具有方塊脈衝歸屬函數模糊類神經網路之研究及其應用

Abstract

模糊類神經網路(FNN)過去十年來廣泛應用在即時智慧型控制、影像處理等不同的工程領域，歸功於其自我學習及訓練能力。而更早的查表法技術(TL)則有無需繁複的計算且收斂快速等優點，所以若能結合兩者優點將是一種運用在即時人工智慧控制極具潛力的工具。 本論文提出與查表法技術同等簡易的具有非重疊方塊脈衝歸屬函數(BPMFs)之FNN控制器架構，並將兩者之隱藏連繫關係作一完整探討，提供一套基於FNN設計觀念來設計TL控制器的方法；而此一改良的架構也應用到汽車渦輪引擎噴水控制系統及高轉速無刷直流馬達(BLDC)控制器上，透過即時學習來克服系統非線性，並透過實作驗證本系統在汽車渦輪引擎高溫惡劣操作環境及無刷直流馬達高速三相轉換時間下仍保有極佳的性能，顯示以具有BPMFs的FNN為基礎的TL控制器是一強健的控制系統。實作中發現傳統模糊類神經網路的模糊規則需依專家制定及實驗改良，本文提出一套直接產生模糊規則的設計方法，在BPMFs條件下不僅保證輸入空間向量一對一映射至輸出空間向量並且可加快線上學習及減少即時運算時間。

This thesis presents an alternative method to design Fuzzy Neural Network (FNN) using a set of non-overlapped block pulse membership functions (BMPF's), and this FNN with non-overlapped BPMF's will be shown to be equivalent to the conventional Table Lookup technique. Therefore, the relationship between Table Lookup and FNN techniques are presented in this thesis which provides a methodology to design a Table Lookup controller based on FNN design concept. In order to do so, a new direct formula is first developed to generate the fuzzy rules from the premise part in FNN. This direct formula not only guarantees a one to one mapping that maps the fuzzy membership functions onto the fuzzy rules, but also alleviates the coding effort during hardware implementation. It is further elaborated that the FNN with non-overlapped BPMF's has the advantage of faster on-line training which requires less computation time, but at the cost of more memory requirement to store the fuzzy rules. In addition, this dissertation also shows the approach described above has been applied successfully in both the water injection control of a turbo-charged automobile and the fuzzy logic controller (FLC) for high-speed sensor-less brushless dc (BLDC) motor with excellent results.