Abstract
This paper presents a comprehensive diagnostic framework for electrical machines, based on the application of artificial neural networks (ANNs) for the analysis of electrical and vibration signals. The proposed method leverages deep learning architectures to automatically extract informative features and achieve high fault classification accuracy. The framework integrates signal pre-processing, neural network training, and a condition evaluation module, enabling the implementation of a predictive maintenance system suitable for industrial applications. A multi-sensor diagnostic system is proposed, combining CNN-LSTM architectures with a graph neural network (GNN) for correlational analysis of currents, vibrations, and thermal parameters. This approach allows early detection of inter-turn short circuits and bearing faults, improving diagnostic accuracy by 7–12% compared to existing state-of-the-art methods. The framework demonstrates robustness under varying operating conditions, including transient and self-excitation regimes, and provides physically interpretable results, bridging the gap between data-driven and physics-informed diagnostics.