基于多输入层卷积神经网络的滚动轴承故障诊断模型  

A fault diagnosis model for rolling bearings based on a multi-input layer convolutional neural network

作　　者：昝涛[1] 王辉[1] 刘智豪 王民[1,2] 高相胜 ZAN Tao;WANG Hui;LIU Zhihao;WANG Min;GAO Xiangsheng(Beijing Key Laboratory of Advanced Manufacturing Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Electrical Discharge Machining Technology,Beijing 100191,China)

机构地区：[1]北京工业大学机电学院先进制造技术北京市重点实验室,北京100124 [2]电火花加工技术北京市重点实验室,北京100191

出　　处：《振动与冲击》2020年第12期142-149,163,共9页Journal of Vibration and Shock

基　　金：国家自然科学基金(51575014,51875008,51505012);北京市教委科技计划项目(KM201410005026);国家留学基金(201806545032)。

摘　　要：针对滚动轴承信号易受噪声干扰和智能诊断模型鲁棒性差的问题,在一维卷积网络的基础上,提出基于多输入层卷积神经网络的滚动轴承故障诊断模型。相比传统卷积神经网络诊断模型,该模型具有多个输入层,初始输入层为原始信号,以最大化地发挥卷积网络自动学习原始信号特征的优势;同时可将谱分析数据在模型任意位置输入模型,以提升模型的识别精度和抗干扰能力。通过滚动轴承模拟试验,进行可行性和有效性验证,同时与人工神经网络(Artificial Neural Network,ANN)、支持向量机(Support Vector Machine,SVM)和典型的卷积神经模型进行对比,证明了所提出模型的优势;向测试集中加入噪声来检验模型的鲁棒性,并且运用增量学习方法提升模型在强噪声环境下的识别性能;通过滚动轴承故障实例,验证模型的识别性能和泛化能力。试验结果表明,所提出的模型提升了传统卷积模型的识别率和收敛性能,并具有较好的鲁棒性和泛化能力。Aiming at the problem that rolling bearings signal is susceptible to noise interference and the poor robustness of an intelligent diagnosis model,a fault diagnosis model for rolling bearings based on a multi-input layer convolutional neural network was proposed on the basis of a one-dimensional convolutional network.Compared with the traditional convolutional neural network diagnosis model,the model had multiple input layers.The data of initial input layer was the original signal,in order to maximize the advantages of the convolutional network to automatically learn the original signal characteristics.The spectral analysis data could be input into the network at any position of the model,in order to improve the recognition accuracy and anti-jamming ability of the model.Firstly,through a simulation test of rolling bearing,the feasibility and validity of the proposed method were verified.Then,the robustness of the model was tested by adding noise to the test set,and the recognition performance of the model in strong noise environment was improved by using an incremental learning method.Finally,through the example of rolling bearing fault,the recognition performance and generalization ability of the model were verified.Experimental results show that the proposed model can improve the recognition rate and convergence performance of the traditional convolutional model,and has good robustness and generalization ability.

关 键 词：深度学习 卷积神经网络 滚动轴承 故障诊断 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3