Research And Application Of Data-Driven Soft Sensor Modeling For Complex Industrial Processes

In recent years,mechanism-driven modeling approaches have become very difficult,due to the continuous innovation in production technology,the increasing scale of the process industry and the growing complexity of the process technology.In order to realize timely and efficient monitoring and control of industrial processes,it is essential to measure the key quality indicator.However,due to the harsh measurement environment and high measurement cost,direct measurement of the key quality indicator becomes very diffcult.To solve this problem,soft sensor modeling technology is used to establish a mathematical model between some easy-to-measure auxiliary variables and difficult-to-measure key quality indicator,and to predict the estimation of key quality indicator through easy-to-measure auxiliary variables,thus providing important assurance information for the control and operation optimization of production process.In the actual production process,the industrial processes and data often exhibit one or more characteristics such as high complexity,strong nonlinearity,large scale,dynamics,and non-completeness,which will seriously affect the soft sensor model performance.In this thesis,we take process industry as the research object,analyze the complex characteristics of process data,and carry out a series of data-driven soft sensor modeling research.The research works are mainly shown below:(1)A novel soft sensing data augmentation method for target-relevant autoencoder based on noise injection is presented to solve the problem of data scarcity in industrial process modeling.Firstly,the target-relevant autoencoder is used to extract the feature representation of process data,which can ensure the consistency of the feature space in the input and output data.Then,to increase the diversity of generated samples,virtual samples can be generated at the output layer of the target-relevant autoencoder by randomly injecting gaussian noise into the extracted feature representation,and the generated virtual samples are added to train the soft sensor model to improve the model performance caused by the problem of data scarcity.Finally,the effectiveness and superiority of the presented method can be verified by a numerical simulation and an ethylene industrial case.The experimental results show that that data augmentation technique can not only effectively alleviate the problem of process modeling data scarcity,but also significantly improve the generalization performance of the soft sensor model.(2)A novel soft sensor modeling method using whale optimization algorithm and kernel regularization-based functional link neural network is presented to deal with the problem of multicollinearity and strong nonlinearity of process data.Firstly,the original input variables are nonlinearly mapped to high-dimensional space to obtain the high-dimensional feature representation,and the high-dimensional feature representation and the key quality variable are used to establish the regression model by using regularization method.Then,the kernel technique is introduced to reduce the computational complexity and improve modeling accuracy,and an enhanced whale optimization algorithm is employed to search for the best parameter sets of the soft sensor model.Finally,a numerical simulation and a real-world industrial case called purified terephthalic acid solvent system are performed to verify the model performance.According to the experimental analysis,the proposed model has advantages in dealing with multicollinearity nonlinear process modeling,not only with high model accuracy,but also with fast convergence speed.(3)A novel soft sensor modeling method based on deep layer-extended target-relevant autoencoder is presented for modeling nonlinear industrial process with multiple characteristics such as high complexity and large scale.Firstly,to address the problem of weakened implicit feature representation caused by the accumulation of reconstruction errors in the hierarchical stacking training of stacked autoencoder,a layer-extended target-relevant autoencoder model is proposed.The layer-extended target-relevant autoencoder can not only ensure the consistency in the feature space of the input and output variables,but also effectively restrain the influence of reconstruction error accumulation during the model training by the introduction of extended layer.Then,a deep layer-extended target-relevant autoencoder is designed to minimize the influence of the model reconstruction error accumulation through the joint training of the extended layer and the original input,so that the extracted deep feature representation can better express the original input information,and the original input and the extracted implicit feature representation can jointly participate in the soft sensor model.Finally,a numerical simulation case and a debutanizer column industrial case are carried out to validate the effectiveness and superiority of the proposed method,and the experimental results show that the proposed soft sensor model has superior performance and high modeling accuracy.(4)A novel ensemble learning soft sensor modeling method based on singular value decomposition for echo state network is presented for modeling nonlinear industrial process with dynamic characteristic.Firstly,using the echo state network as the base model for dynamic soft sensor modeling,in order to solve the problems of redundant information and ill-conditioned solution of linear regression in soft sensor modeling,the singular value decomposition method is designed to compress the feature information of the reservoir while ensuring minimal loss of the reservoir feature information,which can eliminate the redundancy of the reservoir information.Meanwhile,in the process of establishing the regression model,the singular value decomposition based linear regression method can be effective in overcoming the problem of ill-conditioned solution caused by the traditional linear regression method.Then,to improve the stability of the model,a dynamic window bootstrap method is designed.In order to ensure the dynamic characteristic of the process data,the original input data can be dynamically sampled in a small range to increase the diversity of the training data,and an ensemble learning approach is used to establish the soft sensor model.Finally,through an SRU case and a high-density polyethylene industrial case,it is verified that the proposed method has good modeling performance,low computational effort,high information space utilization and good stability.