一种基于嵌套式深度学习模型的电机异常检测方法  

A motor anomaly detection method based on nested deep learning models

作　　者：岑跃峰 李旭成 岑岗[1] 赵澄 CEN Yuefeng;LI Xucheng;CEN Gang;ZHAO Cheng(School of Information and Electronic Engineering,Zhejiang University of Science and Technology,Hangzhou 310023,China;College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江科技学院信息与电子工程学院,浙江杭州310023 [2]浙江工业大学信息工程学院,浙江杭州310023

出　　处：《浙江工业大学学报》2025年第2期189-196,共8页Journal of Zhejiang University of Technology

基　　金：浙江省教育厅一般科研项目(Y202249481)。

摘　　要：针对电机生产检测环节中对异常检测方法提出的高准确率及高泛化性需求,提出了一种基于嵌套式深度学习模型的方法来实现电机的异常检测。首先通过卷积模型(Convolutional neural networks,CNN)和自注意力模型(Self-attention)组成集成学习,提取相邻时序的特征;然后引入残差网络丰富原始振动特征,以提高模型泛化性;最后通过时序卷积模型(Temporal convolutional networks,TCN)输出最终检测结果。实验结果表明:笔者所提出方法的平均准确率达到了95.93%,且模型具有较好的泛化能力,所提出的异常检测方法能很好地在小波变换、傅里叶变换、小波包分解等特征提取方法无法生效的场景中发挥作用,为工业电机自动化生产提供参考。A nested deep learning method is proposed to analyze the motor vibration signal to achieve anomaly detection in the motor production inspection process.Firstly,the adjacent time-series features are extracted by the ensemble learning composed of convolutional neural networks and self attention model.Secondly,the original features are enriched by the introduction of residual network in order to improve the generalization of the model.Finally,the classification is performed by temporal convolutional networks based on the newly obtained features.The experimental results show that the average accuracy of the proposed method reaches 95.93%,and the model has good generalization ability.The proposed anomaly detection method can work well in scenarios where feature extraction methods such as continuous wavelet transform,Fourier transform,and wavelet packet decomposition cannot take effect,and provides some reference value for the application of industrial motor automated production.

关 键 词：电机 异常检测 深度学习 

分 类 号：TM341[电气工程—电机] TP183[自动化与计算机技术—控制理论与控制工程]