基于稀疏自编码的深度故障诊断方法与研究  

Deep Fault Diagnosis Method based on Sparse Auto-Encoder

作　　者：张海天 陈斌[1] 刘程 瞿珊瑚 ZHANG Hai-tian;CHEN Bin;LIU Cheng;QU Shan-hu(College of Electronic Engineering,Naval Univ.of Engineering,Wuhan Hubei 430033,China)

机构地区：[1]海军工程大学电子工程学院,湖北武汉430033

出　　处：《通信技术》2018年第10期2436-2442,共7页Communications Technology

摘　　要：针对目前短波发信机故障诊断主要采用人工规则系统、专家系统、监督学习方法提取特征的现状,提出了一种基于稀疏自编码深度神经网络的方法进行无监督特征提取。该方法首先将收集的发信机信号通过稀疏自编码进行训练获得参数,其次通过神经网络进行分类,采用反向传播的方法对算法进行微调以提高准确度,最后用训练的网络对短波发信机各功放模块的故障进行识别。实验结果表明:相比于传统的诊断方法,提出的算法具有更高的准确率和稳定性。At present, the fault diagnosis of short-wave transmitters mainly relies on artificial rule systems, expert systems, and supervision learning methods for features extraction. Aiming at this situation, an unsupervised feature extraction based on sparse auto-encoder depth of the neural network is proposed. The method first trains the collected transmitter signals through sparse auto-encoding and acquires the parameters. Then, the neural network is used for classification, and the algorithm is fine-tuned by back- propagation so as to improve the accuracy. Finally, the fault of each power amplifier module for the short wave transmitter is identified by the trained network. The experimental results indicate that the proposed algorithm has higher accuracy and stability than traditional diagnostic methods.

关 键 词：稀疏自编码 深度神经网络 短波发信机 故障诊断 准确率 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]