Studies On Soft Sensor Modeling Methods And Their Applications In Industrial Process

The higher performance requirements of control systems for modern industrial process have been promoting the development of soft sensor technique. In modern complicated industrial process, some variables are very hard to be measured or even cannot be measured on-line by existing instruments and sensors. Soft sensor is an effective means of implementing the on-line evaluation of these variables. At present, soft sensor technique has become one of the most important research areas in process control field. Combining technics knowledge of chemical engineering process, some modeling methods of soft sensor technique are studied intensively in this dissertation. Furthermore, the problems of soft sensors implementing for industrial processes are also studied and solved. The main contributions are described as follows:1) A new soft sensor modeling method based on improved FasBack neuro-fuzzy system is developed. Levenberg-Marquardt algorithm is used to train some parameters in the model, while the residual parameters are still trained by BP algorithm. Based on practical process data, the proposed improved FasBack neuro-fuzzy system is applied to build soft sensor model of 4-CBA concentration of purified terephthalic acid (PTA) product. Simulation results indicate that the proposed model is precise and efficient and it is possible to realize the quality control of PTA product in the commercial reactor. Because of the powerful clustering and nonlinear regression capability, FasBack neuro-fuzzy system is also very suitable to multi-input multi-output (MIMO) soft sensor modeling. So, the proposed model is also used to build a MIMO soft sensor model to evaluate the three quality variables simultaneously in compound fertilizer process. Simulation results indicate the proposed model possesses good convergence and prediction precision. It is a useful try in MIMO soft sensor modeling method.2) In view of the problem that the three quality variables in compound fertilizer production need to be monitored and controlled simultaneously, a new modeling method of multi-inputs multi-outputs (MIMO) soft sensor, which is constructedbased on hybrid modeling technique, is proposed for these interactional variables. The technics information of the process is fully used in this modeling method, that is combing the simplified first principle of the process and data-driven modeling method toghther to build the MIMO soft sensor model of the contents of compound fertilizer process. The pertinence and redundancy of the data are considered at the same time, so limited memory PLS algorithm that is very powerful in solving pertinence and redundancy is chosen as the data-driven modeling method. Futhermore, a new variance recursive algorithm is adopted in the limited memory PLS algorithm, therefore the PLS model can be updated on-line. Simulation results based on practical process data indicate that the proposed model is fast, precise and efficient and it is possible to realize the on-line quality control for compound fertilizer.3) The MIMO soft sensor model that is based on hybrid modeling method is applied for a practical compound fertilizer process to evaluate the content of nitrogen, P₂O₅ and K₂O on-line simultaneously, and therefore process monitoring is implemented. The detail procedures are described in detail, including the software and hardware platform of soft sensor implementing, data collection, data pretreatment, time alignment matching between the primary variables and the process variables, the modeling steps of the soft sensor, and the on-line soft sensor model rectification. In this dissertation, the on-line evaluating values are presented. The results indicate that the proposed model can satisfy the requirement of on-line evaluating of the three quality variables in compound fertilizer process.4) In order to overcome the problem that soft sensor models cannot be updated with the process changes, an adaptive soft sensor modeling algorithm based on hybrid modeling method is proposed. In this hybrid modeling method, the training samples are firstly clustered by fuzzy c-means (FCM) algorithm, and then by training each clustering with SVM algorithm, sub-model is built to each clustering in order to improve the evaluation precision of the soft sensor model. When an incremental sample that represents new operation information isintroduced in the model, incremental learning is applied to the corresponding SVM sub-model in order to reducing computing time and increasing the model's adaptive abilities to various operation conditions. Because the computing complexity of SVM depends on the number of the support vectors, when a new support vector is added, an old support vector chosen by heuristic sample displacement method was then discarded from the sub-model to control the size of the working set. The proposed method is applied to predict the p-xylene (PX) purity in an adsorption separation process. Simulation results indicate that the proposed method actually increases the model's adaptive abilities to various operation conditions and improves its generalization capability.