Multi-target Regression Based On Stochastic Configuration Networks

In recent years,driven by the Made in China 2025 strategy,traditional industries have accelerated their upgrading and transformation,and industrial processes have changed from testing parameters such as pressure and flow to testing parameters related to production efficiency and product quality.However,in the actual industry,these parameters are often difficult to measure,and usually need to use process data and rely on soft-sensing technique to implement online estimation.Currently,soft-sensing techniques based on various neural network models have emerged,among which stochastic configuration networks are one of the typical representatives,which are gradually used in industrial processes due to their ability to quickly and autonomously determine the network structure and parameters based on data during training.There are often coupling properties among difficult metrics in industrial processes,so their soft-sensing can be essentially recognized as a multi-target regression problem.Since the traditional stochastic configuration networks are a single-hidden layer feedforward neural network,it only considers the influence of the input on the output（target）and ignores the correlation between the outputs,leading to its performance degradation in multi-target data analysis tasks.To this end,this thesis focuses on stochastic configuration networks with the goal of improving multi-target prediction capability and investigates multi-target learning methods,and the main work and innovations are as follows:（1）A multi-target learning method for stochastic configuration network based on L_2,1-norm structure matrix is proposed.Firstly,the correlation between the targets is handled from the global method,the structure matrix term is added to improve the multi-target prediction performance of the model;then the supervision mechanism is improved to ensure the approximation properties of the model;finally,an update strategy is used to resolve the structure matrix and the output weights alternately.（2）A Two-stage Learning Method to Multi-target Regression for Stochastic Configuration Network is proposed.The first-stage learning method models the complex relationships of inputs and outputs,and the second-stage learning method uses the prediction information of the first-stage learning method to learn the correlation between targets through the structure matrix,thus improving the prediction performance.And the structure matrix and output weights are decoupled,thus greatly improving the modeling efficiency.（3）The two improved multi-target stochastic configuration network algorithms proposed in this thesis are applied to a typical industrial example to establish a crude oil Atmospheric and vacuum distillation quality parameter estimation model,and experimental simulations are carried out,and the results show that both of the two stochastic configuration network multi-target learning methods can achieve the requirements for quality improvement.In summary,this thesis focuses on the study of multi-target modeling theory based on stochastic configuration networks and applies it to a typical industrial process.Based on the benchmark data and the actual data of industrial processes,it is proved that the method developed in this thesis has good predictive performance in multi-target application scenarios and has certain theoretical research and practical application value.