Soft Sensor Modeling Based On Feature Extraction

With the rapid development of modern industrial technology,many industrial processes have more and more measurement requirements for some quality variables that affect the production.However,there are many key variables in the actual production process cannot be measured in time and effectively by sensors or other hardware measurement devices,and the measurement of these key variables has a great impact on the operation reliability and energy efficiency.As an effective method to solve the problem of online estimation of dominant variables that are difficult to directly measure in complex processes,soft-sensor technology plays an important role in improving the efficiency and product quality of industrial production,and has always been a research hotspot in the field of process control.In this paper,feature extraction and multi-model modeling in soft-sensor modeling research are studied.The existing feature extraction techniques are improved,and combined with clustering algorithm and integrated learning algorithm to increase the estimation accuracy of soft-sensor model and reduce the complexity of modeling.The soft-sensor technology is used to estimate some quality variables in a production process of bisphenol A in a factory.The specific research contents are as follows:(1)For the neighborhood graph construction problem in the isometric mapping algorithm,an adaptive neighborhood construction method is proposed.The method uses the Euclidean distance to calculate the sample similarity coefficient.Subsequently,the density exponential function is constructed based on the local density and average density of each sample,and the neighbor number of samples is adjusted adaptively according to the density exponential function so as to construct a reasonable neighborhood graph.This method was applied to the soft sensor modeling of the dehydration tower unite in a Bisphenol A production device.The adaptive neighborhood construction method was used to construct the neighborhood graph in the isometric mapping algorithm,and the Gaussian regression process was used to develop the model.The simulation results show that the estimation accuracy and generalization ability of the model are both improved.(2)Aiming at the poor robustness,poor topology stability of isometric mapping in manifold learning,a construction method of kernel matrix is proposed with constant translation which called Kernel Isometric mapping so as to improve the learning efficiency and topological stability of Isomap algorithm.In addition,the neighborhood graph construction method combing K-nearest neighbor algorithm and the ?-radius is applied to the kernel isometric mapping to develop the Gaussian process regression soft-sensor model.The simulation results show that the proposed method has higher estimation accuracy and learning efficiency compared with other methods of feature extraction soft-sensor modeling methods.(3)Due to the complex and changeable working conditions in chemical production process,a single soft sensor model cannot satisfy the requirements of estimation accuracy.A multi-manifold soft sensor modeling based on modified expanding search clustering algorithm is proposed.This algorithm uses the distance between manifolds instead of the Euclidean distance to adaptively determine the neighborhood radius,and introduces the local density to determine the center of clustering.The features of sub-manifold obtained after clustering are extracted by kernel isometric mapping method in manifold leaning respectively,and develop Gaussian process regression models.The simulation results show that the clustering effect and the estimation accuracy of the model are both improved.(4)In order to improve the model accuracy and generalization ability of complex chemical processes modeling,a multi-model soft-sensor modeling method based on kernel principal component analysis and modified stochastic gradient boosting algorithm is proposed.Aiming at the shortcomings of the traditional stochastic gradient boosting algorithm,the algorithm uses Gaussian process regression as the base learner,and uses the KPCA method to extract the features of the data.According to the feedback of the weak learning machine in each iteration learning,the learning rate is adaptively adjusted,and the weighting factor is introduced as the ranking index to prioritize the training data of under-learning,which improves the learning efficiency of the SGB algorithm and the estimation accuracy of the model.The simulation results show that the method not only has higher learning efficiency,but also improves the prediction accuracy.