Abstract:
For industrial processes with high dimensions and multiple modes, we propose a multi-model soft sensor modeling method based on an improved affinity propagation (AP) clustering algorithm. First, we apply the principal component analysis and differential evolution method to improving the performance of the traditional AP algorithm and to removing the influence of redundant information. The most accurate sub-datasets are obtained based on the optimized parameters. Second, we use Gaussian process regression to construct local models. Finally, we utilize the prediction variance of the new data to calculate the posterior probability, and obtain the prediction result through a combination of different local models. The effectiveness of the proposed algorithm is verified through simulation results of two benchmark datasets and a sewage treatment process.