Multilevel Few-Shot Model With Selective Aggregation Feature for Bearing Fault Diagnosis Under Limited Data Condition

Abstract

Diagnosing bearing faults is an important issue in the field of electrical machines, where approximately 40 $\%$ of faults in electrical machines are caused by bearings. With the development of deep learning, diagnosing bearing faults from vibration signals helps reduce costs and time while increasing diagnostic accuracy. However, traditional deep learning models need to be trained from large and diverse datasets to be able to provide good diagnostic results, which is not suitable for specific data such as bearings because it can be difficult to collect data and require expensive resources. In this letter, a new diagnostic method is proposed based on few-shot learning to overcome the data problem. The proposed method synthesizes information from both spatial-level and channel-level to find information in the condition of only little training data, improving diagnostic accuracy. Besides, selective aggregation feature extraction is proposed to replace the traditional convolution neural network to extract condensed features that carry more information. For instance, with only 30 training samples, the model achieves 86.67% accuracy on the CWRU dataset, this suggested method obtains State-of-the-Art results, demonstrating its efficacy.