【作者】 李谷芳；

【导师】 宗成阁；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2007， 硕士

【摘要】 常规防空雷达受其固有低重复频率等性能的影响,将其应用于飞机目标自动分类与识别的问题一直是目标识别研究领域的难点。目前,国内装备有大量的常规防空雷达,如果能在常规防空雷达目标识别上取得突破,将具有重大的军事与经济意义。如何准确实时地提取飞机目标的有效特征并对多传感器信息准确融合是解决目标自动分类与识别问题的关键技术。本文的主要工作是研究了小波分析提取飞机目标发动机周期调制特征的方法以及构造实现飞机目标分类识别的智能化类型融合模型和算法。由于飞机旋转部件对雷达电磁波的调制,目标回波会产生周期调制特征。常规防空雷达可以提取周期调制特征来分类、识别。由于传统提取方法如复倒谱法、无偏自相关估计、AR功率谱法等或者要求雷达有较高重复频率或者存在提取误差大、计算量大等缺点。本文通过对直升机、螺旋桨飞机和涡扇喷气飞机回波调制特性参数模型的研究,分析了这三类飞机所产生的回波信号特点。提出采用小波分析方法有效获取目标回波的周期调制特征,并详细分析了小波基的选取和几个重要雷达参数对其提取结果的影响并给出相应的仿真试验结论。在多传感器信息融合方面,本文采用了基于多级神经网络的融合模型。该模型分为传感器子网和融合子网。传感器子网通过多个传感器获得的目标特征信息,实现对各类目标类型的置信度分配。融合子网将结合各传感器子网的置信度对各传感器子网的输出结果进行融合,最终得到对目标类型的判断。本文重点研究了该多级神经网络融合模型中传感器子网的模型和算法。采用了一种基于专家规则的模糊神经网络,网络结构和各个节点都有确切的含义。并改进其算法——将改进的遗传算法与传统后向传播算法结合得到性能更优的GA-BP分类算法并应用该算法对网络进行了训练和检测。更多还原

【Abstract】 Because of the influences of some inherent performances, such as low repeat frequency, it is always very difficult to apply convention air defense radars to airplane goal automatic sorting and recognition in the target identification research area. At present, massive convention air defense radars are equipped in domestic. Obtaining the breakthrough in the convention air defense radar’s target identification will have momentous military and economical significance. Extracting valid airplane goal characteristics exactly and rapidly from the echo and multi-sensor information fusion are two pivotal techniques. The prime tasks of this paper are airplane periodic signature of modulation extraction with wavelet decomposition as well as structuring intellectualized type fusion model and algorithm of target identification.Because the airplane’s revolving part can have cyclical modulation feature to the radar electromagnetic wave modulation, convention air defense radar could extract periodic signature of modulation to distinguish airplanes of different types effectively. There are some disadvantages in conventional extract methods, such as differential cepstrum analysis, estimate of unbiased autocorrelation, AR power spectrum and so on. Some need high radar repeat frequency, some have biggish error and some have to do a mass of computation. According to the research of modulating feature parameter model of helicopter, propeller-driven aircraft and turbofan jet airplanes’echo, in this paper, we analyze the echo signal feature and put forward the algorithm of extracting target echo’s periodic signature of modulation by wavelet analysis. In this paper, we also discuss the question of how to choose appropriate wavelet function and the influence of several radar parameters to the result of extract in detail and give corresponding results and conclusions of simulations.In the aspect of multi-sensor information fusion, a multistage neural network fusion system based on neural network, fuzzy reasoning and expert system will be used. The neural network includes sensor subnet and fusion subnet. The sensor subnet obtains target feature informations through many sensors and gives the likelihood of every target type. The fusion subnet unifies each sensor’s confidence level to fulfill fusion task with the output results of sensor subnet and finally distinguishes the type of target. This paper puts emphasis on the model and algorithm of sensor subnet of the multistage neural network fusion system. The sensor subnet is one kind of fuzzy neural network based on expert rules. The architecture and each node of the network all have accurate meanings. Also in this paper, the algorithm of the sensor subnet will be optimized, the genetic algorithm optimized and back propagation algorithm will be fused to obtain a new algorithm named GA-BP algorithm which has more excellent performances. At last the sensor subnet will be trained and tested with this algorithm.更多还原