A Frequency-Aware Transformer for Multiscale Fault Diagnosis in Electrical Machines

Abstract

Motor fault diagnosis is a critical technology for ensuring the reliability of industrial equipment and enabling predictive maintenance. However, conventional fault diagnosis methods struggle to effectively capture complex time-frequency patterns, limiting their ability to perform early fault detection and accurate classification. To address these challenges, this study proposes the Frequency-Aware Motor Fault Transformer (FAMFT), which integrates a self-attention mechanism with multi-scale feature analysis to comprehensively analyze the time-frequency characteristics of multidimensional power quality data, including voltage, current, and harmonics. FAMFT overcomes the limitations of conventional CNN- and RNN-based models through three key innovations: 1) Multi-scale feature extraction via parallel analysis of fine-, intermediate-, and long-term temporal scales, 2) Selective feature enhancement through a frequency gating mechanism, and 3) An interpretable fault diagnosis framework based on SHAP (SHapley Additive Explanations). Experimental results demonstrate that the proposed model achieves 99.9% diagnostic accuracy by maintaining an exceptionally low false alarm rate and missed detection rate, thereby ensuring high reliability. Notably, FAMFT exhibits consistent performance across various load conditions, demonstrating a level of robustness suitable for direct implementation of real-time predictive maintenance systems in industrial environments. This study introduces a novel transformer-based fault diagnosis paradigm, contributing to the stable operation of power systems and improving maintenance efficiency.