基于Attention-CNN的振动信号电机转子断条识别  

Vibration Signal Used Motor Broken Rotor Bar Identification Based on Attention-CNN

作　　者：申海锋 石颉 杜国庆 吴宏杰[1,2] SHEN Haifeng;SHI Jie;DU Guoqing;WU Hongjie(School of Electronics and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009;Jiangsu Provincial Key Laboratory of Building Intelligent Energy Conservation,Suzhou University of Science and Technology,Suzhou 215009)

机构地区：[1]苏州科技大学电子与信息工程学院,苏州215009 [2]苏州科技大学江苏省建筑智慧节能重点实验室,苏州215009

出　　处：《电气工程学报》2024年第2期9-15,共7页Journal of Electrical Engineering

基　　金：国家自然科学基金(62073231);江苏省研究生实践创新计划(SJCX22_1581)资助项目。

摘　　要：针对基于振动信号的转子断条诊断技术依赖于人工特征选择,泛化能力差,以及常规卷积神经网络(Convolution neural network,CNN)模型在对时序信号自动特征提取时忽略序列信息的问题,利用Attention机制对局部特征在整体表达时的度量能力,提出了一种Attention-CNN网络模型。首先通过Attention在原始信号上分配注意力,其次结合CNN进行特征提取构建网络,然后利用粒子群优化算法(Particle swarm optimization,PSO)进行网络超参数寻优,训练转子断条识别模型,最后从整体和局部两个方面进行模型评价。试验结果表明,所提出的识别模型能够达到传统诊断水平,且泛化能力高于现有方法,更适用于通过振动信号进行电机转子断条识别。Aiming at the problems that the vibration signal-based broken rotor bar diagnosis technique relies on manual feature selection with poor generalization ability and conventional convolution neural network(CNN)models that ignore sequence information in automatic feature extraction of time-series signals,an Attention-CNN network model is proposed using the Attention mechanism for the metric ability of local features in the overall expression.Firstly,attention is assigned on the original signal by Attention,secondly,the network is constructed by combining CNN for feature extraction,then the particle swarm optimization(PSO)algorithm is used to perform network hyperparameter search and train the broken rotor bar recognition model,and finally,the model is evaluated from both overall and local aspects.The experimental results show that the proposed recognition model can reach the traditional diagnosis level and has a higher generalization capability than existing methods,and is more suitable for motor broken rotor bar recognition through vibration signals.

关 键 词：Attention-CNN 振动信号 转子断条 泛化能力 

分 类 号：TM930[电气工程—电力电子与电力传动]