Study On Data Reconciliation And Gross Error Detection Methods And Applications

Process data play a vital role in industrial processes, which are the basis for monitoring, process control, optimization and business decision making. However, it is inevitable that process data measurements will be corrupted by random errors even gross errors. Therefore, data reconciliation and gross error detection techniques have been developed to improve accuracy of process data by reducing the effect of random and gross errors. In this dissertation, taken the process industry as the background, several solutions and an application in plant-wide mass balance system about data reconciliation and gross error detection are discussed.First of all, the deficiency of the existed frameworks of data reconciliation and gross error detection approaches are analysed. Then a novel framework and an approach are proposed. With the sequence of the development from linear steady systems to nonlinear dynamic systems, the proposed framework and the approach are improved step by step so as to cover several aspects of the data reconciliation and gross error detection problems. As a result, a complete theory framework of data reconciliation and gross error detection is constituted. At the last, the proposed simultaneous data reconciliation and measurement bias, leak estimation approach within the theory framework has been applied in a practical plant-wide mass balance system. Meanwhile, the framework and the approach are verified through the practical application to prove the effectiveness.The main contributions in this dissertation are listed as follows:1. A novel recursive strategy using mixed integer non-linear program (MINLP) for linear systems is presented based on the Akaike information criterion (AIC) to identify and estimate gross errors. In the approach a recursive strategy is used to reduce the computational burdens of the original approach, and a statistical test is applied to improve the performance of gross error detection. Simulation results show that the proposed strategy possesses better performance of identifying gross errors and costs less time than the MINLP approach. Although it is usually considered that using the complete form of the AIC to address data reconciliation and gross error detection problem will provide superior performance, the simulation results show that the performance is unsatisfied. Therefore, a novel framework with superior performances for data reconciliation and gross error detection should be proposed.2. To address the problem arisen from the previous research, instead of the AIC, the statistic learning theory is introduced as a framework of data reconciliation and gross error detection. Furthermore, a support vector regression (SV regression) approach is proposed for simultaneous data reconciliation and gross error or outlier detection, which considers gross errors and outliers as model complexity so as to remove them. For data reconciliation, the SV regression approach is robust to outliers in data sets, because its impact function is bounded. Moreover, it is not so strict to tune the coefficients of the SV regression approach because of the robustness of the coefficients for the reconciled results. Finally, a number of process and literature system simulation results show the features of the SV regression approach proposed.3. Gross errors can be divided into two categories, namely measurement related such as malfunctioning sensors and process related such as process leaks, and the latter one could also deteriorate the reconciled results. However, few approaches could handle process leaks. In the statistic learning theory framework of data reconciliation and gross error detection, the SV regression approach is extended to simultaneously detect and estimate measurement biases and process leaks so as to make the proposed framework and the approach be able to cover both kinds of gross error detection problems. The simulation and the comparision results indicate the effectivness of the approach.4. For nonlinear dynamic systems, the augmented unscented recursive nonlinear dynamic data reconciliation (AURNDDR) algorithm is proposed, which concerns uncertainties of states, measurements and parameters in estimation so as to provide accurate estimates. Furthermore, the SV regression approach is applied in nonlinear dynamic systems to establish the SV regression AURNDDR (SVRAURNDDR) approach, which can overcome the effect of outliers to simultaneously provide accurate estimates of states and parameters. Meanwhile, with the SV regression approach located in the control loop, improved close-loop control performances could be obtained even when outliers are presented.5. In any modern petrochemical plant, the plant-wide data rendering the real conditions of manufacturing is the key to the operation managements such as production planning, production scheduling and performance analysis. Because of the characteristic of data reconciliation and gross error detection, it is quite suitable to address plant-wide mass balance problem using data reconciliation and gross error detection techniques. For practical applications, a plant-wide mass balance system architecture compliant to the international standard of Manufacture Operation Management (MOM) is formed. Based on the architecture, the proposed SV regression approach for simultaneous data reconciliation and measurement bias, leak estimation is applied to a plant-wide mass balance system so as to make the system be able to simultaneously address measurement biases and material movement report mistakes effectively. The system has been applied in a refinery belonging to the SINOPEC and a chemical plant in Dalian. Moreover, the system is popularized to the following projects. The practical applications show that the SV regression approach and the architecture proposed are effective and standard, and they are easy to be applied in practice.