多尺度多任务注意力卷积神经网络滚动轴承故障诊断方法  

Rolling bearing fault diagnosis with multi-scale multi-task attention convolutional neural network

作　　者：王照伟 刘传帅 赵文祥[1] 宋向金 WANG Zhaowei;LIU Chuanshuai;ZHAO Wenxiang;SONG Xiangjin(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏大学电气信息工程学院,江苏镇江212013

出　　处：《电机与控制学报》2024年第7期65-76,共12页Electric Machines and Control

基　　金：国家自然科学基金杰出青年科学基金(52025073);国家自然科学基金青年科学基金(52007078,62002140);江苏省自然科学基金(BK20200887)。

摘　　要：针对振动信号时间尺度不一、故障特征分布差异及信息冗余等问题,提出一种多尺度多任务注意力卷积神经网络(MSTACNN)的滚动轴承故障诊断方法。首先,在参数共享单元构建多尺度卷积神经网络,提取多任务之间共享信息的多尺度通用特征;其次,利用多任务学习机制对故障类型、故障尺寸以及运行工况同时训练,规避单任务学习效率低下问题;然后,采用注意力机制对多尺度特征信息进行筛选,识别并保留有效特征;最后,设计了一种自适应损失权重算法,动态调整子任务的损失权重,控制不同任务的学习进度,实现了对轴承故障类型、故障尺寸以及运行工况同时识别的目标。在西储大学数据集和帕德博恩大学数据集分别对方法有效性进行验证,其中故障类型的识别准确率分别达到了99.95%和98.41%。实验结果表明,所提方法均展现出较高的识别准确率、良好的收敛速度和稳定性,具有较强的特征提取学习能力和泛化性能。Aiming at the problems of different time scales,inconsistent characteristic distribution,and information redundancy of vibration signals,a rolling bearing fault diagnosis method with a multi-scale multi-task attention convolutional neural network(MSTACNN)was proposed.Firstly,a multi-scale convolutional neural network was constructed in the parameter sharing unit,and multi-scale common features containing information shared between different tasks in vibration signals were extracted.Secondly,the multi-task learning mechanism was employed to simultaneously accomplish three tasks:fault type,fault size,and operation conditions.Thus,the inefficiency of single-task learning was solved.Then,the attention mechanism was used to enhance the feature expression and the influence of useless information was eliminated.Finally,an adaptive loss weight algorithm was designed to dynamically adjust the loss weight and the learning progress of three tasks,the goal of simultaneously identifying bearing fault type,fault size,and operating conditions was achieved.The effectiveness of the proposed method was verified in the dataset of Western Reserve University and the University of Paderborn,respectively.The recognition accuracy of fault types achieved 99.95%and 98.41%in different datasets.The experimental results show that the proposed method shows high recognition accuracy,good convergence speed and stability,which proves that the proposed method has strong feature extraction learning ability and generalization performance.

关 键 词：多尺度卷积 注意力机制 多任务学习 自适应损失权重 故障诊断 

分 类 号：TM34[电气工程—电机]