基于非监督贝叶斯学习雷达性能指标动态评估

Dynamic evaluation of radar performance index based on unsupervised Bayesian learning

针对传统雷达性能指标评估方法相对“机械”、缺乏理论约束,需要多次重复实验,导致评估效率较低,评估成本较高等问题,提出基于非监督贝叶斯学习方法的雷达性能指标动态评估算法,在一定雷达探测目标先验假设下,结合典型回波观测数据模型,建立雷达性能指标后验概率模型。考虑到先验知识与观测数据可能存在的非共轭特性,针对先验概率模型建立分层贝叶斯模型,从而保证雷达性能指标后验概率密度函数的可解性。此外,为了保证后验概率密度函数的闭合解析解,应用变分贝叶斯期望最大化(variational Bayesian expectation maximization,VB-EM)方法,基于高斯赛德尔迭代策略分别计算性能指标及其超参数的后验概率密度函数。最终,利用后验概率密度函数计算结果,可获得相应性能指标解析估计值及其置信区间和置信度,从而实现对指标动态变化的解析指示。相比传统蒙特卡罗评估方法,所提方法仅需一次实验数据便可获得定量的、解析的指标评估结果,可以大大缩减评估成本,提升评估效率,同时可对指标动态变化给出定量指示。应用仿真数据对雷达定位、测高精度以及目标检测概率指标进行了验证,相比传统方法,评估处理增益获得了有效提升。

In view of conventional evaluation method of radar performance index is relatively tedious and lack of theoretical bounds,which requires enough repeated experiments.This results in low efficiency and high cost of implementation.A fully dynamic evaluation algorithm for the radar performance index is proposed based on the unsupervised Bayesian statistical learning method.Combined with typical observed data model,the posterior probability model of radar performance index is established under a certain prior assumption of radar target.Considering the possible non-conjugated characteristic of prior and observation data,a hierarchical Bayesian model for the prior probability model is introduced,which guarantees the solvability of the intended posterior probability function for the radar performance index.In addition,to ensure closed form solutions of the posterior probability function,the method of variational Bayesian expectation maximization(VB-EM)is utilized,so that the posterior probability function of the performance index and the hyper-parameters can be calculated through Gauss-Seidel iterative strategy.Finlly,the analytical estimates of corresponding performance indexs,and their confidential interval and confidential degree can be acquired by using the results of posterior probability function,which can realize the analytical indication of the dynamic change of index.Compared with the traditional Monte Carlo evaluations,the proposed method can obtain quantitative and analytical index evaluation results with single piece of experimental data,which can greatly reduce the evaluation cost and improve the evaluation efficiency.At the same time,it can give a quantitative indication of the dynamic changes of the index.The simulation data is applied to verify the accuracy of radar positioning and height measurement as well as target detection probability index.Compared with the traditional methods,the gain of the evaluation can be improved in a great extent.