基于过程神经网络的稀疏数据过程建模方法

Modeling of Sparse Data Processes Based on Process Neural Network

针对一类因检测困难而导致检测数据稀少的连续工业过程,提出了基于离散Walsh变换的过程神经网络建模方法。在对稀疏样本数据进行预处理的基础上,采用递推式非邻均值生成法对样本数据进行扩充,以此建立可产生任意密集预测数据的过程神经网络模型,并采用在线滚动学习的方法进一步提高所建立的预测模型的精度。以味精发酵过程菌体浓度预测为例,验证了所建立的过程神经元网络预测模型可以得到非常高的预测精度。

The modeling method of the process neural network （PNN） based on Walsh conversion was proposed, which aimed to solve the problem of sparse sampled data for some industry processes. On the base of the pretreatment for the sparse sampled data, they were extended to train the PNN model, which could produce any required time-interval data, according to the recursive unneighbored average value. The PNN model was trained by the way of rolling learning to revise the error timely. The germ concentration in the fermentation process of the glutamic acid was pre-estimated by the PNN model. The simulation results show that the process neural network has good generalization ability.