類神經模糊網路應用於雷達脈波信號壓縮處理及機動目標追蹤之研究

Abstract

摘 要 本論文最主要之目的是要以類神經模糊網路的方法，應用於雷達信號處理器及目標追蹤器上，以解決當今雷達系統在執行目標偵測及進行目標追蹤時，所遭遇到的兩大難題，亦即長的雷達發射脈波寬和高解析度之間兩難之困境及追蹤目標具機動運動(maneuvering)時所導致之追蹤漏失(miss-tracking)問題。經由對問題的研究與分析，將類神經模糊網路分別設計並結合在雷達信號處理器及目標追蹤器上，建立智慧型現代化雷達系統之新架構，並考量所遭遇之不同場景及環境，模擬各種仿真範例，進行性能驗證、評估及比較，結果顯示本文所建構的方法在處理上述兩大問題上，除了優於其他方法外，更能達到現代化雷達高性能之需求。 無論是民生或國防上之應用，雷達為遠距離目標偵測、追蹤及其資訊獲取最佳工具之一。雷達要能獲取正確的目標資訊，須具備良好之目標追蹤能力，而目標追蹤品質之良窳，決定在雷達信號處理器是否提供良好之目標信號，因此目標之偵測與追蹤能力，事實上決定了現代雷達的性能。然而雷達信號處理器在偵測目標時存在有兩難之困境，那就是以較短之雷達發射脈波發射，雖具有較高之距離解析度，但也因較短之雷達發射波，而具較低之發射功率，所以只有較近之有效偵測距離；反之，以較長之雷達發射脈波發射，雖有較遠之有效偵測距離，相對地卻也因較長之發射脈波而有較低之距離解析度。在目標追蹤之問題上，精確之目標追蹤以獲取正確之目標資訊，始終是目標追蹤系統設計最主要目的。對於一個等速運動之目標，一般的傳統追蹤技術，諸如 濾波器及標準卡爾曼濾波器(Kalman filter)即足以精確的追蹤，然而對於突然具有加速度運動之機動目標，其何時開始機動及機動發生時之加速度值是多少，一般之追蹤技術皆無法測知進而加以補償，也因此常使目標無法繼續鎖定追蹤而造成脫鎖，此為雷達目標追蹤系統最重要且困難的問題。因此若能建構出解決此兩大難題之方法，則現代雷達系統之性能將大幅的提昇，這是本論文主要的研究動機。 在高解析度雷達中要克服長的雷達發射脈波和高距離解析度之間兩難之困境，必須以脈波壓縮(pulse compression)來處理，其主要之功能是提高信號對旁波比值(signal-to-sidelobe ratio) 及降低總體旁波程度(integrated sidelobe level, ISL)，以提昇雷達之距離解析度，並增加雷達之分辨及偵測能力。為要實現雷達脈波壓縮之功能，使用13位元巴可碼(13-element Barker code)之傳統技術，諸如直接自我關聯濾波器(autocorrelation filter, ACF)、最小平方反向濾波器(least squares inverse filter, LS)及線性規劃濾波器(linear programming filter, LP)已經被開發使用；近來類神經網路之技術亦被提出。然而傳統技術皆未能達到現代雷達之高信號對旁波比值的需求，而一般之類神經網路如倒傳遞(backpropagation, BP)網路，則又產生低收斂速度及易受都卜勒頻移(Doppler frequency shift)所影響等問題。為克服上述之缺點，本論文利用神經網路之學習能力提出一種新的方法，以類神經模糊網路處理雷達系統中脈波壓縮。在文中，一種具監督式學習技術之六層自我建構類神經模糊推理網路(self-constructing neural fuzzy inference network, SONFIN)將用以分別結合13位元之巴可碼及20位元之組合巴可碼(combined Barker code, CBC)，模擬結果顯示此種類神經模糊網路脈波壓縮(neural fuzzy network pulse compression, NFNPC)技術在抗干擾特性、距離解析能力及都卜勒容忍度(Doppler tolerance)上，具有較傳統及BP方法為佳之優點，同時比BP方法有更快之收斂速度，因此非常適用於高解析度雷達系統。 雷達在追蹤機動目標時，所存在的重要且困難問題：如何偵測目標之機動資訊以便能快速地反應進而避免追蹤漏失？傳統之目標機動偵測技術，諸如變維濾波器(variable dimension filter, VDF)及輸入估測(input estimation, IE)等皆具有大計算量及即時實現上有困難之缺點；近來類神經網路之技術亦被提出，然而一般之類神經網路，如BP網路又產生另外之低收斂速度及網路過於龐大等問題。為克服上述之缺點，本論文研究發展卡爾曼濾波器結合SONFIN，稱之為KF-SONFIN技術來實現雷達追蹤器。藉由產生具有機動資訊之目標可能軌跡，用以訓練SONFIN，使完成訓練之SONFIN具有測知目標之機動何時發生、機動發生時之加速度值及機動何時結束之能力，由SONFIN所獲取目標之加速度值，便可即時且直接的輸入至卡爾曼濾波器以為補償。因此，此一新的追蹤技術除有不須改變卡爾曼濾波器之基本結構，亦具不須求取機動目標模式的有利條件，且SONFIN本身又具自我建構為最經濟有效之網路兼具快速學習之優點，而由模擬實驗之結果顯示，更具有比VDF及IE等方法較佳之追蹤精度。 總之，本論文最重要的貢獻是：(1)於信號處理器中提出NFNPC技術來解決目標偵測時之困境，(2)於目標追蹤器中提出KF-SONFIN以解決機動目標追蹤問題，由模擬實驗得到比既有方法有較佳之結果，完全克服現代高性能雷達所面臨的兩大難題。

