臺灣博碩士論文加值系統

類神經網路的蓬勃發展及對多肢體系統 (Mutibody Systcm)的興趣，使得近年來在結合這二個領域的研究發展非 常迅速。多肢體系統的動態模型因模型的簡化，而比完整的動 態模型具有更強大的計算效率，但傳統的比例微分控制器(PD Controller)對此系統腳的狀態變換無法即時反應，而造成運動時各關節角的誤差﹔在此系統中加入類神經控制，希望藉由類神經網路的優點使整個系統更具有彈性，當腳舉起或著地之運動狀態改變時系統的穩定性可達到更佳的狀態。
本文提出以倒傳遞網路(Back-Propagation Neural Network)作為類神經控制器之基礎，並以此類神經控制器作為原系統之前饋控制器(Feedforward Controller)，經過模擬後確實可改善腳之運動狀態改變時所造成的不穩定現象，而且改善了各關節角度的誤差值。但受限於處理器對模擬程式的計算速度及倒傳遞網路本身的特性，所以仍無法作到即時(Real Time)之動態模擬。

For the flourishing development of the neural network and the interest on the multibody system, the progress of the recent researches that combine these two fields is rapid. By the model simplification dynamic simulation of the multibody system has gained more computational efficiency than the complete model. But the classical PD controller can not response immediately when the state of walking leg in the system has changed. For this reason, it will cause some errors of each joint angle. We expect to improve the stability of the system with the merits of neural network and to make the system flexible no matter the legs are in stance or recovery state.
In this thesis, the back-propagation neural network is employed to constitute the neural controller which will be used as the feedforward controller in the hexapod system. The simulation results indicate that the neural controller have improved the unstable state of walking legs and decreased the errors of every joint angle. Nevertheless, the real-time simulation is still unreachable due to the speed of the processor and the property of the back-propagation network.