中文摘要：

神经网络集成可以显著提高神经网络的泛化性能。传统的集成方法中大都采用将训练的所有网络直接进行组合的方式形成集成网络,而实际上这些网络可能具有一定的相关性。为此,选择性神经网络集成成为目前研究的热点,它能够进一步提高集成网络的泛化性能。本文提出了一种利用网络权值计算网络模型之间差异度的新的选择性神经网络集成方法DWSEN。UCI数据测试表明,与流行的集成方法Bagging和Boosting比较,本方法有着更好的泛化能力和稳定性。将DWSEN应用于精对苯二甲酸（PTA）溶剂系统脱水塔装置的建模过程,结果显示,利用该方法训练得到的集成模型具有更好的泛化性能,能够较好地模拟生产运行过程。

英文摘要：

Neural network ensemble could dramatically improve the generalization performance of neural network. In traditional ensemble processes, all the trained networks are directly combined to the integrated network. However, these networks may have certain correlation in fact. Therefore, selective neural network ensemble has become a hot issue recently, by which the generalization ability of neural network ensemble can be further improved. Thus, the authors propose a new selective constructing approach to neural network ensemble named DWSEN through measuring the diversity of individuals according to weights of networks. Compared with some prevailing ensemble approaches such as Bagging and Boosting, testing of UCI data sets illustrated that the DWSEN approach has higher generalization ability and stronger stability. This method is further validated by the modeling of solvent dehydration tower of purified terephthalic acid （PTA） solvent system. Case study shows that the obtained model has better generalization performance, and can simulate the production process better.