Research Of Online Multi-model Modeling Based On Locally Weighted PLS Algorithm

In the chemical production process, there are some process variables closely related to the quality of products, they reflect the important features and index of actual production process, and must be monitored and controlled. Due to limitations of the detection or costs, usually without appropriate online instrument to measure the process variables directly and rapidly, so the establishment of soft sensor model with high precision becomes very important. For the highly nonlinear, dynamic and multiple operating points of industrial process, the general nonlinear modeling method may appears mismatch of model characteristic, worse extrapolative performance and so on. To describe the characteristics of this process, a multi-model modeling method is employed. On the basis of Bayesian classifier and locally weighted partial least squares(LWPLS) algorithm, the research of modeling algorithm is carried out, which via the expectation maximization(EM) algorithm to identify the LWPLS parameters with the idea of semi-supervised learning(SSL).For the characteristics of high nonlinearity, time-varying and multiple operating modes in the Tennessee Eastman(TE) process, dividing the training data from the TE process into different sub-databases and assigning the new coming data by Bayesian classifier. In the following step, it employs LWPLS to establish the online local models using the idea of just-in-time(JIT). The output estimation is the combination of each local model with respect to the posterior probabilities. The simulation results show the effectiveness of the algorithm and high prediction accuracy.As small proportion of labeled data obtains from industrial process, a semi-supervised LWPLS online soft sensor based on EM algorithm is proposed. Firstly, labeled and unlabeled historical data are accumulated to construct training database. For the newly coming query data, the corresponding similarity index is computed towards each sample points in the database, which acts as importance weight at the same time. Then the semi-supervised LWPLS model is constructed, and an EM algorithm is employed to estimate the parameters of the timely updated semi-supervised LWPLS model. Online prediction finally achieves simulation results of debutanizer distillation processes, which suggests that the proposed method has good prediction accuracy and stable generalization performance.Under the framework of semi-supervised learning theory, the multi-model modeling algorithm of semi-supervised LWPLS is studied by combining multiple operating points characteristics of the production process. Taking the data identity and principal component vector of sub models as latent variables of EM algorithm, local semi-supervised LWPLS model parameter is identified. The simulation results of debutanizer distillation processes show that the method has good modeling performance, which has practical reference value in handling multi-model modeling problem by small proportion of labeled data in industrial process with multiple operating points.