近年來,人工智慧的研究及發展愈來愈受注目,而類神經網路是人工智慧的一個研究領域。在利用類神經網路進行問題處理時,常會因為針對的問題不同,而有不同的類神經網路運算架構。本論文開發出一套根據倒傳遞類神經網路運算模式而建構出來的硬體運算架構,只要設定所需要運算的節點數,即可完成倒傳遞類神經的運算架構。由於本系統是在硬體上開發,因此在移植性及運算速度上將會比在軟體上開發還要來的便利及快速。 由於倒傳遞類神經網路需要透過轉移函數進行運算,但轉移函數是一個非線性函數,對於硬體而言,要實現非線性函數是一個挑戰。而且在硬體中沒有專門的演算法可以直接完成轉移函數運算,因此只能用近似化的方法規劃出非線性函數模型。在本論文中,以三種近似化轉移函數的演算法,分別為泰勒級數展開法、片段近似法及查表法進行分析與討論,進而完成轉移函數的硬體設計。 本文最後在Altera公司所提供NiosII StratixII EP2S60F1020C4開發平台上加以實現。並透過XOR問題及過電流保護電驛反時曲線模擬當作系統驗證與效能分析。
Recently, the research and development of artificial intelligence has been studied by many researchers. Artificial Neural Network (ANN) is one of the fields in artificial intelligent which has been studied and applied to real problems. When applying ANN to different problems, structure needs to be modified in order to fit the characteristic of the problems. This thesis proposed a circular structure which requires less modification for adapting to different problems. The development of the ANN is implemented on FPGA. Thus, the portability and performance of calculation of the proposed system is better than software based ANN systems. Transfer function is an important element in Back-Propagation ANN. Because this function is a non-linear function, it is a challenge to implement such function in hardware. This thesis studies three models, Taylor series, Cut-line and look up table, to implement the transfer function. Finally, the proposed ANN system in this thesis is implemented on a NiosII StratixII EP2S60F1020C4 developmental board by Altera. Two examples, XOR and over current protected relay time-current curve are used to validate the correctness and verify the performance of the proposed sysstem.