Research On Signal Sorting Algorithm For Communication Reconnaissance

With the continuous development of the communication technology, increasingly intensive communication environment makes the wideband data intercepted by wideband communication reconnaissance receiver possibly include various kinds of signals with different characteristics, such as fixed-frequency narrowband communication signal, frequency hopping signal, burst signal, sweeping frequency signal, spread spectrum signal and many kinds of artificial and non-artificial interference signals. So many interwined signals make it increasingly difficult to monitor these interested communication signals. Therefore, researching on how to sort all kinds of communication signals in complex communication environment, eliminate interference and noise, find interested communication signals, estimate the associated characteristic parameters, and reduce the further treatment burden of the communication reconnaissance system, not only is a very challenging issue, but also has become one of the urgent and difficult tasks for current communication reconnaissance area.This dissertation is focused on the problem of sorting fixed-frequency conventional communication signal (fixed frequency signal) and frequency hopping signal. The main achievements of this dissertation are given as follows:Firstly, the hardware-efficient architecture of polyphase DFT filter bank is derivated and when only white Gaussian noise is poured into the filter bank, the effect of the designed prototype low-pass filter on the correlation of the filter bank output is discussed in detail. When noise variance is known, the false alarm probabilities are obtained for two different conditions (practical condition and ideal condition), respectively. When noise variance unknown, some classical algorithms, such as cell-averaging(CA), order statistics(OS) and foreword consecutive mean excision(FCME) algorithm, are firstly proposed to performing detection in frequency domain, and then the false alarm probabilities of the CA detector are discussed under the above two different conditions. Simulated results show that the detection performance under the above two different conditions are very close to each other, which means that the correlation of the filter bank output can be ignored in practical application. Moreover, the detection performance of the above three detectors are compared by simulation.For the problem of sorting fixed frequency signal, two sorting methods are proposed, which are based on single antenna and array antennas, respectively. The proposed methods include three main parts: narrowband signal detection, characteristic parameter estimation, and narrowband signal tracking. In the aspect of narrowband signal detection, to reduce the computation complexity of the original localization algorithm based on double thresholding (LAD), a modified LAD algorithm is proposed. To impove the detection performance, combining the above modified LAD algorithm and morphological processing together, an enhanced LAD algorithm is proposed. For characteristic parameter estimation, a single antenna based estimation method is firstly discussed. When array antennas are used, the idea of clustering analysis is introduced and a new measurement set segmentation based estimation method is proposed. When it comes to narrowband signal tracking, two different narrowband signal decision methods are discussed and a narrowband signal library update method is studied.According to the time-domain, frequency-domain and spatial-domain information used in sorting methods, three frequency hopping signal sorting methods are proposed, which are based on frequency-spatial-domain(FSD), time-frequency-domain(TFD) and time-frequency-spatial-domain(TFSD) information, respectively. The FSD based signal sorting method is very similar to the above fixed frequency signal sorting method when array antennas are used. The difference between the above two methods is only in case of narrowband signal tracking. In this method, a density estimation clustering based measurement set segmentation algorithm and a frequency hopping signal tracking method are studied. For the TFD based signal sorting method, a morphological processing based time-frequency detection and interference eliminate algorihm is firstly proposed, and then a characteristic parameter estimation method is discussed. Finally, a frequency hopping period estimation method is investigated in detail, which dilutes the characteristic parameter set, and uses cumulative difference histogram (CDIF) and center time (CT) transform algorithm to deinterleave these CT sequences included in each diluted subset and estimate frequency hopping period. To solve frequency collision problem, the spatial information is introduced and a TFSD based sorting method is developed. Compared with TFD based method, the TFSD based sorting method only has small difference in the aspect of interference eliminate and characteristic parameter estimation. The last method need to segment the corresponding measurement set of each connected region in TFD, and then estimate characteristic parameters. In this method, a morphological watershed based clustering algorithm is studied.The problems of solving frequency collision and characteristic parameter set dilution can be regarded as processing measurement set segmentation problem. Hence, a Gaussian mixture model based measurement set segmentation algorithm is proposed. Firstly, the conception of Gaussian mixture model and standardexpectation-maximization (EM) algorithm is simply introduced. A new competitive stop EM (CSEM) algorithm is proposed, which is on the basis of normality test based EM algorithm. The proposed algorithm does not know the model order in prior, is robust to the initialized parameters of model, and suitable for compact and hyperellipsoidal sets. At last, simulation results vertify the validity of the proposed CSEM algorithm; its application in measurement set segmentation is also discussed by simulation. To improve the robustness to outliers and the shape of characteristic parameter set, two new support vector clustering(SVC) based measurement set segmentation algorithm are developed. The first algorithm is a kernel parameter estimation based multi-sphere support vector clustering (KPMSVC) algorithm, which is proposed after carefully analyzing the effect of spread factor and penality factor on clustering results. Because of the computation complexity of cluster assignment for large sample set, the proposed KPMSVC algorithm is only suitable for small sample set. In order to apply SVC for large sample set, a statistical histogram based multi-sphere support vector clustering (SHMSVC) algorithm is proposed. Simulation results vertify the validity of the above two algorithms. The above two algorithms are also applied to solve frequency collision and characteristic parameter set dilution problems.