A Real-Time Framework for Low-Load Fault Detection and Multi-Level Fault Severity Assessment Using an Energy-Modulated Transformer Autoencoder

Abstract

Early fault detection and fault severity assessment are essential for ensuring the reliable operation of induction motors. While stator winding faults can often be identified through current-based monitoring techniques, reliable characterization of fault severity remains challenging, particularly for incipient faults under low-load operating conditions. To address this problem, this study proposes a low-latency fault diagnosis framework that combines healthy-only anomaly detection with multi-level fault severity assessment. At its core, an Energy-Modulated Transformer Autoencoder (EM-TransAE) is developed to model healthy operating behavior by incorporating instantaneous current energy into the self-attention mechanism. The framework adopts a two-stage learning strategy. First, EM-TransAE is trained exclusively on healthy data within a one-class learning paradigm, where fault detection is performed through reconstruction-error-based anomaly assessment. The proposed model achieves a reconstruction error of 0.000484, indicating high-fidelity representation of healthy operating conditions. Second, encoder-derived latent features and reconstruction error are combined to form a diagnostic feature set for XGBoost-based multi-level fault severity assessment, achieving an accuracy of 96.93%. To support practical deployment, the framework is integrated within an edge-based streaming architecture for real-time operation. Experimental results on stator inter-turn faults across multiple low-load operating conditions demonstrate superior reconstruction performance compared with representative recurrent, transformer-based, and graph-based autoencoder models while maintaining computational efficiency suitable for low-latency deployment.