飛彈防衛系統之任務分配及軌跡規劃

Abstract

針對非線性系統的控制問題，本論文發展兩個嶄新的控制架構。首先探討多自主系統(MAS, multi-agent system)，其為一包含多個自主運作的群體系統，能接受來自其他機構或者中央的命令來採取行動，並在我們所設定的條件內達成任務。飛彈防衛系統(MDS, missile defense system)即為一多自主系統的延伸架構，環境內有來自敵方的多個攻擊飛彈的威脅、有限多個據點將遭受戰損與有限多個防衛飛彈發射並各自攔截攻擊飛彈等，均是本論文第一部份的探討重點。在飛彈的任務規劃部份，傳統的任務分配使用窮舉法來做多對多的配對，雖然最終能找到滿足條件的最佳配對，但是當對象數量增加時所花費的計算量更是龐大，因此本論文提出自組織映射(SOM, self-organizing map)，其優勢在於降低配對的計算量並可根據多個據點的最小總戰損為目標來進行多個防衛飛彈與多個攻擊飛彈之間的配對，如此不僅可以加速配對時間也可以滿足所設定的最小總戰損條件。在控制器的部份，本論文根據飛彈導引法則為基礎建構提出智慧型模糊類神經網路(FNN, fuzzy neural network)控制器架構，相較於CMAC(cerebellar model articulation controller)在飛彈導引的運算時間過長及失誤距離過大，更進一步改善飛彈導引系統的即時性；本論文所提出的控制器架構可經由李亞普諾夫穩定性證明來保證系統的穩定性，同時藉由參數學習的機制來克服系統的非線性。最後由模擬成果可明顯看出自組織映射在任務分配與模糊類神經網路控制器在飛彈導引控制的結合度極高，因此可以完整建構出一飛彈防衛系統。本論文之第二部份為高階霍普菲爾神經網路(HOHNN, high-order Hopfield-base neural network)應用於動態系統之鑑別。高階霍普菲爾神經網路中的函數型連結網路(FLN, functional link net)能提供額外的輸入給網路之各神經元。相對於傳統的霍普菲爾網路(HNN, Hopfield neural network)，本論文所提出的函數型連結網路具有系統化階次的數學表示法，具有較快的收斂速度及較少的計算負載。另外，函數型連結網路之權重更新，亦可藉由李亞普諾夫穩定理論來保證在非線性即時系統的鑑別收斂。針對各種基於霍普菲爾神經網路架構的比較，可由模擬結果及計算量分析顯示我們所提出的高階霍普菲爾神經網路在未知動態系統的鑑別具有較高的效能。

In this dissertation, two novel control schemes are proposed to solve the control problems of nonlinear systems. The first is the multi-agent system (MAS) consists of multiple autonomous systems which can activate, interact, and communicate with each others or from central command, and eventually complete some missions under the desired conditions. Missile defense system (MDS) is a suitable application of MAS: threat from multiple attacking missiles, some limited assets are under attack, and multiple defense missiles (or agents) are launched to intercept the associated attacking missiles (or targets), and a fuzzy neural network (FNN) controller with self-organizing map (SOM) for MAS are investigated in the first part of this thesis. The presented approaches are better than traditional exhausted method which can find the optimal solution though time-consuming when the processing data increases. The advantage of SOM is the less computational load under the condition of minimal total asset damages. Therefore, SOM can be adopted to not only dispatch the agents toward the targets, but also lower the computational load under the desired condition. Based on the missile guidance law, the proposed FNN can deal with the problems of large computational load and miss distance by the cerebellar model articulation controller (CMAC). Finally, the proposed SOM-based FNN controller adopted in the highly nonlinear MDS can be guaranteed stable and the parameters can be updated via Lyapunov stability criterion. From the experimental results, it can be demonstrated the possibility of applying the proposed intelligent control method in MDS. In the second part of the thesis, the high-order Hopfield-based neural network (HOHNN) is proposed to the dynamical system identification. The functional link net (FLN) in HOHNN has additional inputs for each neuron. In comparison with the traditional Hopfield neural network (HNN), the compact structure of FLN with a systematic order mathematical representation combined into the proposed HOHNN has additional inputs for each neuron for faster convergence rate and less computational load. In addition, the weighting factors in HOHNN are tuned via the Lyapunov stability theorem to guarantee the convergence performance of real-time system identification. The simulation results and computation analysis for different Hopfield-based neural networks are conducted to show the effectiveness of HOHNN in uncertain dynamical system identification.