计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (13): 165-171.DOI: 10.3778/j.issn.1002-8331.1805-0269

• 模式识别与人工智能 • 上一篇    下一篇

雷达目标检测深层自编码器自适应优化算法

侯  旋1,2,陈  涛3,王唯良1   

  1. 1.西北政法大学 新闻传播学院,西安 710122
    2.空军工程大学 航空工程学院,西安 710038
    3.电子科技大学 电子科学技术学院,成都 611731
  • 出版日期:2019-07-01 发布日期:2019-07-01

Deep Atuoecoder Adaptive Learning Algorithm on Radar Target Detection

HOU Xuan1,2, CHEN Tao3, WANG Weiliang1   

  1. 1.School of Journalism and Communication, Northwest University of Politics and Law, Xi’an 710122, China
    2.College of Aeronautics Engineering, Air Force Engineering University, Xi’an 710038, China
    3.College of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu 611731, China
  • Online:2019-07-01 Published:2019-07-01

摘要: 研究了现阶段雷达低小慢目标探测技术的难点与方法。分析了深层自编码器基本模型与算法,通过引入自适应学习理论,提出了基于Rumelhart函数的深层自编码器自适应算法(RDAAA),并证明了算法的收敛性。优化算法避免了网络训练过程中出现惩罚过度的现象,克服了学习速率过高导致网络振荡发散,或学习速率过小降低网络收敛速度等缺陷。利用两种数据集对RDAAA、基于交叉熵函数的深层自编码器学习算法(CDAA)与误差反向传播算法(BPA)进行模式识别能力分析,结果表明在确定限定误差与选取最佳学习速率的情况下,RDAAA相对于CDAA与BPA收敛速度最快,正确识别率更高。围绕雷达目标检测与深度学习理论,分析了低小慢目标特性,将目标检测问题转化为模式分类问题,利用上述三种算法进行目标检测仿真实验,结果表明RDAAA与CDAA的性能明显优于BPA,且RDAAA的检测率更高,特别是处于低信噪比阶段,仍可保持较高的发现概率。

关键词: 目标检测, 低小慢目标, 深度学习, 自动编码器, 自适应优化算法

Abstract: The difficulties and methods of radar low-small-slow target detection technology are studied. It analyzes the basic model and algorithm of deep autoencoder. By introducing adaptive learning theory, Rumelhart function-Deep Atuoecoder Adaptive learning Algorithm(RDAAA) is proposed, and the convergence of the algorithm is proved. The optimization algorithm avoids the phenomenon of excessive punishment in the network training process, overcomes the disadvantages of excessively high learning rate leading to oscillating and divergent in network or low learning rate leading to reduce network convergence speed. Two types of data sets are used to analyze the pattern recognition ability of RDAAA, Cross-entropy function-Deep Autoencoder learning Algorithm(CDAA) and error Back Propagation Algorithm(BPA). In the case of determining the limit error and selecting the optimal learning rate, the results show that RDAAA has the fastest convergence rate and higher correct recognition rate than CDAA and BPA. Focusing on radar target detection and deep learning theory, the characteristics of low-small-slow target are analyzed, and the target detection problem is transformed into a problem of pattern classification. Using the above three algorithms for target detection simulation experiments, the results show that the performance of RDAAA and CDAA is significantly better than that of BPA, and the detection rate of RDAAA is higher, especially in the low signal-to-noise ratio stage, and the high probability of discovery can still be maintained.

Key words: target detection, low-small-slow target, deep learning, autoencoder, adaptive optimization algorithm