Data Reconciliation as a Framework for Chemical Processes Optimization and Control

Abstract

This thesis presents, discusses and compares a set of methodologies and several appropriate combinations of them, to provide accurate estimation of process variables, either for steady-state or dynamic systems. <br/><br/>Firstly, the accuracy of estimated measurements is improved through the proposal of novel Data Reconciliation techniques. The proposal combines data-based and model-based filtering and also consider the presence of time-delays between sampled data.<br/><br/>Secondly, measuring network design and its optimal use are addressed. Thus, the measuring device number, their type and their location for optimum reliability and accuracy of measurement at lowest possible cost are determined.<br/><br/>The first part of this thesis provides procedures for accuracy estimation in dynamic evolving processes. These procedures rely on combining data-based filtering and model-based filtering. <br/><br/>One technique combines a Moving Average filter and a steady-state Data Reconciliation technique sequentially. The resulting estimator presents the important statistic feature of being unbiased. <br/><br/>Additionally, this estimator provides high accuracy estimation and good tracking for dramatic dynamic changes of process variables, when compared with other techniques. The other technique performs a wavelet analysis as a former step for reconciling dynamic systems. The wavelet technique catches or extracts the process measurement trends that are later made consistent with the dynamic process model. As a consequence of this technique high estimation accuracy is provided. Additional advantages of applying this technique over the current techniques are the easy handling of distinct sample times and evaluating the variance of dynamic variables. Furthermore, this thesis addresses an important aspect regarding dynamic Data Reconciliation: how to improve the accuracy estimation when the process is faced with the presence of time-delay. This problem was overcome in a simple and efficient way by proposing a time-delay estimation method that works in conjunction with the Measurement Model adopted within the Data Reconciliation technique.<br/><br/>The presented time-delay estimation method determines the existing delay by maximizing the correlation of the process variables using genetic algorithms.<br/><br/>The second part of this thesis addresses the design of sensor networks, the proposed strategy allows the optimal selection and placement of measuring devices. The proposal deal with different sensor placement aspects: variation in design, retrofit, hardware redundancy and available sensor type.<br/><br/>The sensor placement procedure was extended to deal with dynamic systems by taking advantages of dynamic variable classification and dynamic Data Reconciliation.<br/><br/>The procedure to locate sensors in dynamic systems aims at maximizing the performance of Kalman filtering using accuracy as its main performance index. To accomplish this, both the measurement noise and the observation matrices are manipulated.<br/> <br/>The solution strategy has been implemented in academic and in the Tennessee Eastman challenge problems showing promising results. The resulting optimization problem was solved satisfactorily either by exhaustive search or using genetic algorithms based optimization. The profile of the relative increase of the system performance along the sensor network and the associate investment cost gives the designer all the alternatives for making an adequate decision.<br/><br/>Additionally, reliability is considered by combining quantitative process knowledge and fault tree analysis, providing an efficient way to improve its evaluation. It is important to state that the possibility to use inferential sensors based in an Artificial Neural Network model instead of physical sensors, and their incorporation within reliability and reconciliation procedures was a paramount consideration throughout this work.<br/><br/>Finally, this thesis also provides two frameworks, one for sensor placement and the second for Data Reconciliation. Both proposed frameworks have been designed, specified and validated following the guidelines of the new standards and trends in developing component-based application (e.g. UMLTM, CAPE-OPEN). These frameworks can include the above mentioned algorithms and can be extended to include other existing or futures approaches efficiently.