微特电机 ›› 2026, Vol. 54 ›› Issue (9): 73-77,83.

• 生产技术 • 上一篇    下一篇

边缘计算微特电机振动实时监测与智能诊断方法

李可成,甘  景,周雪会   

  1. 广西工业职业技术学院,南宁 530001
  • 出版日期:2026-09-28 发布日期:2026-09-28
  • 作者简介:李可成(1980—),男,本科,副教授,研究方向为电气自动控制技术。 甘景(1995—),男,硕士,教师,研究方向为智能控制技术。 周雪会(1981—),女,硕士,副教授,研究方向为工业互联网技术。
  • 基金资助:
    2023年度广西职业教育教学改革研究项目 (GXGZJG2023B037) ; 2025 年广西高校中青年教师科研基础能力提升项目(2025KY1553) ;自治区科技厅2025 年“ 八桂之光” 访学研修计划(科技人才专项)项目(桂科发〔 2025〕121号)

Vibration Monitoring and Intelligent Diagnosis for Micro Special Motors Based on Edge Computing

LI Kecheng,GAN Jing,ZHOU Xuehui   

  1. Guangxi Vocational & Technical Institute of Industry,Nanning 530001,China
  • Online:2026-09-28 Published:2026-09-28

摘要:

针对微特电机振动信号幅值小、频带宽,以及现有深度学习模型在边缘侧推理延迟高、噪声与小样本下鲁棒性不足的问题,提出一种边缘计算架构下的实时监测与智能诊断方法。 构建端-边-云三层物理架构,设计信号预处理流水线与时域、频域轻量化特征,采用幅度阈值与马氏距离两级筛选降低存储负载。研发低比特量化全连接神经网络,融合预测熵与综合预警分数实现分级预警。实验表明,本方法在-5 dB(A) 噪声下准确率为 83. 5%,1%小样本下准确率为81. 2%,边缘推理时间均值为 0. 089 ms、标准差为 0. 007 ms,显著优于多尺度卷积稀疏注意力变换器模型( multi-scale convolutional sparse attention Transformer,MCSAT) ,满足微特电机采样周期内实时诊断需求,为资源受限边缘端高可靠监测提供有效方案。

关键词: 微特电机, 振动监测, 边缘计算, 轻量化神经网络, 故障诊断

Abstract:

To address the challenges of small amplitude, wide bandwidth, high inference latency of existing deep learning models on edge sides,and insufficient robustness under noisy and few-shot conditions for vibration signals of micro special motor, this paper proposes a real-time monitoring and intelligent diagnosis method based on an edge computing architecture. A three-layer end-edge-cloud physical architecture is constructed, along with a signal preprocessing pipeline and lightweight time-domain and frequency-domain features. A two-stage screening mechanism combining amplitude threshold and Mahalanobis distance is adopted to reduce edge storage load. Furthermore,a low-bit quantized fully connected neural network is developed, integrating prediction entropy and a comprehensive warning score to achieve hierarchical alerting. Experimental results demonstrate that the proposed method achieves an accuracy of 83. 5% under - 5 dB noise and 81. 2% in the 1% few-shot scenario, with an average edge inference time of 0. 089 ms and a standard deviation of 0. 007 ms,significantly outperforming the multi-scale convolutional sparse attention transformer( MCSAT) method. It meets the real-time diagnostic requirement within the sampling period of micro special motor and provides an effective solution for highly reliable monitoring on resource-constrained edge sides.

Key words: micro special motor, vibration monitoring, edge computing, lightweight neural network, fault diagnosis

中图分类号: