摘要
神经网络具有优良的非线性映射逼近能力,广泛应用于化工过程建模,但神经网络建模方法属于黑箱法,所获得的模型缺乏透明性,各变量的解释性差,限制其指导化工企业优化技术决策。结合神经网络释义图、连接权法和改进的随机化测验三种方法,对复杂化工过程神经网络模型进行透明化研究。首先利用神经网络释义图可视化模型,再用连接权法对决策参数贡献率定量分析,最后利用改进的随机化测验,对模型的连接权、决策参数的综合贡献度和相对贡献率进行显著性检验,进而修剪模型。通过对复杂化工过程氢氰酸生产模型验证研究,结果表明该方法获取了过程变量的内部信息,极大地提高了模型的"可理解"能力。因此,本研究为复杂化工过程神经网路模型的透明化提供了一条很好的途径。
Neural networks are widely used in chemical process modeling, due to the excellent approximation capability of nonlinear mapping, while the modeling approach is a black-box method, so that the model obtained by the method is lack of transparency, each variable has poor explanation ability, thus it is limited to guide and optimize technical decisions in process indusry. Three methods that network interpretation diagram, connection weights and improved randomization texts are combined to improve transparency of neural network model in complex chemical process. Fistly, network interpretation diagram is used to visualize black-box of the network mod- el ; then connection weights method is employed to calculate the relative contribution ratio of decision variables ; lastly, the connection weights, the overall contribution and relative contribution rate of decision parameters are included to implement significance tests using the suggested randomization tests for trimming the model. Verification is carried by the case of hydrocyanic acid production; the results show that the proposed method could obtain internal information of process variables and greatly improve the understandable capacity of NN model. Therefore, the study provides a fine way to improve the transparency of neural network model in complex chemical process.
出处
《控制工程》
CSCD
北大核心
2013年第1期115-120,共6页
Control Engineering of China
基金
国家自然科学基金(51075418)
国家自然科学基金(61174015)
重庆市自然科学基金(CSTC2010BB2285)
关键词
神经网络模型
透明化
化工过程
网络释义图
随机化
neural network model
transparency
chemical process
network interpretation diagram
randomization
