Soft Sensor Theories And Methods For Complex Industrial Processes Based On Deep Learning

There are many key variables in industrial process,which are closely related to the production efficiency or quality.However,it is an enormous challenge to measure some key variables directly for some reasons such as economic issues or the lack of corresponding hardware sensors.In order to resolve this problem,soft sensors estimate the key variables indirectly by constructing a mathematical model.Some easily measured variables are used as the input,which appear to be strong relevant to the target variables,and the target variables are used as the output.Then,the key variables can be estimated by establishing the mathematical relationships with the input variables.Currently,studies on the soft sensors are mainly limited to the modeling methods.However,for the complex industrial processes with large structures,complex mechanisms and numerous parameters,the conventional modeling methods are not appropriate for the parameter modeling.Meanwhile,other relevant investigations on auxiliary modeling links are in the early stages.Considering these shortcomings,the present thesis intends to focus on the key issues such as difficulties in measuring information,establishing a model and implementing an effective control in complex industrial systems.To this end,the model structure and training method of soft sensor are considered as the research core and the deep learning method is applied to establish soft sensors for parameter estimation in complex industrial applications.Simultaneously,the research is also conducted in other key issues in the modeling process to resolve operational problems,including the low reliability and slow implementation of conventional auxiliary variable selection and the problem of insufficient data samples.It is worth noting that this study is of great significance for resolving key scientific problems in industrial process control,breaking through core technologies and promoting the development of the soft sensor technology.The main contents are as follows:1.Two data-driven soft sensors are proposed based on deep learning.By stacking support vector regression(SVR)on top of the deep neural network,a SAE-SVR model based on stacked autoencoders(SAE)and a DBN-IPSO-SVR model based on deep belief network(DBN)are established.During the training process,the loss functions are redefined and the optimization algorithms are also improved.These models are applied in a complex industry process for predicting the rotor thermal deformation of an air preheater in thermal power plant.Obtained results indicate that compared with other conventional methods,the training stability and prediction accuracy are significantly improved.2.Based on the SAE-SVR model,a knowledge-and-data-driven soft sensor,called L-SAE model,is further proposed.This model not only refers to the mechanism background and expert knowledge of industrial processes but also uses the field data to reflect the process information more comprehensively.The prediction results of the rotor thermal deformation of the air preheater under the same experimental background show that the L-SAE model has higher prediction accuracy and better model generalization ability compared with common knowledge-driven or data-driven methods.3.In the modeling process of soft sensor,the quantity or quality of the acquired samples are often very poor,which is a limitation for employing soft sensor technology in industrial measurement and control.A deep generative model VA-WGAN combining with generative adversarial network(GAN)and variational autoencoder(VAE)is proposed.By learning the distribution of the real data,new samples are generated to supplement the dataset for the soft sensor training.Obtained results show that the VA-WGAN model with a fast and smooth convergence process is a powerful method to solve difficult training problems of the original GAN.It is found that the artificial samples generated by VA-WGAN more closely resemble the real samples compared with other common generative models.4.In order to select the soft sensor auxiliary variables more effectively,a variable selection method based on the mutual information and prediction errors,called the MI-ME method is proposed.Meanwhile,a novel update method of the neural network weight is proposed to reduce the computational expenses.Obtained results indicate that compared with other commonly used variable selection methods,the variable set selected by the MI-ME contains the least auxiliary variables for the same soft sensor prediction errors.It is found that applying the variable set selected by the MI-ME as the input of the soft sensor can simplify the model structure and reduce the learning burden of the model while ensuring the prediction accuracy.