Experimental Verification of Vibration-Based Unbalance Detection in a PMSM Drive Using a CNN-Based Classification Framework

Abstract

This paper presents an experimental verification of a convolutional neural network (CNN) method for vibrationbased unbalance diagnosis (also known as imbalance diagnosis) in a PMSM drive test stand. Vibration data were acquired using an industrial accelerometer and converted into grayscale image representations for CNN classification using a band-pass-filter-based signal-to-image pipeline. The study considers three variants of the moment of inertia, each measured at three sampling frequencies (1, 8, and 48 kHz) and under two operating conditions: normal and with added unbalance on a metal disk. The results confirm high classification accuracy for the training data at 900 rpm, but reduced performance at unseen speeds (800 and 1000 rpm), indicating limited generalization across operating points and the need for improved robustness of the method.