机载雷达杂波抑制的智能网络与学习系统设计方法研究

Clutter suppression is the core problem for the signal processing of airborne radar. Space-time adaptive processing (STAP) is the core technology of the clutter suppression and the moving target detection for airborne radar. However, in practical applications, it suffers the problem that its training samples are insufficient. Aiming at the core difficult problem in the applications of STAP: the difficulty to convergence with insufficient training samples, this proposal carries out new research directions, and explores the design principle of the architecture of intelligent clutter suppression for airborne radar. The following researches are included in this proposal. One is that the design method of STAP intelligent network and learning system with insufficient training samples is proposed. It designs the embedded function, the input and output of the network, and the constraint-loss function of distance compensation for the STAP intelligent system, which makes the unknown distance compensation matrix of STAP participate in the training and adjustment of the system. It breaks through the core problem of insufficient training samples of STAP in applications. The other one is that the design method of intelligent clustering analysis for the elimination of the non independent identically distributed samples, and the design method of intelligent network and learning for the dimension reduction of STAP, are proposed. They implement the key technical requirements in the core problem of the clutter suppression, overcome the defects of traditional methods, and reduce the sensitivity of clutter suppression to complex clutter environment and harsh conditions. This proposal aims to lay a foundation for the breakthroughs of the related theory and the key technology in clutter suppression and target detection of airborne radar systems, under the circumstance of the insufficient of training samples.

杂波抑制是机载雷达信号处理的核心问题，空时自适应处理(STAP)是机载雷达杂波抑制与运动目标检测的核心技术，但其在实际应用中存在训练样本不足问题。本项目针对STAP应用中面临的核心难题——训练样本不足STAP难以收敛，开展新方向研究，探索机载雷达智能杂波抑制架构设计原理。研究包括：提出训练样本不足情况下的STAP智能网络与学习系统设计方法，对STAP智能系统的嵌入函数、网络输入输出和距离补偿约束—损失函数进行设计，使STAP的未知距离补偿矩阵参与系统训练与调整，突破STAP实际应用训练样本不足的核心难题；提出非独立同分布样本剔除的智能聚类分析设计方法、智能网络与学习的STAP降维设计方法，实现杂波抑制核心问题中的关键技术需求，克服传统方法缺陷，降低杂波抑制对现实复杂杂波环境与苛刻条件的敏感性。本项目旨在为训练样本不足条件下机载雷达系统杂波抑制和目标检测相关理论与关键技术的突破奠定基础。

杂波抑制是机载雷达信号处理核心问题，空时自适应处理（STAP）是机载雷达杂波抑制与动目标检测核心技术，但其实际应用存在训练样本不足问题。本项目针对样本不足STAP难以收敛这一核心难题，开展新方向研究，探索机载雷达智能杂波抑制架构设计。本项目按原研究内容和目标计划充分完成了项目预期目标：I）提出了“样本不足条件下基于自动编码器神经网络的机载雷达杂波抑制方法”，针对研究内容，实现了样本不足下杂波抑制智能网络与学习系统设计，并落实了杂波抑制空时二维稀疏自编码智能网络设计新思路。所提方法比传统方法改善因子平均提高了10dB。在正侧和非正侧视阵雷达下都具明显优越的改善因子，大大减小了样本不足引起的杂波谱展宽，提高了自适应性、更好地同时适应非均匀杂波和样本不足等非理想样本条件；II）提出了“非独立同分布样本条件下机载雷达杂波抑制与目标检测的无监督近邻传播聚类智能化方法”，实现了研究内容中非独立同分布（IID）样本的智能聚类分析新思路设计，并落实了不同智能网络，亦具有降维后的合理运算量。所提方法比传统方法具有明显更高的检测概率和较大优势，计算时间合理，更好地克服杂波距离依赖，更好地适应非IID样本；III）提出了“样本不足条件下基于极限学习机的机载雷达杂波抑制方法”，实现了智能网络不同设计与降维。所提方法将杂波抑制所需IID样本数量降低到了NK/2，即降低到原1/4倍，性能较好，且将运算时间降到角度-多普勒补偿方法的3倍，降低了机器学习方法较大运算量；IV）提出了“基于极限学习机线性嵌入的机载雷达运动目标检测方法”，实现了降维及不同网络设计。其计算速度快，泛化性和鲁棒性较好，可更好适应样本数的降低。在样本不均衡、复杂、标注不足下，显著提高了动目标检测杂波抑制性能；V）为雷达构建了可用于其它智能方法应用的通用训练数据集。本项目为雷达智能化应用拓展更大可能，为实现复杂杂波环境和苛刻条件下雷达杂波抑制和检测能力的提高提供关键基础和技术支撑。