面向回转机组电机小样本复合故障的多源异构自适应迁移学习  

Motor Compound Fault Adaptation Transfer Learning of Rotary Machinery for Multi-source Signal and Small Samples

作　　者：巩晓赟 智泽恒 杜文辽 韩明[2] 胡亚凯 罗双强 GONG Xiaoyun;ZHI Zeheng;DU Wenliao;HAN Ming;HU Yakai;LUO Shuangqiang(School of Mechanical and Electrical Engineering,Zhengzhou University of Light Industry,Zhengzhou Henan 450002,China;Anyang Cigarette Factory,China Tobacco Industry Co.,Ltd.,Anyang Henan 455004,China)

机构地区：[1]郑州轻工业大学机电工程学院,河南郑州450002 [2]河南中烟工业有限责任公司安阳卷烟厂,河南安阳455004

出　　处：《机床与液压》2024年第3期209-216,共8页Machine Tool & Hydraulics

基　　金：国家自然科学基金项目(52275138);河南省留学择优资助项目(20221803);河南中烟工业有限责任公司科技创新项目(AW2023024)。

摘　　要：针对单源信号对回转机组电机多点复合故障信息表征不充分及复合故障信号小样本问题,提出一种小样本下电机复合故障的多头卷积神经网络迁移学习模型,实现小样本下电机复合故障的多源异构迁移诊断。将动力装置中电流、振动等多源原始数据作为输入,构造超参数优化的多头卷积神经网络模型。将大样本单故障的原始数据集作为源域,构建目标域下以原始数据为输入的电机小样本复合故障迁移网络模型。将正则化惩罚项应用到迁移学习模型中,构建模型目标函数参数更新准则,实现模型对源域与目标域参数的自适应更新配适。试验结果表明:单源信息的诊断可靠性依赖于数据源的选取,多源信号的多头卷积神经网络模型可有效融合电流、振动信号并实现特征提取。通过与多个模型比对,所提方法在小样本下对电机复合故障的识别精度显著提升,且收敛时间缩短近2/3。To solve the problem of insufficient representation in single source signal for different position compound fault information and small sample of compound fault signal in rotary machinery motor,a multi-head convolutional neural network transfer learning model for motor compound fault based small samples was proposed,to realize the multi-source heterogeneous migration diagnosis of motor compound fault under small samples.Taking the original data of current,vibration and other sources in the power units as input,a model of multi-head convolutional neural network by hyperparameter optimization was constructed.Then,the original dataset of the large sample single fault was taken as the source domain,and the small sample compound fault transfer learning network model of the motor with the original data as the input under the target domain was constructed.Finally,the regularization penalty term was applied to the transfer learning model,and a criterion for updating the model's objective function parameters was constructed,to achieve adaptive updating and adaptation of the model parameters between the source and target domains.Experimental results show that the diagnostic reliability of single source information depends on the selection of data sources,the information fusion and feature extraction of current and vibration signals is realized by the multi-head convolutional neural network mode.Compared with other models,the proposed method can significantly improve the recognition accuracy of motor compound fault based small samples,and the convergence time is shortened by nearly 2/3.

关 键 词：感应电机 复合故障 小样本 多头卷积神经网络 迁移学习 

分 类 号：TH17[机械工程—机械制造及自动化] TM346[电气工程—电机]