Fault Diagnosis of Rotating Machinery Bearings Based on Multi-Scale Attention Feature Fusion under Few Shot and Complex Working Conditions

Abstract

Accidents caused by failures in rotating machinery bearings have heightened attention in bearing failure diagnosis. Scholars have explored methods to build fault diagnosis models to tackle the challenge of creating diagnostic models with few bearing failure samples available. However, many diagnostic models may suffer from overfitting issues under insufficient samples, affecting fault diagnosis performance. Additionally, rotating machinery operates under changing, complex conditions and noise interference, further deteriorating the effectiveness of fault diagnosis. This paper proposes a fault diagnosis model based on multi-scale attentional feature fusion (FD-MSAFF) to address the issue above. This model comprises a multi-scale feature extraction module, an attentional feature fusion module, and an MMD-based weighted prototype network. The FD-MSAFF model improves few-shot learning by using its multi-scale feature extraction and depth-wise separable attention modules to blend multi-scale features effectively with contextual information. It also tackles classification issues in few-shot sizes with its MMD-weighted prototype network, which is less susceptible to noise. Simulated tests under complex scenarios, such as changing conditions, noise variations, cross-bearing conditions, and comparisons with other algorithms, have proven the FD-MSAFF model's superior accuracy and generalization in diagnosing bearing faults in rotating machinery.