基于卷积神经网络与CatBoost的轴承故障诊断算法  

Bearing fault diagnosis algorithm based on CNN and CatBoost

作　　者：鲁夕瑶 张成彬[1] 皋军[1] 徐燕萍[1] 邵星[1] LU Xi-yao;ZHANG Cheng-bin;GAO Jun;XU Yan-ping;SHAO Xing(College of Information Engineering,Yancheng Institute of Technology,Yancheng 224051,China;College of Mechanical Engineering,Yancheng Institute of Technology,Yancheng 224051,China)

机构地区：[1]盐城工学院信息工程学院,江苏盐城224051 [2]盐城工学院机械工程学院,江苏盐城224051

出　　处：《机电工程》2023年第5期715-722,共8页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金资助项目(61375001,61502411)。

摘　　要：使用一般诊断算法对滚动轴承进行故障排查时,需要对数据进行特征提取,而在其特征提取过程中存在数据量庞大、手动提取和选择受限的问题,为此,提出了一种基于卷积神经网络(CNN)与CatBoost的混合分类模型(方法),以进行轴承的故障诊断。首先,将预处理后的数据经过CNN提取的特征作为输入量,输入到该模型中,提取了训练后输出的模型参数;然后,使用CatBoost方法对滚动轴承数据集进行了分析,进一步研究了不同学习模型在同一数据集下对分类精度的影响;最后,通过降低过拟合的风险,运用4种相关系数指标进行了对比实验,研究了CNN-CatBoost混合分类模型对滚动轴承故障数据的分类效果。研究结果表明:基于CNN与CatBoost方法进行轴承故障诊断的平均准确率达到98%以上,该方法的有效性得到了验证;采用少量的数据训练样本即可达到较好的轴承故障数据分类效果,与单一深度学习模型和一些典型机器学习模型相比,该模型具有更好的性能。When using general diagnostic algorithms for fault ranking of rolling bearings,feature extraction of the data was required,and there were defects in feature extraction in terms of large amount of data and limited manual extraction and selection.Aiming at this problem,a hybrid classification model based on convolutional neural network(CNN)and CatBoost was proposed.Firstly,the pre-processed data extracted by CNN was fed into this model with the features extracted as the input quantity to extract the model parameters output after training.Then,the rolling bearing dataset was analyzed using CatBoost method to further investigate the effect of different learning models on the classification accuracy under the same dataset.Finally,by reducing the risk of overfitting and applying four correlation coefficient indicators,a comparative experiment was conducted to study the classification effect of CNN-CatBoost hybrid classification model on rolling bearing data.The results indicate that the average accuracy of the proposed method is over 98%,which validates the effectiveness of the proposed method;a small number of data training samples can produce better classification results for bearing fault data,and has better performance in the Case Western Reserve University(CWRU)public bearing dataset compared with a single deep learning model and some typical machine learning models.

关 键 词：卷积神经网络 CatBoost算法 故障特征提取 故障分类精度 深度学习模型 训练时间 

分 类 号：TH133.33[机械工程—机械制造及自动化]