Gear Fault Diagnosis in Geared Motors Based on Frequency Adaptation Graph Prototype Network with Limited Data

Detecting and classifying gear faults with limited labeled data is essential for effective fault diagnosis in electromechanical systems. This paper presents a novel Frequency Adaptation Graph Prototype Network (FAGPN) for classifying various gear faults in geared motors. FAGPN employs customized low-pass and high-pass filters, integrated with an attention mechanism, to enhance multi-scale feature fusion from vibration signals processed through various frequency-time methods. Additionally, FAGPN projects embeddings onto hyperspherical space with a consistency constraint to improve accuracy. Experimental results on real-world datasets involving multiple gear sets in different states demonstrate the superior performance of FAGPN. Remarkably, with a training ratio of 1%, FAGPN achieves over 99% accuracy.