ABSTRACT The main aim of this thesis is to apply neural fuzzy networks to solve two significant and difficult problems existing in the radar signal processor while detecting a target, and the target tracker while tracking a target in the modern radar system, respectively. The first problem is the dilemma between long radiated pulse width and high resolution. The second is the miss-tracking problem occurring in target tracking. After exploring and analyzing the problems, neural fuzzy networks are designed and combined with a radar signal processor and a target tracker to solve the problem by constructing an intelligent modern radar system. Various scenarios and environments are used in simulation examples. The performances of the proposed schemes are verified, evaluated, and compared with other methods. Simulation results show that the proposed schemes are superior to existing methods in solving the above two problems, and they meet the high performance requirements of a modern radar system. In both civil and military applications, radar is one of the best instruments to detect a long-range target, track the target and then acquire its information. To acquire accurate target information, radar systems must have an excellent target tracking ability. Moreover, to get excellent tracking quality, a radar signal processor must also provide a high quality signal. However, there exists a dilemma for the radar signal processor while detecting a target. Although the radar has higher range resolution when shorter pulses are radiated, these shorter radiated pulses have lower radiated power, which leads to a shorter effective detection range. Conversely, a radar has longer detection ranges when longer pulses are radiated; a lower resolution also occurs because longer pulses are radiated. In the problem of target tracking, it is always the main purpose of a the target tracking system to track the target precisely for acquiring accurate target information. The conventional algorithms, such as filter and standard Kalman filter, can track a constant speed target adequately. However, they cannot detect the time when an abrupt maneuver occurs and the acceleration value of a maneuvering target to be compensated thereafter. This is the main reason for failing to keep tracking a target continuously and thus leading to loss of the target eventually when using the conventional algorithms. This is the most significant and difficult problem for a radar target tracking system. Therefore, if there exists an approach to solve these two difficult problems, the performances of the radar are improved substantially. This is the main motivation of this paper. Pulse compression is used to overcome the dilemma between a long radiated pulse width and high range resolution. By increasing the signal-to-sidelobe ratio and decreasing the integrated sidelobe level (ISL), pulse compression can not only increase the mainlobe magnitude and decrease its width but also decrease the sidelobe magnitude, which results in enhancing the range resolution and thus the abilities of target discrimination and detection. To achieve this, the traditional algorithms such as direct autocorrelation filter (ACF), least squares (LS) inverse filter, and linear programming (LP) filter based on 13-element Barker code have been developed. Recently, the neural network algorithms were issued. However, the traditional algorithms cannot meet the requirement of high signal-to-sidelobe ratio for modern radar system and the normal neural networks such as backpropagation (BP) network usually produce the extra problems of low convergence speed and sensitivity to the Doppler frequency shift. To overcome these defects and to make use of neural learning ability, a new approach using a neural fuzzy network to deal with pulse compression in a radar system is presented in this paper. The 13-element Barker code and 20-element combined Barker code (CBC) used as the signal codes are carried out by a six-layer self-constructing neural fuzzy network (SONFIN) with a supervised learning algorithm. Simulation results show that this neural fuzzy network pulse compression (NFNPC) algorithm has significant advantages in noise rejection performance, range resolution ability, and Doppler tolerance, which are superior to the traditional and BP algorithms. It also has faster convergence speed than the BP algorithm, so it is suitable for high-resolution radar systems. In radar target tracking systems, there exists an important and difficult problem while a maneuvering target is being tracked: how to detect target maneuvers and fast response to avoid miss-tracking? The traditional maneuver detection algorithms, such as variable dimension filter (VDF) and input estimation (IE), are computation intensive and difficult to implement in the real time. To solve this problem, neural network algorithms have been proposed recently. However, conventional neural networks such as backpropagation networks usually produce the problems of low convergence speed and large network size. To overcome these defects and to make use of neural learning ability, a developed standard Kalman filter with a self-constructing neural fuzzy inference network (KF-SONFIN) algorithm for target tracking is presented in this paper. By generating possible target trajectories including maneuver information to train the SONFIN, the trained SONFIN can detect when the maneuver occurred, the magnitude of maneuver values, and when the maneuver disappeared. Neither having to change the structure of Kalman filter nor modeling the maneuvering target, this new algorithm, SONFIN, also has the advantage of finding itself an economic network size with a fast learning process. Simulation results show that the KF-SONFIN is superior to the traditional IE and VDF methods in estimation accuracy. In short, the contributions in this thesis are (1) the NFNPC algorithm is proposed to solve the dilemma occurred in target detection, and (2) the KF-SONFIN algorithm is proposed to solve a maneuvering target tracking problem. The simulation results are superior to the existing algorithms. The significant and difficult problems encountered in modern system are overcome completely.