類神經模糊網路於信號測向度估測及鑑別之應用

Abstract

本論文主要目的在針對(1)信號測向度估測及(2)信號鑑別的問題，加以研究分析，分別提出有效解決方法，並經由環境模擬之仿真範例進行性能驗證及評估，結果發現本文所建構網路及演算法均優於其他方法。 在陣列信號處理領域中，信號測向度估測問題非常重要。許多傳統式估測方法，諸如：「多信號分析法」(MUSIC)與「最大可能性估測法」(MLE) ，因而被提出。二種估測法中，「最大可能性估測法」其估測性能雖佳，但由於估測過程中需處理多變數非線性極大值問題，致使運算量過於龐大，並不普遍被使用。相較下，當信雜比及抽樣數量不小的條件下，此時只要處理一維極大值問題及尋找信號或雜信子空間的「多信號分析法」就顯得較受歡迎。雖然如此，此方法於進行特徵分解運算時，仍存在高運算量問題。總之，二種估測法具有高運算量及無法即時性處理的缺點。為改善此缺點，本論文提出不同觀點，亦即將測向度估測的問題視為映射問題加以分析處理。首先提出一個具有線上學習兼具高抗雜訊能力之六層式類神經模糊網路 (NFN) 架構，此網路架構結合專家知識 (模糊法則) 及類神經網路學習能力特色。初始時，網路本身並無法則存在。法則的產生與調整乃是由線上同時進行的架構與參數學習完成。就架構學習而言，網路的前件部是根據正型分群法 (aligned clustering-based) 對輸入空間進行自動線上彈性分割。後件部的學習，初始是依據分群法給予每一條法則的單值。前件部與後件部的學習可產生一個有效率、動態自我增長的網路。以此架構為基楚，推導出網路相關參數學習法則。前件部參數由倒傳遞演算法調整，後件部參數則採用最小平方法 (LMS) 或遞迴式演算法 (RLS) 調整。架構與參數學習同時進行的結果，使本網路具快速的學習能力。故在學習過程中，僅需由外界給予適當訓練資料即可迅速完成架構及參數學習，而不需要任何事先的知識 (法則) 。最後，為提高系統估測精確度，網路輸入資料取自干涉儀的相位差輸出，如此，即可實現一個智慧型最佳化類神經模糊 (neuro-fuzzy) 估測系統，以改善傳統式估測器低精確度及高運算量缺點，並解決即時性信號測向度估測問題，最後再與向徑式基底函數網路 (radial basis function network) 就估測精確度、調整參數量及抗雜訊靈敏度項目作一比較。 此外，在信號鑑別方面，仍將運用上述問題處理模式，將信號鑑別視為一種映射問題而加以分析處理，於是建構一種三層式向量類神經網路 (VNN)，網路輸入變數則取自信號源重要特性參數，諸如：射頻 (radio frequency)，脈波寬 (pulse width) 及脈波來復時間 (pulse repetition interval) 等之上下限值，基於監督式學習，文中將介紹傳統向量倒傳遞式(CVTBP)及新型向量倒傳遞式(NVTBP)二種演算法，兩者演算法分別就不同定義之誤差函數 (error function) 所產生，並推導出網路參數學習法則，以進行網路參數的調整，並尋求系統中最佳化連結權值，達到高鑑別率目的。文中提出多種範例，針對二者演算法進行模擬驗證及比較，結果顯示新型向量倒傳遞式演算法，在收斂速度及鑑別準確性上，均優於倒傳遞演算法，並證明所建構之向量類神經網路確實可處理信號鑑別的問題。為因應未來信號複雜化，可考慮增加信號參數做為網路輸入變數以進行信號鑑別。 以上所提出的類神經模糊網路或向量類神經網路，均由仿真信號 (模擬無雜訊或有雜訊干擾環境下測試信號) 範例，進行電腦模擬驗證，並與其它方式作比較，結果均驗證出所提方法，無論在信號測向度估測精確度或信號鑑別準確性上，皆具有高估測能力及高鑑別率優點﹔因此，在未來研究中，所建構網路應可廣泛應用於信號偵測、定位及追蹤上。

