多重分形符号相位转移熵理论及其在海面漂浮小目标检测中的应用

Under the condition of low signal-to-noise ratio, the detection of weak radar targets on the sea surface faces big challenges and is a hot and difficult point of present researches. This project studies the radar weak target detection problem based on nonlinear random fractal, symbolic phase transfer entropy and deep learning. A new feature extraction, feature compression and classification algorithm based on random multifractal and symbolic phase transfer entropy is proposed to solve the weak target detection problem under complex background. The researches include: ①Two-dimensional cross-correlation singularity power spectrum (2D-CSPS) theory and algorithm are proposed,and based on this, the SAR ship target detection method in the background of sea clutter is proposed.②The concept of multifractal symbolic phase transfer entropy (MF-SPTE) analysis is proposed.The MF-SPTE estimation algorithm and feature extraction method of complex scene echo are constructed.③The detection method of sea-surface small floating target based on deep learning and MF-SPTE is studied. Based on MF-SPTE and multiscale symbolic phase transfer entropy algorithm, extract the combined feature vector of sea clutters, combine with improved AE deep learning method and "dropout" technology to obtain deep features to improve the ability of target detection of sea-surface small floating target. The research results will enrich the theory of singularity spectrum analysis and provide theoretical and algorithmic basis for weak target detection.

低信杂比条件下海面漂浮弱小雷达目标检测面临巨大的挑战，是当前研究的热点和难点。本项目研究基于非线性随机分形、符号相位转移熵和深度学习的雷达弱小目标检测问题，提出新的基于随机多重分形和符号相位转移熵的特征提取、特征压缩和分类算法，拟解决复杂背景下的微弱目标检测问题。研究包括：①提出二维互奇异性功率谱（2D-CSPS）理论和算法，并基于此提出海杂波背景下SAR舰船目标检测方法；②提出多重分形符号相位转移熵（MF-SPTE）的概念和理论，构造MF-SPTE估计算法和复杂场景回波特征提取方法；③研究基于深度学习和MF-SPTE的海面漂浮小目标的检测方法，基于MF-SPTE和多尺度符号相位转移熵算法，提取海杂波信号序列的组合特征向量，结合改进的AE深度学习方法和“dropout”技术获得深度特征，提高海面漂浮小目标的检测能力。研究成果将丰富奇异性谱分析理论、为弱小目标检测提供理论和算法基础。

本项目系统研究了多重分形算法、样本熵、多尺度样本熵、排列熵、多尺度排列熵、权重排列熵、转移熵、相位转移熵的算法，并将其应用在生物医学信号、轴承信号、海杂波信号的特征分析与分类中，取得较好效果。当前，目标检测较好的分析方法，就是融合多个特征(基于融合时域和频域多特征)的海上小目标检测方法，在一定程度上提高了检测精度，但是多特征融合的方法要求各统计特征之间具有好的互补特性，在低信杂比(Signal Clutter Ratio, SCR)和短时观测积累情况下，这类方法也难以达到理想效果，而且使用固定的人工选择的统计特征难以使算法具有好的泛化能力。为了能够发挥深度学习在图像分类领域的优势，本项目通过将雷达回波多普勒谱序列编码为图像，把时间序目标检测问题转化为图像分类问题，并基于改进的EfficientNet网络模型提出了一种新的海杂波背景下的漂浮小目标检测方法。为了找到多普勒谱序列最佳的图像编码方法，本项目探索了2种不同的编码方式: 格拉姆角求和场(Gramian Angular Summation Field, GASF)和格拉姆角求差场(Gramian Angular Difference Field, GADF)。为了强调编码图中主要位置纹理信息的重要性，本研究尝试向EfficientNet这一轻量级网络模型中MBConv模块中添加坐标注意力(Coordinate Attention, CA)机制，并采用自适应AdamW优化算法来优化模型收敛。最后，通过在GADF和GASF数据集上进行验证，表明了所提算法不仅具有较高的识别准确率，还具有较少的模型参数量，在轻量化计算的基础上所提算法具有明显优于同类最优算法的准确率。