Clustering-Based Hybrid Soft Sensor for an Industrial Polypropylene Process with Grade Changeover Operation

Abstract

A new methodology is proposed to design a soft sensor for a polypropylene (PP) process with grade changeover operation. In contrast to the general polyolefin process, the PP process usually produces more than 100 different grades of products. Its reaction mechanism, based on seven catalysts, is so complex that neither mechanistic nor empirical models have been successful in describing full-scale industrial applications. The proposed methodology was developed based on the hybrid modeling of novel clustering and black-box and mechanistic models. Clustering based on critical to quality enables the soft sensor to handle the complexity of many different grades. Hybrid modeling offers good predictive power for transient behaviors as well as normal behaviors. The methodology also allows us to reduce the cost of building and updating the model. The developed soft sensor was successfully applied to a real industrial process. The accurate and reliable monitoring of the melt index in the PP process helped to significantly reduce the amount of off-specification product generation.

Three-Stage Decomposition Modeling for Quality of Gas-Phase Polyethylene Process Based on Adaptive Hinging Hyperplanes and Impulse Response Template

“…8,18,19 As shown in eq 6, the easy-to-measure inputs U(t) and the difficult-to-measure quality MI c (t) are modeled in the regression form, in which the sampling interval of the low-rate variable is p (p ≫ 1).…”

“…Since operating conditions and polymer properties vary over a wide range and significantly depend on the grade being produced, it is difficult to accurately model the dynamic nonlinear characteristics of the quality variables under frequent operations. 18,19 To overcome the difficulties caused by multirate-sampled data, there are two simple technologies: down-sampling to the slow rate of quality measurements and up-sampling to the fast rate of operational measurements. 8 However, down-sampling may lose a great deal of useful information of operation conditions, while up-sampling of the sparsely quality data may incur great noise.…”

“…8 Meanwhile, to overcome the difficulties incurred by nonlinear dynamic characteristics, some researchers have designed a clusteringbased preprocessor before modeling to divide all the data into different subdomains and then construct local submodels over every domain, which correspond to different operating states of the process. 7,18,19 Kaneko et al proposed a discriminate model that detects the completion of transitions and regression models constructed for each product grade. 17 However, the prediction accuracy of the models established by data-driven methods significantly depends on the data training and may be difficult to reflect different operating conditions over a wide range, especially when there is great uncertainty in the process.…”

Extracting dynamic features with switching models for process data analytics and application in soft sensing

“…As for modeling an input-output system, the model identity of each data point is usually determined together with parameter estimation. [20][21][22] As for the latent feature extraction, several structures have also been proposed from different assumptions. If switching behaviors are assumed to follow a Markovian transition, Hidden Markov Models 23,24 and other extended hierarchical structures 25,26 can be used to estimate the model identity sequence.…”

Inferential Estimation of Polymer Melt Index Using Sequentially Trained Bootstrap Aggregated Neural Networks

“…The inferential estimation of polymer quality, based on empirical models, has been investigated in order to overcome this difficulty [5,6]. The empirical models can be developed from process operation data.…”