Research Of Soft Sensor Modeling Based On K-means Clustering

Soft sensor technology is an important method to estimate the key variables which arehard to measure. It’s also a kind of novel testing technology which has vast developmentpotential. The limitation of traditional instrument detection has been broken and real-timeonline supervising has been made possible by its existence. Since practical industrialprocesses have characters such as complexity, multivariable, nonlinearity and time varying,the single soft sensor model is often far away from the ideal result. The process properties ofthe object can be better described by multi-model soft sensor models and furthermore theprediction accuracy as well as the generalization performance of the soft sensor model can beimproved. Necessary handling of the collected data is an important step before modeling. Inthis paper, the phenol content at the exit of dissolving tin in condensation reaction unit duringBisphenol A production process is taken as a background. The traditional K-means algorithmin data mining is stressfully researched and improved. Then the comparison among singlemodel, multi-model soft sensor models based on traditional K-means algorithm andmulti-model soft sensor models based on three improved algorithm are made andcorresponding conclusion are obtained. Three aspects of research are made in the followingparts:(1) Traditional K-means clustering algorithm has the advantage of briefness, speedinessand strong theoretical property. But clustering results of traditional K-means clusteringalgorithm easily fluctuates with random initializing cluster centers. At the same time, theclustering number K needs to be settled manually, which will directly affect the predictionaccuracy of soft sensor model if the value of K has not been accurately settled. Therefore, anovel K-means algorithm based on maximum distances product algorithm is promoted. A newtarget function is defined so as to do the optimization selection of the value of K. Theclustering accuracy is improved with the method and then a combination model based onsupport vector machine (SVM) is established. The simulation result shows that the predictionaccuracy of soft sensor model is improved the potential application in industry field isdemonstrated.(2) Traditional K-means clustering algorithm easily fall into local extremum and finallylead to poor prediction precision. In order to make up such shortage, the combination ofsimulated annealing (SA) and K-means clustering algorithm is promoted. The SA algorithmhelps traditional K-means algorithm jump out of local optimization and finally achieve globaloptimization. The new method is applied to a practical industry process in Bisphenol Aproduction for related quality index. From the result of a conclusion can be made that thechange trend of phenol content is tracked effectively, the complex industry process is adaptedand the generalization performance as well as the prediction precision of the soft sensormodel are improved.(3) Clustering assumes that each feature of the samples is the same to the contribution ofclustering, which is not corresponding to practical production. Considering the particularcontributions of different features, a feature-weighted K-means clustering algorithm is proposed. This algorithm adjusts feature weights gradually on the basis of the iterativeK-means clustering, and improves the clustering result finally. Gaussian process is arandomized model with less optimized parameters, quick learning rate and fine convergence.The combination of Gaussian process and feature-weighted K-means clustering algorithm ispromoted and the simulation result shows that the change trend of phenol content is trackedeffectively by the improved method which demonstrates the effectiveness and the potentialapplication in industry field.