Online Bearing Fault Diagnosis using Support Vector Machine and Stacked Auto-Encoder
Online Bearing Fault Diagnosis using Support Vector Machine and Stacked Auto-Encoder
复制标题
DOI：
10.1109/icphm.2018.8448775
复制
发表时间：
2018-06
期刊：
影响因子：
--
通讯作者：
Wentao Mao;Siyu Tian;Xihui Liang;Jianliang He
中科院分区：
文献类型：
--
作者：
Wentao Mao;Siyu Tian;Xihui Liang;Jianliang He
得益于物联网等先进测量技术的快速发展，旋转机械在线故障诊断受到学术界和工程界的广泛关注。与目前大多数方法对大量现有振动数据进行诊断不同，在线诊断需要从顺序收集的数据中查找是否发生故障或裂纹。在这种情况下，需要利用新收集的数据来更新诊断模型。此外，为了保持在线诊断的准确性，需要快速提取最具代表性的故障特征。针对这两个任务，本文提出了一种基于增量支持向量机（ISVM）和堆叠式自动编码器（SAE）深度学习算法的在线轴承故障诊断方法。该方法包含两个阶段。离线阶段，利用SAE自动提取故障数据的代表性特征，然后利用ISVM构建离线诊断模型。在在线阶段，对于每个新到达的数据块，引入离线SAE模型的网络权重来直接生成在线故障特征向量，然后通过整合这些特征来更新在线诊断模型以合并新数据。在CWRU和IMS轴承故障数据集上的实验结果表明，与典型的传统特征提取和诊断方法相比，该方法能够快速达到令人满意的诊断性能。
Thanks to the quick development of the Internet of Thing and other advanced measuring techniques, online fault diagnosis for rotating machinery has received much attention in the fields of academy and engineering. Different from the most of current methods which make diagnosis on a bunch of existing vibration data, online diagnosis needs to find whether a fault or crack occurred from the data which are collected sequentially. In this scenario, it needs to update the diagnosis model by means of the new collected data. Moreover, to maintain the online diagnosis accuracy, it needs to extract the most representative fault features quickly. Following these two tasks, a new online bearing fault diagnosis method is proposed in this paper based on Incremental Support Vector Machine(ISVM) and Stacked Auto-Encoder(SAE) which is a kind of deep learning algorithm. This method contains two stages. At offline stage, the fault data are utilized to automatically extract the representative features by means of SAE, and then an offline diagnosis model is constructed by using ISVM. At online stage, for each new arrived data block, the network weight of offline SAE model is introduced to generate online fault feature vectors directly, and then the online diagnosis model is updated to incorporate the new data by integrating these features. Experimental results on CWRU and IMS bearing fault data sets show that, the proposed method can reach the satisfactory diagnosis performance at quick speed compared to the typical traditional feature extraction and diagnosis methods.