湖北农业科学 ›› 2026, Vol. 65 ›› Issue (11): 176-181.doi: 10.14088/j.cnki.issn0439-8114.2026.11.030

• 信息工程 • 上一篇    下一篇

基于时频分析和迁移学习的茶叶烘干机风机故障智能诊断技术研究

梁玉玥1, 雷道锐1, 李火生2   

  1. 1.恩施职业技术学院智能制造学院,湖北 恩施 445000;
    2.武汉纺织大学外经贸学院机电工程学部,武汉 430200
  • 收稿日期:2026-06-29 发布日期:2026-11-17
  • 通讯作者: 李火生(1973-),男,湖北大悟人,副教授,高级工程师,硕士,主要从事机器人工程、数字孪生等研究,(电子信箱)381035608@qq.com。
  • 作者简介:梁玉玥(1987-),女(苗族),湖北咸丰人,副教授,博士,主要从事农业机械智能装备研究,(电子信箱)274655882@qq.com
  • 基金资助:
    湖北省自然科学基金联合基金项目(2026AFC0245); 恩施州“启航专项”科技计划项目(D202407); 湖北省职业技术教育学会科学研究课题(2026ZJGB050)

Research on intelligent fault diagnosis technology for tea dryer fan based on time-frequency analysis and transfer learning

LIANG Yu-yue1, LEI Dao-rui1, LI Huo-sheng2   

  1. 1. School of Intelligent Manufacturing, Enshi Polytechnic, Enshi 445000, Hubei, China;
    2. Department of Mechanical and Electronic Engineering,College of International Business and Economics,WTU, Wuhan 430200, China
  • Received:2026-06-29 Online:2026-11-17

摘要: 为了克服茶叶烘干机风机在变工况条件下故障诊断准确率低、模型泛化能力不足的问题,提出一种融合自适应时频分析与深度迁移学习的智能故障诊断方法(VCRJ-Net)。该方法首先采用变分模态分解联合小波包阈值对实测振动信号进行降噪处理,并通过连续小波变换将信号转换为时频图像;然后构建基于残差网络和卷积块注意力模块的特征提取网络,引入联合最大平均偏差度量源域与目标域之间的分布差异,结合梯度反转层实现对抗式域适应,从而在目标域无标签条件下完成跨工况故障识别。结果表明,工况A条件下VCRJ-Net的诊断准确率达96.21%,优于其他对比方法。在3种变工况迁移任务(A→B、A→C、B→C)中,VCRJ-Net的诊断平均准确率达94.50%,较未引入域适应的ResNet提升了12.22个百分点。消融试验中,VMD联合降噪模块和JMMD域适应模块分别贡献了4.13、9.04个百分点的准确率增益,证明信号预处理与域对齐在农用装备智能诊断中均不可或缺。VCRJ-Net有效提升了茶叶烘干机风机故障诊断的泛化能力和识别精度,为茶叶加工装备的智能化运维提供了可行的技术方案。

关键词: 时频分析, 迁移学习, 茶叶烘干机, 风机故障诊断, 智能诊断技术

Abstract: To overcome the problems of low fault diagnosis accuracy and insufficient model generalization ability of tea dryer fans under variable operating conditions, an intelligent fault diagnosis method (VCRJ-Net) integrating adaptive time-frequency analysis and deep transfer learning was proposed. This method first used variational mode decomposition combined with wavelet packet thresholding to denoise the measured vibration signals, and converted the signals into time-frequency images through continuous wavelet transform; then a feature extraction network based on a residual network and a convolutional block attention module was constructed, and a joint maximum mean discrepancy was introduced to measure the distribution difference between the source domain and the target domain, and adversarial domain adaptation was achieved by combining a gradient reversal layer, thereby completing cross-operating-condition fault identification without labels in the target domain. The results showed that under operating condition A, the diagnostic accuracy of VCRJ-Net reached 96.21%, outperforming other comparison methods. In three variable-operating-condition transfer learning tasks (A→B, A→C, and B→C), the diagnostic accuracy of VCRJ-Net reached 94.50%, which was 12.22 percentage points higher than that of ResNet without domain adaptation.In the ablation experiments, the VMD combined denoising module and the JMMD domain adaptation module contributed accuracy gains of 4.13 and 9.04 percentage points respectively, proving that both signal preprocessing and domain alignment were indispensable in intelligent diagnosis of agricultural equipment. VCRJ-Net effectively improved the generalization ability and identification accuracy of fault diagnosis for tea dryer fans, and provided a feasible technical solution for intelligent operation and maintenance of tea processing equipment.

Key words: time-frequency analysis, transfer learning, tea dryer, fan fault diagnosis, intelligent diagnosis technology

中图分类号: