中文摘要：

集成建模方法能显著提高软测量模型的预测性能,其中选择性集成通过剔除一些性能不佳的子模型,能进一步提高整体软测量模型预测性能。针对目前选择性集成研究中因忽略了数据间的差异性而导致模型预测性能不佳的问题,提出了一种动态选择性集成神经网络软测量建模方法。首先将原始数据集分为训练集和验证集,采用bootstrap算法对训练集进行差异性扰动,建立了多个神经网络子模型;然后对每个待测样本,采用K-最近邻搜索算法从验证集中找到一个最近邻子集,用该子集评估各神经网络子模型的预测性能,为待预测样本筛选合适的神经网络子模型;最后根据各子模型的预测性能合理分配组合权重,从而建立集成模型,并实现待预测样本的预测。将该建模方法应用于聚丙烯熔融指数软测量研究中,研究结果表明：与单一神经网络、常规全集成和静态选择性集成神经网络模型相比,基于动态选择性集成神经网络的熔融指数软测量模型具有更佳的预测精度。

英文摘要：

Ensemble modeling method can significantly improve the prediction performance of the soft sensor model. Selective ensemble makes further improvement on the prediction performance of the whole model by rejecting some badly-behaved sub-models. Aiming at the problem existing in present selective ensemble research that ignoring the difference between data leads to a poor performance of the model, a modeling method based on dynamic selective ensemble neural networks was proposed. Firstly, the original sample data was divided into training set and validation set, and bootstrap algorithm was applied to differently disturb the training set and several neural network sub-models were developed. Then a nearest neighbor data set from the validation set which formed by K-nearest neighbor search method was used to evaluate each sub-model and select appropriate neural network sub-models for the query data. Finally, the combination weights were properly allocated for the ensemble model according to the prediction performance of the sub-models and the established ensemble model was used to predict the query data. The proposed modeling method was applied to develop polypropylene melt index soft sensor. The result indicated that the polypropylene melt index soft sensor based on dynamic selective ensemble neural networks had better prediction accuracy than the single neural network model, fully ensemble neural network model and the static selective ensemble neural network model.