the inevitable noise poses a severe test on deep learning methods. To solve the above problems, a novel fault diagnosis network based on a refined prototype and correlation weighting Manhattan distance (RPCMN) is proposed. Specifically, a multiscale feature extraction (MSFE) module and a sparse non-local attention (SNLA) module are developed to comprehensively extract key classification information. Moreover, a prototype interactive refinement mechanism (PIRM) is established to refine the position of prototypes to make them more representative. A correlation weighting Manhattan distance (CWMD) is designed to accentuate the correlation between different prototypes. The superiority of our method is verified on two standard datasets and one vibration dataset in practical industrial applications. We found that the diagnosis accuracy is 99.12 % at the training set size of 20. Meanwhile, at different noise levels (−6 to 6 dB), the diagnostic accuracy is higher than 90 %.