基于深度迁移学习的滑动轴承-转子故障诊断  

Fault Diagnosis of Journal Bearing-rotor Systems Based on Deep Transfer Learning

作　　者：朱琰 黄敏 王小静[1] 郑成东 ZHU Yan;HUANG Ming;WANG Xiaojing;ZHENG Chengdong(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China;Shanghai Marine Equipment Research Institute,Shanghai 200031,China)

机构地区：[1]上海大学机电工程与自动化学院,上海200444 [2]上海船舶设备研究所,上海200031

出　　处：《噪声与振动控制》2022年第4期107-114,共8页Noise and Vibration Control

基　　金：国家自然科学基金资助项目(52075311);上海市智能制造及机器人重点实验室资助项目。

摘　　要：深度神经网络在滚动轴承故障诊断领域得到广泛应用,而将深度神经网络应用于滑动轴承-转子系统故障诊断的研究较少。其中大多数研究假设轴承故障的训练数据与测试数据分布相同,基于该训练数据得到的神经网络能较好地对轴承故障进行描述,但当轴承转子系统的结构和工况发生变化,原神经网络就不能对故障进行准确诊断。提出一种基于改进型联合分布差异(Improved Joint Distribution Discrepancy,IM-JDD)方法的深度卷积迁移学习框架(Deep Convolutional Transfer Learning Network,DCTLN),该框架采用二维振动图像作为网络输入,通过深度卷积神经网络提取图像的可迁移特征,提出的改进型联合分布差异方法实现了不同结构及工况下滑动轴承-转子系统故障特征的迁移学习。最后在结构不同的滑动轴承-转子实验台上进行测试,结果表明,本框架在不同工况下和不同机器间对无标记故障样本具有较强的诊断能力,并优于其他竞争方法。Deep neural network has been widely used in the field of rolling bearing fault diagnosis while there are few researches of the fault diagnosis for journal bearing-rotor systems based on deep neural network.Most of these studies assume that the distribution of the training data for bearing failures is the same as that of the test data.The neural network based on the training data can describe the bearing faults well.However,when mechanical structure or working condition changes,the original network cannot accurately classify faults.This paper proposes a Deep Convolutional Transfer Learning Network(DCTLN)framework based on the Improved Joint Distribution Discrepancy(IM-JDD)method with two-dimensional vibration images as the network input.The IM-JDD method can not only realize the transference of the fault diagnosis knowledge between the sliding bearing-rotor systems with different structures and working conditions,but also adapts the joint probability distributions of the features extracted by the deep convolutional neural network(CNN).Finally,extensive experiments were carried out on the journal bearing-rotor test bench with different structure to validate the effectiveness of the proposed framework.The results show that the framework has a strong diagnostic ability for unlabeled fault samples under different working conditions and different machines,and is better than other competitive methods.

关 键 词：故障诊断 卷积神经网络 改进型联合分布差异 振动图像 滑动轴承-转子系统 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]