基于机器学习的超宽带穿墙雷达三维成像理论研究

In order to satisfy the practical application of through-the-wall imaging radar, the two key problems which are ambiguities of wall parameters and real-time imaging must be solved simultaneously. The two problems are hot point and difficulty at present. Our research group have realized real-time location of a single target under unknown wall parameters in two-dimensional situation, and want to solve complex and three-dimensional problems on this basis. In this program, a through-the-wall scenario is simulated using finite difference time domain(FDTD), and the propagation characteristic of ultra-wide band (UWB) pulse when it encounters the walls is studied. Thus the UWB signal suitable for the through-the-wall imaging is designed. The receiving signal is analysed, and the scattered signal of the target is extracted from it and the clutter signals are filtered. The scattered signal are raw data for the imaging algorithms. Then the problem of through-the-wall imaging is converted to a regression one. The relationship between the scattered signal of the target and the unknown parameters such as the location and the shape of the target, the wall parameters is established. The relationship is nonlinear and improperly posed. To determine the relationship quickly and efficiently, the training process is done and the approximate linear formulation is obtained based on the powerful ability of learning the machnine learing. Therefor the methods can invert the unknown parameters according to the scattered signal. Finally the two- dimensional and three- dimensional real-time location and imaging problems about the multiple targets under unknown wall parameters are realised.

穿墙雷达成像要满足实际应用的要求，就必须同时解决实时性和墙体参数（结构、厚度、电磁参数）未知这两个关键问题。这两个问题是目前研究的热点和难点。项目组已实现了二维的墙体参数未知时单目标实时定位问题，拟在此基础上解决复杂问题和三维问题。本项目利用时域有限差分法仿真模拟穿墙场景，研究超宽带脉冲信号遇到墙体时的传播特性，设计出适用于穿墙成像的超宽带信号。分析接收机的接收信号，从中滤除杂波信号提取目标的散射信号，为后面的成像算法提供数据。再把穿墙成像问题转换为一个回归问题，建立目标散射信号和未知参数（目标的位置、形状、墙体参数等）之间的关系，这个关系是非线性的、不适定的。为了快速、高效地确定这个关系，利用机器学习方法强大的学习能力，通过训练得到它的近似线性表示式，从而可以根据目标散射信号反演未知参数。最终实现二维问题和三维问题中墙体参数未知情况下的多目标的实时定位和成像。

穿墙雷达利用电磁波的传播特性和信号处理技术，对墙后目标进行定位和成像。它是一种新型的无损探测技术，被广泛应用于军事和民事领域。本项目主要围绕如何解决穿墙雷达问题中两个关键问题——实时性和墙体参数未知进行展开研究。主要研究内容及结果包括以下几个部分：.一、利用接收机接收到的信号，分析墙体参数和信号特征之间的关系，通过机器学习找到两者之间的关系，从而对未知墙体参数进行预测。通过训练数据和测试数据所处环境的不同，验证了机器学习的泛化性和鲁棒性。.二、在实现二维墙后单目标快速定位的基础上，实现三维情况下墙后单目标的快速定位。结果表明，在维数增加的情况下，精度几乎不受影响。.三、当墙后有多个目标存在时，目标之间的回波常会相互干扰，影响定位。因此在三维情况下，利用每个目标的回波信号具有双曲线特性这一特点，对回波信号进行处理，使得目标回波和目标之间实现一一对应关系，从而大大提高了多目标定位的精确性和准确性。.四、利用神经网络学习方法——极限学习机对墙后目标进行定位。因极限学习机能在极短的时间内实现训练和测试，从而实现墙后目标的实时定位。相比机器学习，极限学习机具有更快的速度和更高的预测精度。. 总体而言，机器学习和神经网络能实现三维问题中墙体参数未知情况下多目标的实时定位，对穿墙雷达的定位技术有很重要的指导意义。