Scholarlink

Condition monitoring of induction motors is critical for ensuring the reliability and operational efficiency of industrial systems. This paper presents an end-to-end deep learning framework designed to diagnose stator inter-turn short-circuit faults, a prevalent and potentially catastrophic failure mode. The proposed methodology utilizes a one-dimensional convolutional neural network (1D CNN) that directly processes raw three-phase current signals to classify seven distinct machine health states, spanning both incipient and severe fault conditions. Through a systematically validated data augmentation strategy incorporating noise injection and amplitude jittering, the model effectively mitigates overfitting and demonstrates enhanced generalization capabilities. Upon evaluation with a public benchmark dataset, the model achieves a classification accuracy of 99.32% on an independent test set (95% CI [98.9%, 99.7%]). This outcome validates the efficacy of a well-regularized 1D CNN in delivering a highly accurate and automated fault diagnosis solution, eliminating the need for manual feature engineering. Furthermore, a rapid inference time (~4.3 ms/sample) and a compact quantized model size of 4.05 MB position the framework as a scalable and efficient solution for predictive maintenance in Industry 4.0 environments, confirming its suitability for real-time monitoring on edge computing devices.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust 1D CNN-Based Approach with Data Augmentation for Incipient and Severe Stator Fault Classification in Induction Motors

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hanen Berriri,
  • Mohamed Faouzi Mimouni

摘要

Condition monitoring of induction motors is critical for ensuring the reliability and operational efficiency of industrial systems. This paper presents an end-to-end deep learning framework designed to diagnose stator inter-turn short-circuit faults, a prevalent and potentially catastrophic failure mode. The proposed methodology utilizes a one-dimensional convolutional neural network (1D CNN) that directly processes raw three-phase current signals to classify seven distinct machine health states, spanning both incipient and severe fault conditions. Through a systematically validated data augmentation strategy incorporating noise injection and amplitude jittering, the model effectively mitigates overfitting and demonstrates enhanced generalization capabilities. Upon evaluation with a public benchmark dataset, the model achieves a classification accuracy of 99.32% on an independent test set (95% CI [98.9%, 99.7%]). This outcome validates the efficacy of a well-regularized 1D CNN in delivering a highly accurate and automated fault diagnosis solution, eliminating the need for manual feature engineering. Furthermore, a rapid inference time (~4.3 ms/sample) and a compact quantized model size of 4.05 MB position the framework as a scalable and efficient solution for predictive maintenance in Industry 4.0 environments, confirming its suitability for real-time monitoring on edge computing devices.