Bi-CovaSC: A Few-Shot Learning Model with Multi-Scale Feature and Covariance Bi-Metric Learning for Bearing Fault Diagnosis

The growth of modern industry comes with the inevitable application of electrical machines and their indispensable ball-bearing parts, which significantly improve the machinery's efficiency, stability and lifespan. To ensure the quality of the machines, one has to take account of the condition of the ball bearings, particularly to identify any possible faults, which is challenging. With that problem in hand, a deep-learning few-shot model named “Bi-CovaSC” is proposed, which could efficiently diagnose abnormalities in the ball bearings without much expertise knowledge and data provided. In this model, by utilizing the “MAFE” mechanism, the model could extract the feature patterns of different ranges and then learn their long-range relationship. To further highlight the extracted information both spatial-wise and channel-wise, the BAM attention block is employed before the information is passed through the Covariance-Bi-metric classification module. The “Bi-CovaSC” model is evaluated on the Paderborn University (PU) dataset and yielded high accuracy, which puts the model in a potential place for real-life application. The code will be available at: https://github.com/viet2715/fewshot-code