| In process industries,the real-time monitoring of system status,the timely tracking of parameter quality and the accurate and smooth control of the system are important for maintaining safe and efficient industry operations.With the development of computing technology,traditional process industries are gradually moving towards intelligent manufacturing and information management,using soft measurement technologies to replace some hard-to-install sensors for the process monitoring,control and optimization of important processes.Process industries have problems such as complex and difficult-to-measure physical properties of system raw materials,coupled and redundant information,high-dimensional variables and low operational error tolerance.Such factors make it difficult for traditional soft measurement inference models to meet production requirements and severely limit the predictive effect of the models.Therefore,soft-sensing models based on neural networks are widely used in process industries.However,on one hand,redundancies often exist in both data and network structure;on another hand,as neural networks are mostly based on least squares loss,the abnormal data affects the generalization and prediction accuracy of the underlying model.To address these problems,this thesis designs the optimization algorithms of multilayer perceptron(MLP)soft-sensing model based on sparse optimization and robust estimation theories to neural networks to handle non-ideal industrial data.The main research elements of the thesis are as follows:(1)An optimization algorithm for MLP soft-sensing models with sparsity,unbiasedness and robustness(ARdLASSO-MLP)is proposed for complex nonlinear,coupled,redundant and anomalous data.Firstly,the MLP is combined with two-stage lp regularizations to achieve variable selection and structural optimization;secondly,an adaptive mechanism is designed using the maximum information coefficient and embedded in the lp regularizations to achieve unbiased estimation of the model;finally,Huber’s loss is introduced into the sparse optimization algorithm of MLP model to improve the robustness to vertical outliers.A mathematical simulation is able to verify the robust,adaptive,and sparse optimization performance of the proposed algorithm,thereby reducing the model complexity and improving generalization and prediction performance.(2)A globally robust sparse MLP algorithm(WARlp-MLP)based on improved Huber’s loss and two-stage lp regularization is proposed for data with nonlinear,coupling,redundancy and anomalies in both explanatory and response variables.Firstly,the weighted Huber’s loss is designed according to robust scale and embedded into ARdLASSO-MLP model to improve the robustness of Huber’s loss to outliers in explanatory variables.Secondly,a two-stage lp regularization is used to select significant variables and prune hidden layer neurons,so as to achieve sparse optimization of the model and improve the generalization performance of the model.Finally,a numerical simulation shows that the WARdlp-MLP model has better robustness,prediction accuracy and generalization performance for abnormal data in explanatory variable.(3)In this section,the superiority of the proposed models is verified through two modelling examples from real process industry and compared with the other state-of-the-art models.The prediction of octane values in a petrochemical oil cracking process with high-dimensional redundancy and only significant vertical outliers shows that the ARdLASSO-MLP model outperforms other algorithms in all evaluation criteria,indicating that the model is close to the optimal model and has higher prediction accuracy.The WARdlp-MLP model is shown to have better generalization and prediction accuracy,as well as better robustness,by predicting the copper grades from a copper ore metal flotation process with low-dimensional redundancy and both significant vertical and lateral outliers. |
