Accurate induction motor eccentricity detection from phase currents is attractive for industrial monitoring because it can be implemented non-invasively using existing electrical measurements. However, developing robust neural classifiers is often limited by the scarcity of labeled fault data across operating conditions. This work proposes a compact convolutional neural network (CNN) for eccentricity-level classification, trained on a broad set of simulated current signals enhanced to better reflect measurement imperfections, and designed as a foundation for future transfer to real-machine recordings.

Three-phase stator currents were generated with an eccentricity simulation model and enriched using a residual-injection scheme motivated by frequency-domain inspection and correlation analysis, which indicate notable non-ideal components affecting signal consistency. The dataset spans five eccentricity levels (0.0–0.4, step 0.1), steady loads from 0 to 10 (step 2), and 1 Hz sinusoidal load profiles within 0–4, 4–6, and 6–10, for steady speeds of 1500, 1350, and 1200 rpm. From each case, 50 windows of 900 samples were extracted and reshaped into a 30×30×3 representation (three channels for phase currents). The CNN includes three convolutional feature-extraction blocks (convolution, batch normalization, ReLU, max pooling) and a classifier head with adaptive average pooling, dropout, and a fully connected layer.

The proposed network achieved approximately 96% test accuracy, with comparable validation accuracy and ~97% training accuracy across the considered conditions.

A lightweight CNN can accurately classify eccentricity levels using current-only inputs when trained on condition-diverse, residual-augmented simulation data. In future work, the trained model will serve as a pretraining baseline for transfer learning to laboratory measurements, enabling practical eccentricity detection on real motors.