Abstract:  A practical industrial process is often a large-scale complex system with muti-operating modes and nonlinearities, making it difficult to fully mine the data information through a single soft sensor model. To solve this problem, a multi-model soft sensor development approach based on the density peak (DP) clustering and the random forest regression (RFR) is proposed to estimate dominant variables. Firstly, classify the training data by means of the DP clustering algorithm; secondly, establish regression sub-models based on the samples of each category by using the RFR method; finally, apply the switching method for multi-model fusion. The proposed method has been utilized to develop soft sensors of the Tennessee Eastman process and the butane distillation process for estimating the contents of G and propane, respectively. The simulation results illustrate that the estimation accuracy has been improved, which can verify the effectiveness of the proposed method.

Key words:  Soft sensor, density peak, random forest regression, multi-model

摘要： 实际工业过程往往是一个多工况、非线性的大规模复杂系统，使得单一模型软测量建模方法难以充分挖掘数据信息。针对这一问题提出了一种基于密度峰(Density Peak, DP)聚类和随机森林回归(Random Forest Regression, RFR)的多模型软测量建模方法，从而对主导变量进行估计。首先，利用DP聚类算法对训练数据进行划分；其次，采用RFR方法建立各子类的回归子模型；最后采用开关切换的方法进行多模型融合。将提出方法应用于TE过程和丁烷蒸馏过程的软测量建模中，分别对产物G含量和丙烷含量进行估计。仿真结果表明估计精度得到提高，证明该方法是有效的。

关键词: 软测量, 密度峰, 随机森林回归, 多模型