一种可解释性空时模型的风力发电机轴承智能诊断新框架  

A new intelligent diagnosis framework for wind power insulated bearings based on spatio-temporal models of interpretable lightweight

作　　者：李学军 刘治新 杨同光 韩清凯[2] 蒋玲莉 Li Xuejun;Liu Zhixin;Yang Tongguang;Han Qingkai;Jiang Lingli(Key Laboratory of Rotor Vibration Monitoring and Diagnosis Technology for Machinery Industry,Foshan University,Foshan 528225,China;Key Laboratory of Rotor Vibration Monitoring and Diagnosis Technology for Machinery Industry,School of Foshan Graduate Innovation Northeastern University,Foshan 528225,China)

机构地区：[1]佛山大学机械工业转子振动监测及诊断技术重点实验室,佛山528225 [2]东北大学佛山研究生创新学院机械工业转子振动监测及诊断技术重点实验室,佛山528225

出　　处：《仪器仪表学报》2025年第2期51-69,共19页Chinese Journal of Scientific Instrument

基　　金：广东省基础与应用基础研究基金(2023A1515240083);广东省基础与应用基础研究基金(2024A1515240061)项目资助。

摘　　要：针对大功率变频风力发电机轴承故障特征难以挖掘以及现有深度学习模型存在可解释性差的关键难题。开发了一种轻量化空时信息融合模型的智能诊断新框架,命名为BSTA-Net,其着眼于解决实际工程中风力发电机轴承故障难以识别的问题。首先,设计了轴承故障特征空时信息融合模块,并创造性地开发了一种双向的时序信息特征融合新策略,将该策略巧妙运用到所提BSTA-Net框架中,进而充分提取故障数据中的细粒度特征,并在风力发电机轴承状态监测中实现首次尝试。其次,在所提框架中引入特征聚焦模块进行优化,使其能够精准的充分注意到重要的信息,抛弃无用的故障敏感特征,使得所提框架在交变电压冲击和变载荷等复杂工况下,依然具备优秀的学习能力。最后,基于同一数据集,从多个维度对BSTA-Net框架等8种方法的诊断性能进行了对比分析,并将诊断结果与BST-Net等7种方法进行对比分析,结果表明,所提框架具有良好的优越性和泛化性,此项研究为轴承故障识别提供了新思路。将t-SNE和显著性区域检测技术引入所提BSTA-Net框架对故障特征挖掘过程进行物理归因解释,进而提升框架在决策过程中的可信赖性。Mining bearing fault characteristics in high-power variable-frequency wind turbines is challenging,and existing deep learning models suffer from poor interpretability.To address these issues,a new intelligent diagnosis framework of lightweight space-time information fusion model,named BSTA-Net,is developed to enhance bearing fault identification in practical engineering applications.Firstly,a bearing fault feature space-time information fusion module is designed,and a new bidirectional timing information feature fusion strategy is creatively developed.The strategy is cleverly applied to the proposed BSTA-Net framework to fully extract the finegrained features from the fault data,marking the first attempt to apply such an approach to wind turbine bearing condition monitoring.Secondly,the feature focusing module is introduced into the proposed framework for optimization,enabling it to effectively prioritize critical fault-related information while discarding irrelevant or noisy features.This ensures that the model maintains robust learning capabilities even under complex conditions such as alternating voltage shocks and variable loads.Finally,based on the same data set,the diagnostic performance of 8 methods such as BSTA-Net framework is compared from multiple dimensions,and the diagnostic results are compared with 7 methods such as BST-Net.The results show that the proposed framework exhibits superior superiority and strong generalization ability,providing a new idea for bearing fault identification.Furthermore,T-SNE and significance region detection technology are introduced into the BSTA-Net framework to explain the physical attribution of fault feature mining process,thereby improving the reliability of the framework in the decision-making process.

关 键 词：智能诊断 轴承 空时模型 可解释性 

分 类 号：TH165.3[机械工程—机械制造及自动化]