ScholarWorks@동국대학교

초록

Reliable fault detection in electric motors is essential in ensuring operational safety, minimizing downtime, and lowering maintenance costs in modern industrial systems. Long-range temporal relationships and intricate nonlinear patterns found in motor state monitoring data are frequently difficult for conventional signal processing methods and conventional deep learning models to capture. The present manuscript suggests a transformer-based fault detection model with a self-attention mechanism for intelligent electric motor fault diagnosis to address these issues. The proposed methodology enables efficient feature learning without relying on handcrafted features or recurrent neural networks. This is done by directly modeling temporal correlations within multi-sensor motor signals using the Transformer architecture. Under different operating settings, the self-attention mechanism improves sensitivity to subtle fault characteristics by adaptively highlighting signal segments that are relevant to the defect. To learn discriminative fault representations, raw or minimally preprocessed vibration and/or stator current inputs are segmented and embedded before being processed by stacked attention layers. Benchmark motor fault datasets under various operating situations are used to validate the efficacy of the suggested approach. The Transformer-based model consistently outperforms traditional convolutional and recurrent neural network approaches in terms of accuracy, precision, recall, and F1-score, according to experimental results. Furthermore, the model has enhanced generalization performance and great resistance to noise, suggesting its appropriateness for practical industrial applications. These findings demonstrate that attention-based transformer designs provide a potent and scalable solution for sophisticated electric motor status monitoring and problem detection. © 2026 IEEE.

키워드