Research On Multi-Model And Online Soft Sensor Using Semi-Supervised Regression

With development of modern society and demand of process technologies and environmental protection, industrial process has became more and more complex and is characterized by multi-operating conditions, time-variation and strong nonlinearity. In order to control processes accurately, process variables should be able to reflect the state of processes real-timely and accurately, therefore, the measurement of these variables is becoming an important part of industrial control system. If use the traditional method of sensor acquisition to measure these variables, the investment and maintenance costs are usually very high, and even there are delays in some variables'measurements, so sensor acquisition sometimes cannot reflect the changes of processes in real time. Nowadays, soft sensor becomes an effective way to solve the problem of these variables'measurements. For those non-linear and multi-state industrial processes whose operating conditions often change, a single static model usually cannot reflect the actual situation of the process correctly, but if combine soft sensor technology with online modeling technology and multi-modeling technology, the accuracy and reliability of soft sensor measurement can be improved effectively.In this paper, a research of online and multi-model soft sensor system basing on semi-supervised regression is done for those complex industrial processes whose operating conditions often change. Simulations and application have shown correctness and practicality of this research. All of the work is listed as below:1. Elaborated on concepts of soft sensor technology's procedures and modeling methods; introduced semi-supervised technology, online technology and multi-model technology and their developments.2. Proposed an online soft sensor modeling method basing on semi-supervised Local Linear Regression. This method makes Local Linear Regression become a semi-supervised method by importing unlabeled data to its objective function, and simulations proved the semi-supervised method can improve the system's accuracy successfully. For parameters'selecting, this paper proposed an adaptive method which selects Gaussian kernel width by dominate variable's estimated slope in semi-supervised Local Linear Regression. At last, basing on the above research, this paper used rolling time window to propose an online method of SSLLR, and this method makes soft sensor model can be updated in real time and be able to adaptive to input data, simulation result shows that the adaptive parameter selection method makes SSLLR have a better prediction. 3. Researched on multi-model soft sensor modeling. For those industrial processes which have multi-operating conditions and strong nonlinearity, this paper proposed two multi-model soft sensor modeling methods basing on semi-supervised regression: weighted multi-modeling method and switched multi-modeling method. Weighted multi-modeling method uses all sub-models which are created by input data's characteristic to predict the dominate variable at first, then gets fuzzy membership degrees of all input data by fuzzy clustering method, and uses these membership degrees as weights to compute multi-model prediction result by all sub-models'predictions at last; Switched multi-modeling method clusters all input data into different categories at first, and uses only one specific sub-model to predict the dominate variable, and this sub-model is determined by input data's characteristic, finally makes the sub-model prediction result be the multi-model prediction result directly. Both of the two multi-modeling methods are simulated in this paper, and the results show that multi-model can really improve the accuracy of prediction.4. At last, this paper designed a multi-model soft sensor system's application based on semi-supervised regression. The application focus on ultra-supercritical units, and flue gas oxygen content is chosen as the dominate variable, besides, a dozen sets of variables which are obtained from industrial production are chosen as the auxiliary variable. The application is designed into two parts: offline part and online part. Detailed descriptions of these steps, such as working conditions analysis, data preprocessing, are illustrated in this paper. At last, a real flue gas oxygen content prediction is done with this designed application, and the result shows its accuracy and usefulness.