Research Of Multiple Working Conditions Soft Sensor Modeling Based On Geodesic Flow Kernel

Process industry is an important part of manufacturing industry and an important pillar industry of national economic in China.How to achieve optimal control of the process industry has always been a key research area in academia and industry.But the physical and chemical reactions and mechanisms involved in the production process are complex,which makes it difficult to detect key parameters closely related to process control in real time.With the development of science and technology,soft sensor technology has attracted widespread attention.By selecting easy-to-measure variables related to hard-to-measure variables closely,soft sensor technology constructs mathematical model between easy-to-measure variables and hard-to-measure variables to realize the estimation of hard-to-measure variables.As the diversification and customization of the enterprise products,equipment reorganization,production raw material changes and parameter setting adjustment are common in the industrial production process.These operations make the system exhibit multiple working conditions characteristics.It means that the real-time data and the data used for modeling may no longer obey the same probability distribution.In this case,the application of traditional soft sensor conditionls in the process industr may meet challenges to some degree.To solve the problem of soft sensor modeling under multi working conditions,this study introduces transfer learning strategies based on geodesic flow kernel into the field of soft sensor.The study attempts to achieve the purpose of improving the performance of the soft-sensing model by obtaining the available information for the unmodeled condition from the modeled condition.The main research contents are summarized as follows:(1)For the difficulty of adapting to the changes of data distribution with traditional soft sensor modeling,this study aims to introduce transfer learning strategies based on geodesic flow kernel into the soft sensor field.In order to realize the method can deal with the situation where the marginal distribution and the conditional distribution are different,the study establishes label compensation mechanism.This method selects one of the multi-modeled condition as the initial modeled condition,and the remaining modeled condition as the historical modeled condition with true labels.The common subspace and the specific subspace are selected from the initial modeled condition and the historical modeled condition by subspace disagreement measure.The geodesic flow kernel is used to transfer the common subspace of the initial modeled condition to each historical modeled condition to obtain the label deviation values of all historical modeled condition.After the label compensation model is constructed by the specific subspace of the historical modeled condition and the label deviation value,the model is used to predict the label distribution deviation of the unmodeled condition to achieve reliable prediction of key parameters.(2)For the dynamic characteristics of actual industrial processes and the problem that data distribution contains non-Gaussian information and Gaussian information,this study introduces dynamic independent component analysis and dynamic principal component analysis into the geodesic flow kernel and extracts data non-Gaussian information and Gaussian information,respectively.From the perspective of ensemble learning,the final dynamic soft sensor model is fused through the model based on non-Gaussian information and Gaussian information.(3)For subspace disagreement measure does not consider the multiple condition data variation connectionand in label compensation mechanism,and there are dynamic characteristics in the actual industrial process,the joint and individual variation explained extraction algorithm is introduced to extract the joint common information and specific information of the multi-modeled condition and the unmodeled condition.To cope with the dynamic characteristics of the process at the same time,this method constructs an augmented matrix of modeled and unmodeled conditions before extracting information.And under the framework of geodesic flow kernel and label compensation mechanism,the label compensation model is established by using the extracted specific information of the multi-modeled condition and label deviation value,so as to realize dynamic soft sensor model which can effectively deal with the different marginal distribution and conditional distribution.