基于稀疏自动编码深度神经网络的感应电动机故障诊断

Induction Motor Fault Diagnosis Based on Deep Neural Network of Sparse Auto-encoder

针对目前感应电动机故障诊断大多采用监督学习提取故障特征的现状,提出一种将去噪编码融入稀疏自动编码器的深度神经网络,实现非监督学习的特征提取并用于感应电动机的故障诊断。稀疏自动编码器通过自动学习复杂数据的内在特征来提取简明的数据特征表达。为提高特征表达的鲁棒性,在稀疏编码器的基础上融入去噪编码,提取更有效的特征表达用来训练神经网络分类器进而完成整个深度神经网络的构建,并结合反向传播算法对深度神经网络进行整体微调,提升故障分类的准确度。整个训练过程引入'dropout'训练技巧,减少因过拟合带来的预测误差。试验结果表明,相比传统反向传播(Back propagation,BP)神经网络,提出的深度神经网络能更有效地实现感应电动机故障诊断。

To overcome the drawback of using supervised learning to extract fault features for classification in most of current induction motor fault diagnosis approaches, adeep neural network algorithm is presented, which is realized by the sparse auto-encoder combined with the denoising auto-encoder, to achieve unsupervised feature learning for fault diagnosis of induction motors. Sparse auto-encoder can learn the inherent features and extract the succinct expressions from complex data automatically. In addition, the method of denoising auto-encoder can increase the robustness of feature expression, thus improving the performance of the sparse auto-encoder. The extracted features can then be used to train a neural network classifier and complete the deep neural network construction. The back-propagation algorithm is used for fine-tuning the deep neural network with the purpose of improving the accuracy of fault classification. The “dropout” technique is also introduced into to the entire training process to reduce the prediction error caused by “overfitting”. Experimental results have shown that, compared with the traditional back propagation （BP） neural network, the presented deep neural network can realize induction motor fault diagnosis more effectively.