| Free radical polymerization is a main technique in polymer synthesis,which has been widely used in various fields such as chemical engineering,pharmaceuticals and energy industry.With the digital upgrading of manufacturing industry,online monitoring of polymerization process has become an urgent demand.By means of real-time monitoring of key process parameters in the polymerization processes,industrial objectives such as process optimization,reducing energy consumption and quality improvement can be achieved.Since many of the key parameters in polymerization processes cannot be measured directly,machine learning-based soft-sensing has become a hot topic in on-line process monitoring.However,there is a contradiction between cost and accuracy in soft sensing methods.Models based on low-cost modeling parameters struggle to accurately represent the production process,while the in-situ detection devices capable of accurately characterizing the process are not widely adopted in practical production due to their high cost.Therefore,achieving low-cost and high-accuracy soft sensing of polymerization process becomes a challenging and practical research topic.In this thesis,we took the preparation of formaldehyde-free adhesives via polymerization reaction as a research example,and the polymerization progress as the soft-sensing target parameter.In situ infrared spectroscopy was used to monitor the polymerization process in real time.The concentration of each component in the polymerization process was quantitatively analyzed using infrared spectroscopy,so that the soft-sensing could be achieved by modeling the relationship between concentration and polymerization progress.However,the cost of in-situ infrared spectroscopy monitoring the production process is very high.In order to achieve low-cost and high-precision soft sensing,we propose a soft sensing method of polymerization process based on multi-task learning,which takes temperature as the modeling parameter,and infrared spectroscopy prediction as an auxiliary task to predict the reaction schedule.This approach enables accurate soft sensing of the polymerization progress using low-cost temperature data.The main contributions of this thesis are summarized as follows.1.Data acquisition and preprocessing.We designed and implemented a data acquisition scheme for free radical polymerization process modeling,and preprocessed the dataset by Kalman filtering,spectral data enhancement,data standardization,and feature extraction.2.We proposed a quantitative analysis method of solution infrared spectrum based on chain curve-fitting.First,the second derivative spectrum of the solution samples are used to identify the absorption peak positions,and the half-widths are obtained by fitting the second derivative spectra with the constraint of peak positions.Then the peak heights are obtained by fitting the first derivative spectra with the constraints of peak positions and half-widths.Finally,a relationship model between the peak heights and the component concentrations was established using Partial Least Squares(PLS)regression,enabling accurate inversion of component concentrations from solution sample infrared spectra 3.We proposed a soft sensing method of free radical polymerization progress based on multi-task learning,which took the polymerization temperature as the modeling parameter,a two-layer LSTM(Long Short-Term Memory)network as the base learner.The model consists of two tasks:the main task was to predict the polymerization progress,and the auxiliary one was to predict the infrared spectral absorption peak of the reaction solution.Through the auxiliary task,the encoder extracts the implicit correlation information between the infrared spectrum and the reaction progress,which realizes the low-cost and high-precision soft sensing of the aggregation progress.The effectiveness of the proposed methods have been verified by experiments.The results show that the method of solution infrared spectrum based on chain curve-fitting has a coefficient of determination of 0.9912 and the absolute error of 0.0056,and the soft-sensing method based on multi-task learning achieves a coefficient of determination of 0.9688 and a mean square error of 0.0056. |
