推導了隱層節點和輸出節點的學習及修正公式，選擇了初始權值、學習速率和期望誤差等相關學習參數，結合結晶工段實際進行了網絡訓練。為加快網絡的收斂，在保證系統穩定的基礎上，提出了S 函數輸出限幅和變步長法兩種改進方案。采用改進的BP 算法對此BP 網絡進行了訓練，得到的ANN 網絡訓練輸出值與實際輸出值非常逼近。
關鍵詞：BP 算法；網絡訓練；收斂；修正

Algorithmic Choice and Betterment in the BP Network ofCane Sugar Crystallization YANG Jie1,2, YAN Qingdong1 (1. School of Communication Machinery and Civil Engineering, Southwest Forestry University, Kunming, Yunnan 650224, China; 2. School of Mechanical and Vehicle Engineering, Beijing Institute of Technology, Beijing 100081, China ) Abstract: The learning and amendatory formulae of the hidden layer nodes and deferent nodes were deduced. Initial weight value，learning rate and expectation error etc. correlative learning parameters were selected, then the BP network was trained integrating practical instance. In order to quicken convergence, S function output violent change and method of changing step length which two improved project were brought forward in the basis of system stabilization. And the BP network was trained by using improved algorithm, the result showed that ANN
model output extremely approximated to practical output.
Key words: BP algorithm; Network training; Convergence；Amendment