Please use this identifier to cite or link to this item:
標題: 輻射基底函數式類神經模糊網路
RBF Based Neural Fuzzy Network
作者: 黃朱瑜
Huang, Ju-Yi
關鍵字: RBF;輻射基底函數類神經網路;neural-fuzzy;motor position control;on-line training;CMP;Taguchi method;parameters optimal;類神經模糊網路;交流伺服馬達定位控制;線上學習;化學機械研磨;田口法;參數最佳化
出版社: 機械工程學系

In this thesis, a new neural fuzzy configuration that combines the RBF neural network structure and fuzzy logic theory is proposed.
In this new neural fuzzy structure, the conventional six layers neural fuzzy network is simplified to a four layers neural fuzzy network. For single input problem, this new network structure is a kind of RBF neural network. When a multi-inputs problem is applied, it functions similar a conventional neural fuzzy network.
Computer simulation results show that the proposed new neural fuzzy scheme can be successfully applied to the nonlinear function approximation and classification problems.
To fulfill the on-line training requirement, an efficient heuristic learning rule is included. Experimental results show that the proposed approach can be successfully applied to the precise regulating and tracking problems of an AC servo motor system.
For real industrial application, a systematic approach to achieve global optimal CMP process is carried out. In this new approach, orthogonal array technique in the Taguchi method is adopted for efficient experiment design. The RBFNF neural-fuzzy is then used to model the complex CMP process. Signal to Noise Ratio (S/N) Analysis technique used in the conventional Taguchi method is also implemented to find the local optimal process parameters. Successively, the global optimal parameters are acquired in terms of the trained RBFNF network. In order to increase the CMP throughput, a two-stage optimal strategy is also proposed. Experimental results show that the two-stage strategy can perform better then the original approach even though the process time is reduced by 1/6.
Appears in Collections:機械工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.