基于GADF-MDSC的特大型轴承深度迁移故障诊断方法  

Deep transfer fault diagnosis method for extra-large bearings based on GADF-MDSC

作　　者：姜烨飞 王华[1] 潘裕斌[1] 王天祥 傅航 JIANG Yefei;WANG Hua;PAN Yubin;WANG Tianxiang;FU Hang(College of Mechanical and Power Engineering,Nanjing Tech University,Nanjing 211800,China;Soter Transmission Equipment Co.,Ltd.,Changshu 215500,China)

机构地区：[1]南京工业大学机械与动力工程学院,南京211800 [2]索特传动设备有限公司,江苏常熟215500

出　　处：《振动与冲击》2024年第19期10-18,共9页Journal of Vibration and Shock

基　　金：国家自然科学基金(52205106);江苏省自然科学基金青年项目(BK20210547)。

摘　　要：针对工程应用中特大型轴承运行工况复杂以及故障数据匮乏,导致其故障特征提取不全面的问题,提出了一种基于格拉姆角差场-多尺度深度可分离卷积(Gramian angular difference field-multi-scale depthwise separable convolutions,GADF-MDSC)的特大型轴承深度迁移智能诊断方法。首先,构建GADF-MDSC故障诊断网络,该网络分为三大模块:图像转换、特征提取、输出部分。图像转换模块采用GADF编码方式将振动信号转换为二维图像;特征提取模块通过MDSC提取综合故障特征信息,并利用双向门控循环单元筛选融合特征;输出部分由Softmax函数预测轴承故障类型的概率分布。然后,利用源域数据预训练模型,将预训练模型权重参数作为目标域训练模型初始化参数,冻结除底层外的所有参数,使用目标域数据微调模型,实现深度迁移故障诊断任务。最后,通过两种特大型轴承试验对深度迁移模型进行验证。试验结果表明,所提方法在目标域样本仅有5.00%的条件下,仍能保证较高的跨工况精度,达到86.04%,且迁移效果优于其他方法。Here,aiming at the problem of extra-large bearings’complex operating conditions and insufficient fault data in engineering applications to cause incomplete fault feature extracting,a deep transfer intelligent fault diagnosis method for extra-large bearings based on Gramian angular difference field-multi-scale depthwise separable convolutions(GADF-MDSC)was proposed.Firstly,GADF-MDSC fault diagnosis network was constructed,it was divided into 3 modules of image conversion,feature extraction and output.The image conversion module could use GADF encoding method to convert vibration signals into 2D images.The feature extraction module could extract comprehensive fault feature information with MDSC,and use the bidirectional gated recurrent unit to screen fusion features.The output part could predict the probability distribution of bearing fault types using Softmax function.Then,the source domain data were used to pretrain a model,and the pretrained model weight parameters were taken as initialization parameters for the target domain training model.All parameters except bottom layer were frozen,and the target domain data were used to fine tune the model to realize deep transfer fault diagnosis tasks.Finally,the deep transfer model was verified with 2 types of extra-large bearing tests.The test results showed that the proposed method can still ensure higher cross-working condition accuracy to reach 86.04%even when target domain samples have only 5.00%;its transfer effect is better than other methods’.

关 键 词：特大型轴承 故障诊断 迁移学习 格拉姆角差场(GADF) 多尺度深度可分离卷积(MDSC) 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置] TH825[自动化与计算机技术—控制科学与工程]