基于时间近邻拉氏正则的多工况软测量回归

针对流程工业中,工况改变导致传统软测量模型预测精度下降的问题,考虑到工业数据连续性、序列性、多重共线性、数 据量庞大等特殊性对模型建立的影响,提出一种基于时间近邻拉普拉斯正则的多工况软测量回归模型框架。 针对工业数据的 多重共线性,回归框架采用非线性迭代偏最小二乘方法,同时引入域适应正则项改善工况变化对模型的影响,在此基础上,提出 时间近邻拉普拉斯正则项,能够在映射过程中保持住数据的序列结构,并且大幅度减少模型训练时间以满足工业实时性要求。 实验部分以三聚氰胺聚合过程多工况数据集为例,对本文模型的预测有效性以及减少训练时间的有效性进行了实验和分析。 结果表明,与传统方法偏最小二乘回归相比,当目标工况为工况 1 到工况 4 时,本文方法使平均均方根误差分别降低了 30. 3% 、 31. 4% 、29. 3% 和 24. 1% 。 且相较于传统全连接法,时间近邻法构建拉普拉斯正则项能够使得四个工况上模型训练时间分别降 低 14. 11、 1. 01、 26. 43 和 0. 71 s,表明该模型的预测准确性和训练时间均得到有效改善.

Multi-conditions soft sensor regression based on the time-nearest neighbor Laplacian regularization

In the process industry, the change of working condition results in the decrease of the prediction accuracy of traditional softsensing models. Take the impact of industrial data continuity, sequence, multicollinearity, and huge amount of data on the model establishment into account, a multi-conditions soft sensor regression model framework based on the time-nearest neighbor Laplacian regularization is proposed. To solve the multicollinearity of industrial data, the proposed regression framework utilizes the nonlinear iterative partial least squares method. Meanwhile, the domain adaptation regular term is introduced to mitigate the influence of the change of working conditions on the model. On this basis, the time nearest neighbor Laplacian regular term is proposed, which can maintain the sequence structure of the data during the mapping process. And the model training time is greatly reduced to meet the industrial real-time requirement. In the experiment, the multi-conditions data set of the melamine polymerization process is taken as an example. The results show that compared with the traditional method of partial least squares regression, when the target conditions are conditions 1 to 4, the method in this paper reduces the average root mean square error by 30. 3% , 31. 4% , 29. 3% and 24. 1% , respectively. And compared using with the traditional function of total connecting, using the function of time-nearest neighbor connecting to construct the Laplacian regularization could deduce the training time of the four working conditions by 14. 11, 1. 01, 26. 43, 0. 71 s respectively, and indicated that the accuracy and the training time could be improved.