本論文主要是研究與分析結合基因演算法，灰色理論、模糊理論與類神經網路，並將其應用於全車之主動式懸吊系統與倒單擺系統。首先針對GM(1,1)模型提出一些定理，這些都是根據發展係數 與灰色輸入 兩個參數，來簡化求解的型式。接著以基因演算法來對此調適因子進行最佳化之搜尋，來改善預測值之殘差。 第二部分利用灰模糊控制器與類神經網路模糊控制器於全車之主動式懸吊系統設計。我們主要的控制目標是強調去改善車身振盪，經模擬結果證明，類神經網路模糊控制之主動式懸吊系統比傳統的被動式懸吊系統、模糊控制之主動式懸吊系統、及灰模糊控制之主動式懸吊系統更有效降低車身振動。同時吾人並運用灰色關聯度進行車身於不同路況的運動分析，比較各種震盪對地形變化的關聯度，所以灰色關聯度方法也不失為是一種分析工具的選擇。最後吾人結合類神經網路具有學習之優點與灰關聯分析及模糊控制器，以倒單擺系統為實例，來驗證其適應性。模擬結果如下：灰色逆傳遞學習法比傳統逆傳遞學習法的平方誤差小，且性能指標顯示亦較好。傳統逆傳遞學習法，當輸出的模糊集合分割愈細，則性能指標愈好。但灰色逆傳遞學習法與模糊集合分割數目並無絕對的關係。 In this dissertation, some combinations of Genetic Algorithms, Grey Theorem, Fuzzy Control, and Neural Networks are studied. Their applications to an active suspension system for a full-car and an inverted pendulum are illustrated by examples. First, we demonstrate some basic propositions of the GM (1,1) model. The behaviors of the development coefficient and the grey input are proposed to simplify the calculation procedure. Furthermore, the implementation of the Genetic Algorithms in optimizing the generating coefficients of the Grey Model, GM(1,1), improves the prediction values of the modeling with respect to residual errors. Secondly, we present a grey-fuzzy control and a fuzzy neural networks control, respectively, to design an active suspension system for a full-car. Our primary control goal is to emphasize the amelioration of body oscillation. The results clearly indicate that the proposed fuzzy neural networks controller outperforms the passive, the fuzzy logic and grey-fuzzy controllers in providing the desired ride quality. The grey relational method is also applied to evaluate the grade of vehicle oscillation. The simulation data shows that the grey relational analysis method is a good option in the analysis of vehicle oscillation. Finally, back propagation (BP) neural networks in conjunction with a grey relational coefficient is used to discern the optimal partitions of the consequent part in fuzzy neural networks. The learning capability of the proposed method has been demonstrated using an inverted pendulum. The square errors by BP with GRC is much smaller than that of the classical BP in the simulation.
