【摘要】 为了进一步提高深度神经网络(Deep Neural Network,DNN)在轴承故障诊断中的可靠性和稳定性,对深度学习(Deep Learning)中的一些关键技术进行了研究、借鉴以及改进。具体地沿用传统DNN中被广泛用于无监督学习的去噪自动编码器(Denoising Auto-encoder,DAE)进行特征提取,使得特征提取过程不再依赖于先验知识;然后对传统DNN中的DAE进行稀疏化处理,使得特征的提取更加合理、准确;并在DNN中引入核函数运算形成多核结构,提高诊断结果的可靠性以及鲁棒性。最后通过具体的实验,与传统DNN、支持向量机(Support Vector Machine,SVM)等故障诊断方法相对比,来最终反映基于稀疏化DAE的多核结构DNN在轴承故障诊断领域更优越的正确率与稳定性。更多还原

【Abstract】 In order to further improve the reliability and stability of the Deep Neural Network(DNN)in bearing fault diagnosis,some key technologies in Deep Learning are studied,borrowed and improved in it. Specifically,the DAE(Denoising Autoencoder)of the traditional DNN is used for the feature extraction,which has been widely used in unsupervised learning,so that the feature extraction process no longer depends on prior knowledge; then the DAE in the traditional DNN is processed into sparsification,to make the feature extraction more reasonable and accurate; and in order to improve the reliability and robustness of the diagnosis results,the kernel functions are introduced in DNN. Finally,by the specific fault diagnosis experiment and effect contrast of DNN proposed in this paper,traditional DNN and Support Vector Machine(SVM)etc.,this paper finally reflects the high accuracy and stability of this kind of proposed DNN that based on sparisification DAE and multiple kernel structure.更多还原