Research On Multi-operating Chemical Process Data Rectifiction Based On Expectation Maximization

In today’s intelligent process of industrial development,data driven plays an important role in process system control,online real-time monitoring and determination of production decision.However,since the measurement data contains a small number of outliers,and the accuracy of data determines the effectiveness of data-driven method,measurement data generally cannot be directly used in data-driven processes.Data reconciliation technology（DR）is an effective method to ensure the reliability of process data.It can deal with the abnormal data caused by factors such as sensor failures,equipment problems,and production process characteristics in the production process,so as to obtain reliable process data.However,most of the current researches on data reconciliation focus on single stable working condition or dynamic process,and there are few researches on data reconciliation of multi-operating condition process caused by the change of raw material batch or the change of production strategy.This topic studies this field,proposes an improved multi-condition data correction strategy based on historical conditions,and uses the method of expectation maximization to solve the process of condition recognition and data correction.The specific research content are as follows:（1）Aiming at the shortcomings of the multi-cycle data correction model proposed by Alighardashi et al.,such as the difficulty in selecting parameters and the need for multiple calculations,this paper proposes to use data clustering technology to identify the process historical working conditions and obtain the model parameters from the historical working conditions,so as to provide accurate parameters for the multi-cycle data correction model without the need for multiple calculations.（2）Aiming at the shortcomings of the commonly used data clustering methods in the recognition of historical conditions（such as difficult to determine the number of clustering）,an advanced clustering algorithm in the field of artificial intelligence combined with gaussian mixture model was proposed to extract historical multiple conditions from process historical data.（3）To solve the problem that abnormal data,defect data,dynamic data and transition state data in the process historical data have a great influence on the clustering analysis results（namely,the results of condition identification）,the smooth filtering method is used to de-noise the process data,and the steady-state detection method is used to eliminate the unsteady and transition state data.3σ rule and box graph theory are used to identify abnormal data in steady state.The method of linear extrapolation and time series model is used to predict and replace the defect data and abnormal data,thus laying a good foundation for the identification of historical conditions.This topic starts from the process historical data of the actual device,and verifies the proposed multimode process data through four aspects of process data cleaning,steady-state detection,historical multimode identification and multimode process data reconciliation.Correction of the effectiveness of the improvement method.Accurate process data can provide guarantee for operations such as industrial process modeling and process optimization analysis,and lay a reliable data foundation for the construction of smart factories.