Development Of Soft Sensor For The Plant-Wide Process

With the development of modern process industries, the diversity of products’ specification is increasing and system is becoming large and complicated. Meanwhile, the number of operating condition is also increased. The traditional soft sensor using a global model cannot satisfy the requirement of plant-wide process. The adaptive soft sensor based on Divide-and-Conquer strategy is getting more and more attention. In this paper, an adaptive Multi-State Partial Least Squares regression algorithm(MSPLS) and an adaptive Gaussian Mixture Model regression algorithm(GMM) are proposed as follows:In the plant-wide process with multiphase, the variation of process variable is much less than the change in operating conditions. Hence, a Multi-State Partial Least Squares regression algorithm is proposed using the variance change in process variables. Firstly, the variable with maximum variance is set as the state variable and the operating condition is divided into several states comparing the variance of process variables. Then, the variable’s mean of each state is calculated. After eliminating state difference, the PLS model can be built, which removes the influence caused by such conditions and the predictive accuracy of soft sensor is improved. Moreover, an adaptive mechanism using a moving-window is also proposed. Applications on CSTR model and the IPB concentration estimation in the distillation column validate the proposed algorithm.The plant-wide process also has non-phase property because of changes happened in operating condition. An adaptive Gaussian Mixture Model regression algorithm is presented using the Divide-and-Conquer strategy. Firstly, the system is divided into several states using the subtractive clustering algorithm. In each state, a Gaussian model is built and parameters are estimated using expectation-maximization algorithm. Then the global GMM model is obtained using the weighted strategy. Furthermore, a subtractive clustering algorithm is also introduced to reduce the computational burden and improve adaptive capability. At last, the numerical simulation case verified the performance of the proposed method.