PCA-SAE的齿轮箱故障诊断方法研究  

Research on Gearbox Fault Diagnosis Method of PSO-SAE

作　　者：马芸婷 张超[2] 王宇晨 MAYun-ting;ZHANG Chao;WANG Yu-chen(Inner Mongolia Baogang Steel Union Co.,Ltd.,Rail Beam Rolling Mill,Inner Mongolia Baotou 014010,China;School of Mechanical Engineering,Inner Mongolia University of Science and Technology,Inner Mongolia Baotou 01401,China;Inner Mongolia North Heavy Industry Group Co.,Ltd.technology Center Product Research Institute,Inner Mongolia Baotou 014010,China)

机构地区：[1]内蒙古包钢钢联股份有限公司轨梁轧钢厂,内蒙古包头014010 [2]内蒙古科技大学机械工程学院,内蒙古包头014010 [3]内蒙古北方重工业集团有限公司技术中心产品研究院,内蒙古包头014010

出　　处：《机械设计与制造》2022年第3期144-147,152,共5页Machinery Design & Manufacture

基　　金：国家自然科学基金(51565046)于噪声参数最优ELMD方法多特征融合的风电机组齿轮箱状态监测关键技术研究;内蒙古自治区科技计划项目(2018KG007)协整分析下基于大数据的风电机组齿轮箱故障诊断与性能预测。

摘　　要：针对因长时间的信号采集使得振动信号面临数据量大的问题。传统的信号分析方法,已无法解决大数据情况下故障的特征提取与分类,同时采集到的数据样本具有多维度多样本的情况,导致训练网络时在前期导入数据阶段耗费大量时间与硬件的内存,并且会导致网络训练中产生过拟合现象,影响分类准确率。针对以上问题本文提出基于主成分分析与堆叠自动编码机相结合的齿轮故障诊断研究,以实现对齿轮振动信号快速准确的特征提取与分类。首先对原始信号进行主成分析,得到各主成分贡献率,其次,选取主成分贡献率高的前几列作为深度学习网络输入样本。最后深度学习网络即堆叠自动编码机网络对训练数据集进行学习提取数据中的特征并应用测试数据集部分进行分类并计算分类的准确率。最终,实验中将所提深度学习方法与传统的特征提取方法和分类方法进行比较最终识别精度进行比较。实验结果表明本文所提方法最终可以达到98.6%的准确率,实现端到端的故障诊断方法,可以很好的应用于故障诊断领域。For the long-term signal acquisition,the vibration signal faces a large amount of data. The traditional signal analysis method can not solve the feature extraction and classification of faults in the case of big data. At the same time,the collected data samples have multi-dimensional and multi-sample situations,which leads to a large amount of time and hardware memory in the early stage of importing data during training network. And it will lead to over-fitting in network training,affecting the classification accuracy. In view of the above problems,it proposes a gear fault diagnosis based on principal component analysis combined with stacking automatic encoder to achieve fast and accurate feature extraction and classification of gear vibration signals.First,the main signal is analyzed by the main component to obtain the contribution rate of each principal component. Secondly,the first few columns with high principal component contribution rate are selected as the input samples of the deep learning network.Finally,the deep learning network,that is,the stacking automatic encoder network,learns the characteristics of the training data set and extracts the features in the data and applies the test data set to classify and calculate the classification accuracy.Finally,the experiment compares the proposed deep learning method with the traditional feature extraction method and classification method to compare the final recognition accuracy.The experimental results show that the proposed method can achieve 98.6%accuracy and achieve end-to-end fault diagnosis method,which can be applied to the field of fault diagnosis.

关 键 词：故障诊断 主成成分分析 深度学习 堆叠自动编码机网络 

分 类 号：TH16[机械工程—机械制造及自动化] TP260[自动化与计算机技术—检测技术与自动化装置]