北京航空航天大学学报

• •    

基于改进神经网络的航天器电信号分类方法

王全鑫1,李可1,王浚2,孙毅2,宋世民2   

  1. 1. 北京航空航天大学
    2.
  • 收稿日期:2015-03-31 修回日期:2015-09-14 发布日期:2015-10-30
  • 通讯作者: 李可
  • 基金资助:
    航空基金;中央高校基本科研业务费

Satellite Electrical Signal Classification Based On Improved Neural Networks

  • Received:2015-03-31 Revised:2015-09-14 Published:2015-10-30

摘要: 本文根据航天器系统级电性能测试工作中数据量大、任务繁重的特点,设计了基于人工神经网络的智能分类系统,对原始测试数据进行智能化分类,将调试经验以数据的形式储备,可在减少测试工作中对人为经验依赖的同时给航天器信号识别快速提供专家知识。考虑到经典的神经网络系统有训练时间长和对网络初始权值的依赖程度高等不足;为此本文利用主成分分析对数据进行压缩和自动编码技术对网络权值进行初始化。实验数据测试表明:与传统方法相比,该改进学习系统的准确率、稳定性和响应速度均得到显著提高。

关键词: 故障诊断, 神经网络, 模式识别, 自动编码, 电信号, 梯度下降法

Abstract: This paper designs an intelligent classification system based on artificial neural networks. The purpose is to solve the problem of multiple data,arduous task in the aircraft test and intellectualize the management of the testing work. In this way, the system can reduce the workload and the reliance on testing experience as well as can store the nonlinear debugging experience in the form of expert database. This system has many deficiencies, such as, long training time and highly dependent the initial threshold. To this end, the paper uses principal component analysis to compress the raw data and Auto-Encoder that has been successfully applied in deep learning to initialize the network weights. Experimental data indicate: Compared with traditional methods, accuracy, stability and response speed of the improved learning system were significantly increased.

Key words: Fault Diagnosis, ANN, Pattern Recognition, Auto-Encoder, electronic signal, Gradient descent method

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发