Robust fault diagnosis of electric vehicle induction motors via Gramian angular field encoding and metaheuristic-optimized deep transfer learning
TL;DR
An advanced diagnostic pipeline is proposed that transforms one-dimensional time-series current and voltage signals into informative two-dimensional spatial representations using Gramian angular field encoding and Coati optimization algorithm-optimized transfer learning framework provides an accurate, interpretable, and computationally feasible solution for induction motor fault diagnosis in electric vehicle applications.
Abstract
The aim of this study is to develop and evaluate a robust fault diagnosis framework for induction motors used in electric vehicle propulsion systems. Since induction motor faults may compromise operational safety, energy efficiency, and system reliability, an accurate and interpretable diagnostic approach is required for safety-critical electric vehicle applications. This study proposes an advanced diagnostic pipeline that transforms one-dimensional time-series current and voltage signals into informative two-dimensional spatial representations using Gramian angular field encoding. This transformation preserves temporal correlations and enables effective deep feature extraction. Five pretrained convolutional neural network architectures, namely the visual geometry group 19-layer network, the 101-layer residual network, the 169-layer densely connected convolutional network, the extreme inception network, and the efficient network B5 variant, are comparatively evaluated through a transfer learning strategy. To enhance classification robustness, the nature-inspired Coati optimization algorithm is integrated to optimize deep feature selection and weighting. The proposed framework was tested on a multi-fault induction motor dataset including normal operation, overload fault, overvoltage fault, phase-to-ground fault, phase-to-phase fault, and undervoltage fault conditions. The experimental results show that the DenseNet169–Coati optimization algorithm model achieves the best performance, with an accuracy of 97.22%, a precision of 97.62%, and a macro area under the curve of 98.80%. Furthermore, gradient-weighted class activation mapping analysis confirms that the model focuses on physically meaningful signal regions. The proposed Gramian angular field-encoded and Coati optimization algorithm-optimized transfer learning framework provides an accurate, interpretable, and computationally feasible solution for induction motor fault diagnosis in electric vehicle applications.