On-line Measurement Of Product Quality And Process Model Based On Industrial Data

Data-driven process models play a key role in traditional soft sensors technology and modern process analytical technology (PAT) as they can be used to estimate hard-to-measure process variables in real-time, consequently online optimization and control of product quality can be achieved. In recent years much progress has been made in research in data-drive model development, but some literature has used simulation data or laboratory data that cannot fully reflect the challenges posed by industrial applications and the real process data. In fact, many industrial applications of data-driven models encounter difficulties but no readily available solutions can be found in literature. No doubt the most realistic approach to address these challenges is to conduct research on data-driven model building by reference to real industrial applications and data. In this research, on the basis of reviewing state-of-the-art of data-driven methods for process modeling, data-driven model building methods were investigated for two real industrial processes, namely, atmospheric distillation of crude oil and extraction of effective components of a Chinese medicine compound. The main research outcomes are summarized below:A systematic approach based on robust PLS for development of a data-driven static soft sensor is presented and used to predict the final boiling point of reforming feedstock of a crude oil distillation column. The proposed method consists of numerous algorithms, i.e. univariable pretreatment method of wavelet and box figure, determination of lag time using genetic algorithm, identification of outliers and modeling by robust PLS. The final model proved to have achieved excellent prediction performance on unknown data.Dynamic soft sensors models were also established for capturing the dynamic feature of industrial processes. Grid search method was used to optimize maximum lag time and resampling interval of the dynamic soft measurement modeling. The simulation results showed that the prediction error and computation of the soft-sensors decreased obviously. The proposed model based on optimized dynamic PLS is better than a general static soft sensor model.Considering the nonlinear characteristics of the industrial atmospheric distillation column, the results of different modeling methods (PCR, PLS, ANN or SVM) were examined and compared. Best prediction performance was achieved by the PLS method, which may be a result of the continuously steady operation of the process, whose state can be approximated to be linear.In order for the models to be used for over a long period of time during which there might be deviations of process conditions from the original state, a moving window method was incorporated for the automatic adaptation of models. Results show that the models can be adapted to the change of the process and then provide a reference for optimization and control of product quality.The near infrared (NIR) model is established for estimating the polysaccharide concentration of a Chinese medicine extract. First derivative and autoscale methods were employed for the pretreatment of data. Random frog method was then employed to select the key analytical wavelengths. Afterwards PLS was used as the modeling approach for model development. The results obtained in this study indicated that the model of NIR spectroscopy can satisfy the industry’s requirement polysaccharide concentration estimation of the Chinese medicine extract.