基于Fine-tune与DDC的变工况数控设备部件故障诊断  

Fault Diagnosis of CNC Equipment Components Under Variable Operating Conditions Based on Fine-tune and DDC

作　　者：王渤 杨越 陆剑峰[1] 余涛 颜鼎峰 徐煜昊 WANG Bo;YANG Yue;LU Jianfeng;YU Tao;YAN Dingfeng;XU Yuhao(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China;Intelligent Cloud Information Technology Co.,Ltd.,Shanghai 200090,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804 [2]智能云科信息科技有限公司,上海200090

出　　处：《机床与液压》2024年第22期22-29,共8页Machine Tool & Hydraulics

基　　金：国家自然科学基金面上项目(72171173);同济大学“创新设计与智能制造学科群”项目(F2206)。

摘　　要：针对复杂工业环境下的数控设备部件故障诊断数据样本少、变工况诊断困难和准确率不高等问题,提出一种基于模型迁移的故障诊断方法。利用连续小波变换对不同工况下的原始振动数据进行预处理,建立二维时频数据集,并分为源域与目标域;利用源域数据集与CNN进行模型预训练;分别引入微调(Fine-tune)与深度域混淆(DDC)2种迁移学习方式改进模型;最终实现了基于Fine-tune与基于DDC的故障诊断模型的构建。以轴承与数控铣刀2种部件为例进行实验验证,结果证明:Fine-tune与DDC均可以有效提高数控设备部件的故障诊断准确率,其中Fine-tune的泛化能力强,而DDC训练耗时更短且在复杂环境下的性能更优。Aiming at the problems of few fault diagnosis data samples of CNC equipment components in complex industrial environment,difficulty in diagnosing faults under variable operating conditions and low accuracy,a fault diagnosis method based on model migration was proposed.Continuous wavelet transform was used to preprocess the original vibration data under different operating conditions,and a 2D time-frequency dataset was established,which was divided into source domain and target domain.The source domain dataset and CNN were used for pre-training.Then,two transfer learning methods of Fine-tune and deep domain confusion(DDC)were introduced to improve the model.Finally,the fault diagnosis models based on Fine-tune and DDC were constructed.Taking the bearing and CNC milling cutter as example,the results show that Fine-tune and DDC can effectively improve the fault diagnosis accuracy of CNC equipment components,among which Fine-tune has strong generalization ability,while DDC takes less training time and performs better in complex environments.

关 键 词：故障诊断 变工况 卷积神经网络 Fine-tune 深度域混淆(DDC) 

分 类 号：TH17[机械工程—机械制造及自动化]