From engineering point of view, this thesis aims to provide a powerful and effective methodology for direction of arrival (DOA) estimation and emitter identification (EID) in electronic warfare (EW) applications, respectively. Capabilities and performances of the proposed scheme have been verified and evaluated with other methods by various examples. Simulation results show that the proposed networks with associated algorithms are superior to other methods. Estimating the DOA of signals is a significant problem in the field of array signal processing. Many conventional DOA estimation methods have been proposed, including the multiple signal classification (MUSIC) method of Schmidt, and the maximum likelihood (ML) technique. Of both these available methods, the ML technique has the best performance. Nonetheless, because of the high computational load of the multivariate nonlinear maximization problem involved, the ML technique did not become popular. On the contrary, the suboptimal MUSIC method is more prevalent than the ML technique when the signal-to-noise ratio and the number of samples are both not too small, because the suboptimal method involves solving only a one-dimensional maximization problem and finding a subspace (signal subspace or noise subspace). However, the MUSIC has to perform the eigen-decomposition. Hence the major computational burden lies in finding the signal subspace or noise subspace. In summary, these methods are computationally intensive and difficult to implement in real time. The DOA estimation is viewed as a mapping problem from a different view of point in this thesis. Therefore, we propose one scheme to cope with the mapping problems. At first, we propose a six-layered neural fuzzy network (NFN) with on-line learning and anti-noise abilities. The network structure keeps integrating expert knowledge (or fuzzy rules) and neural network's learning abilities. There are no rules initially in the NFN. They are created and adapted as on-line learning proceeds through simultaneous structure and parameter learning. In the structure learning of the precondition part, the input space is partitioned in a flexible way according to an aligned clustering-based algorithm. As to the structure learning of consequent part, only a singleton value selected by a clustering method is assigned to each rule initially. The combined precondition and consequent structure learning scheme can set up an effective and dynamically growing network. Based on the constructing network, the associated parameter learning algorithms are derived. The precondition parameters are tuned by the backpropagation algorithm and the consequent parameters are tuned optimally by either least mean squares (LMS) or recursive least squares (RLS) algorithms. Both structure and parameter learning are done simultaneously to form a fast learning scheme. More notably, only the proper training data need to be provided from the outside world to achieve the structure and parameter learning without giving any initial rule in this learning method. To increase the estimation accuracy, we take the phase differences (PD) from the output of the interferometer as network input. Therefore, we can realize an optimal neuro-fuzzy estimation system to deal with the defects of low accuracy, high computational burden, and real-time estimation for the conventional estimation methods. Finally, the performances comparison of both the NFN and the RBFN in terms of convergence accuracy, estimation accuracy, sensitivity to noise, and network size are performed by various examples. In addition, an optimal identifier is proposed for tackling the emitter identification problems. The associated learning algorithms is also derived in this thesis. Similar to the above processing method, we reconsider the problem of emitter identification as a mapping problem again. Hence, we construct a three-layered network so-called vector neural network (VNN) to identify emitter types. Each of the input variables is selected by the feature vector which is composed of the upper limit and lower limit values of the characteristic parameters in detecting signal. The input feature vectors include the radio frequency (RF), pulse width (PW), and pulse repetition interval(PRI). Based on the supervised learning algorithm, both conventional vector-type backpropagation (CVTBP) and new vector-type backpropagation (NVTBP) algorithms are introduced and used to derive the parameter learning algorithms. The parameter learning algorithms are also used to tune the network parameters, find the optimal connection weights and increase the correction rate of emitter identification. Capabilities and performances of both algorithms (CVTBP and NVTBP) are verified and compared by various computer simulations. Simulation results show that the NVTBP outperforms the CVTBP not only in terms of convergence rate but also in terms of correction rate. It is seen that the constructing VNN is feasible to solve the problem of the EID. For the complex signals, we may add signal parameters as input variables to keep high discrimination rate in the future. Above constructing networks including the NFN and VNN are verified by emulating examples with/without noise conditions through computer simulations. The proposed networks are also compared to other methods by the same examples. In summary, each one of the aforementioned networks (either NFN or VNN) has high accuracy/high correction rate in the DOA estimation and in the EID. With these features, we believe that our proposed networks may be applied for solving the problems of the multiple signals detection, signals localization, signals tracking, and beamforming in the future research.