Static Eccentricity Fault Location Diagnosis in Resolvers Using Siamese-Based Few-Shot Learning
Mahdi Emadaleslami, MohammadSadegh KhajueeZadeh, Farid Tootoonchian
Abstract
Static eccentricity in the resolver, with higher probability and severity occurrence, gradually injects instability into the drive system; thus, its exact location diagnosis is mandatory. Whereas most existing investigations rely on frequency analysis of signal voltages for resolver fault diagnosis, artificial intelligence (AI) is at the center of attention for electric motors fault diagnosis. Yet, sufficient data collection brings some challenges for AI. Therefore, intrigued by the idea of few-shot learning to overcome the data scarcity challenge, a Convolutional Neural Siamese network is employed with limited data to diagnose the static eccentricity location under each stator tooth via raw signal voltages, which is ordinarily employed for rotor angle extraction. Accordingly, a Siamese network learns to yield the probability of signal pairs being similar by distance unit. A lab-built Time-Stepping Finite Element Analysis (TS-FEA) and a resolver test-bench dataset under different training sizes, noise existence, and new unseen locations challenges are employed to assess the employed Siamese network. Comparing the employed Siamese network with One Dimension-Convolutional Neural Network (1D-CNN) shows its proficiency and superiority.