论文部分内容阅读
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。
Ordinary neural network modeling lacks the physical basis of the actual industrial process mechanism to master the lack of information, the model of the effect of extrapolation is not ideal. In view of this situation, a method of integrating expert knowledge into neural network modeling is proposed. Firstly, the mechanism information existed in the industrial process is extracted. In the network model training phase, the sensitivity analysis of the key variables is carried out in advance, and the result of the model sensitivity analysis is compared with the expert knowledge. According to the difference between the two, the model objective function Different degrees of punishment. The simulation results of the crystallization kinetics model show that this method overcomes the blindness of neural network training to a certain degree. Especially for the lack of training data or the existence of noise, this method can effectively improve the generalization ability of neural networks.