Few-shot fault diagnosis of machinery from steady to fluctuating conditions

Abstract Deep learning methods have strict demands on the quantity and quality of data. However, in engineering practice, it is highly difficult to obtain a sufficient number of labeled fault samples. Furthermore, due to the dynamic operating states and external environment factors, the machinery often operates in fluctuating working conditions. As a result, the collected vibration signals are not only limited in quantity but also exhibit considerable distributional shifts, which brings great challenges to fault diagnosis. Therefore, an improved semi-supervised prototype network (ISSPN) is proposed to solve the few-shot fault diagnosis problem of machinery from steady to fluctuating conditions. Multi-head attention-guided residual feature encoder is employed to enhance the feature representation, and adaptive prototype adjustment strategy is designed to improve the alignment with the actual data distribution. The method was validated on bearing and gearbox fault diagnosis cases by simulating various industrial scenarios. The experimental results demonstrate that the proposed ISSPN presents higher accuracy in the few-shot fault diagnosis from steady to fluctuating conditions.