应用专家知识提高神经网络建模的泛化能力的研究  

Application of expert knowledge to improve the generalization ability of neural network modeling

作　　者：张杰[1] 桑志祥[1] 李绍军[1] 

机构地区：[1]华东理工大学,化工过程先进控制和优化技术教育部重点实验室,上海200237

出　　处：《计算机与应用化学》2012年第9期1111-1114,共4页Computers and Applied Chemistry

基　　金：国家自然科学基金资助项目(20976048,21176072)

摘　　要：普通意义上的神经网络建模缺少物理基础,对实际工业过程机理信息掌握不足,模型的外推效果不够理想。针对这种情况,提出将专家知识融合到神经网络建模中的方法。首先提取工业过程中存在的机理信息,在网络模型训练阶段,提前对关键变量进行灵敏度分析,并将模型灵敏度分析的结果同专家知识相对比,根据两者间的违反程度差异对模型目标函数进行不同程度的惩罚。在结晶动力学模型的仿真研究结果表明,这种方法一定程度上克服了神经网络训练的盲目性,特别是针对训练数据缺失或者存在噪声的情况,能够有效的提高神经网络的泛化能力。In ordinary sense, the neural network modeling is lack of physical basis and mechanism information of industrial process, so the prediction effect of model is unsatisfactory. In view of this situation, a method combining expert experience with the neural network modeling was presented in this paper. First, mechanism information of the industrial process was extracted. In the training process of the model, sensitivity analysis on key variables was carried out and the results of sensitive analysis were compared with expert knowledge. According to the violation diversity between them, different degrees punishment was implemented on the model objective function. Simulation results of the crystallization kinetics model show that, this method can overcome blindness in the training process to some extent, and can improve the generalization ability of neural network, especially when the number of training samples is small.

关 键 词：神经网络 专家知识 过拟合 优化 灵敏度分析 

分 类 号：TQ015.9[化学工程] TP391.9[自动化与计算机技术—计算机应用技术]