Multi-condition Soft Sensor Modeling Of Transfer Learning Based On Domain Adaptation

With the increasing requirements for control,monitoring and operational reliability of industrial processes,real-time monitoring of key variables has become particularly important.However,due to the influence of factors such as process mechanism,physical environment and instrument hardware characteristics,some process parameters are difficult to directly measure with sensors,so process monitoring and automatic control cannot be carried out.The soft measurement method has become an effective solution to the above-mentioned problems.It uses the idea of indirect measurement to model through the collected auxiliary process information to realize online estimation of dominant variables.At present,soft measurement methods can be divided into two categories: modeling methods based on process mechanism analysis and data-driven modeling methods.Process mechanism models are easily affected by many factors such as changes in the application environment.At the same time,there are many disturbance factors in the actual industrial process,such as nonlinearity,time-varying and large hysteresis,making it difficult to obtain measured values in real time.The data-driven modeling method relies on the internal connection of the data collected in the process,so it does not need to have a deep understanding of the research object and is suitable for modeling applications in the process industry.However,in the actual production process,due to the reorganization of the system equipment,changes in raw materials and the external environment,etc.,the production conditions will change significantly,making the system present the characteristics of multiple working conditions and multiple modes,resulting in real-time data between working conditions Mismatch with the modeled data distribution.And because of the lack of actual sensor data,it is impossible to obtain effective mark values for modeling,so it is difficult to establish an accurate soft-sensing model after the working conditions change.In order to solve the problem of soft sensor modeling under multiple operating conditions,this paper introduces the transfer learning method to the field of soft measurement,and obtains the available information of the unmodeled operating conditions from the modeled operating conditions to achieve the purpose of improving the performance of the soft sensor model.The main research contents are summarized as follows:(1)Due to the problem that the performance of the original measurement model deteriorates due to the non-linear and time-varying characteristics of the batch process,a feature migration method based on the geodesic stream core is studied.Firstly,the data is mapped to the manifold space to complete the feature migration from the historical working condition to the current working condition;and the real-time data and modeling data are used for distribution adaptation to reduce the distribution difference between the two;finally,the distribution adaptation is used The historical data after preparation establishes a regression model to predict the concentration of penicillin between different batches in the process of penicillin fermentation,and realizes unsupervised batch soft-sensing modeling.The experimental results show that the model can improve the prediction accuracy more effectively than traditional soft-sensing methods.(2)Aiming at the problem of model inaccuracy caused by sudden changes in operating conditions during actual industrial operation,a multi-source domain adaptive soft-sensing method of random weight neural network parameter migration is studied.First,use the maximum mean difference measurement value of the historical main operating condition and multiple historical auxiliary operating conditions to build the model to obtain the output weight parameters of its hidden layer;then,share the parameters to the models of the historical main operating conditions and the current operating conditions;Finally,a related alignment regular term is introduced to constrain the relevance of the hidden layer output,thereby reducing the data distribution difference between the main operating condition and the current operating condition.This method uses the information of multiple source domains to enhance the robustness of the model,and avoids the influence of the labeled data of some target domains in the domain adaptive random weight neural network on the model through parameter migration,and achieves the purpose of unsupervised soft measurement.(3)Aiming at the above methods ignoring the guiding role of the label to the model and the inability to maintain the mapping relationship between the label and the feature in the domain adaptation process,a multi-source domain unsupervised software based on joint distribution alignment and mapping structure maintenance is studied.Method of measurement.Specifically,first use the hypergraph to describe the complex structure of the features and the label,and then use the hypergraph matrix as the two view matrices for multi-view clustering to construct category pseudo-labels;then use dynamic distribution alignment to compare historical conditions and The current operating condition data is adapted to the edge distribution and conditional distribution,and the hypergraph Laplacian operator is introduced to perform manifold regularization constraints to prevent the structure between data features and labels from being destroyed;finally,similar operating conditions are introduced to further Enhance model robustness.Ball mill,TE and CSTR experiments show that this method can effectively improve the performance of the soft-sensing model.