Soft-sensing Modeling Method Based On Clustering Algorithm Of Multi-operating System

The operating points of practical industrial process often change to productdifferent products, which leading to that the industrial production samples are clusteredaround different operating points. In such circumstance, the prediction accuracy andgeneralization performance of single soft-sensing method are unsatisfactory. In order toimprove the performance of multi-operating system, a soft-sensing method based onclustering algorithm is proposed. Its basic idea is to classify the training samples into severalclasses using clustering algorithm, and train the sub-models according to correspondingsub-class samples. The global model which integrated sub-models can cover all operatingpoints. The test samples are assigned to appropriate sub-class, then predicted outputs areestimated by corresponding sub-models. This paper mainly studies soft-sensing method formulti-operating system, aiming at improve clustering algorithms, so as to further improve theestimation accuracy and generalization ability of the global soft-sensing method.In order to use the properties of the samples collected from multi-operating system, asoft-sensing method based on Simulated annealing (SA) algorithm and Least square supportvector machine (LSSVM) is presented. Classify the training samples into several classes bythe improved Kernel fuzzy c-means clustering (KFCM) based on SA to solve theshortcomings of KFCM clustering that it is easy to fall into local optimum and sensitive to theinitial cluster centers. Then, the sub-models are trained by LSSVM according tocorresponding sub-class samples. Finally, the predicted outputs of the test samples areestimated by corresponding sub-models after they are clustered by similarity measurement.The simulation results of Melt Index indicate that the presented method has better predictionaccuracy and generalization performance.A soft-sensing method of multi-operating system based on Adaptive Affinity PropagationClustering Algorithm (adAP) and Least Square Support Vector Machine (LSSVM) ispresented. The adAP clustering can classify the samples without knowing the initial clusteringcenter and the clustering number. Classify the training samples into several classes using theadAP clustering to find the best clustering result, and train the sub-models by LSSVM according to corresponding sub-class samples. The test samples are assigned to appropriatesub-class, then predicted outputs are estimated by corresponding sub-models. The simulationresults of Melt Index indicate that the proposed method has better prediction accuracy andgeneralization performance.