(125d) Integrated Operation Support System (Iopss): the Data Pre-Processing and Data Reconciliation Modules

During the past two decades, process engineers have been working in the development of simpler yet powerful techniques to condition, analyze, and further use of plant data. It is so that methods for gross error and outlier detection, data reconciliation, parameter estimation, and optimization were developed. First, these methods were focused on only one activity at a time. Later, they evolved into methods attacking two or more activities at a time (e.g. simultaneous gross error detection and data reconciliation, parameter estimation and data reconciliation, optimization and parameter estimation). In the same way, the complexity of the problems being solved has been increasing: from steady-state, linear and offline systems to dynamic, nonlinear and online systems. This last kind of problems is today's research objectives in the matter. Their solution is making use of traditional and newer statistical techniques. Unfortunately, for the process engineer in the industry, these techniques represent a far if not an inaccessible tool for their real/practical necessities since their availability to use at the plant is limited. Moreover, these methods do not offer the opportunity to verify the consistency of the data, and at the same time to perform all the other activities in a single, friendly environment.

In a previous work (Rolandi and Romagnoli, 2005) we have proposed an innovative approach for simplifying the problem formulation by incorporating the Problem Definition Environment (PDE) to support current developments in Open Simulation/Optimization Architectures by the computer aided processes engineering community. While offering a user-friendly environment for problem formulation, it gives the opportunity to perform related model-based activities such as simulation, parameter estimation, data reconciliation and optimization using a single model representation.

In this work we discuss our current developments, within this novel environment, in the modules corresponding to the data pre-processing and dynamic data reconciliation activities. In terms of data-preprocessing, we will discuss the development and implementation, within the proposed framework, of an approach based on the Mean Minimun Distance (MMD) for the detection and median replacement for the rectification of outliers. Based on cluster theory, this method provides the tools for detecting outliers in a multivariable system. It takes into consideration that at a particular time, the values of the variables are interrelated (i.e. by the mass and energy balances), so the final result is the location of the time where the outlier occurs, although it does not differentiate which variable(s) is contributing to the presence of the outlier. Furthermore, the extension of the MMD method with median replacement applied to individual variables was considered, allowing the method to detect not only the time where the outlier is present, but also the individual variables contributing to the outlier. These approaches were compared with the modified MT-filter method. This method, developed by Liu et al. (2004), is a modification of the original Martin and Thomson filter (1982), which is based on the previously developed Kalman filter. It allows the rectification of outliers and, at the same time, the reduction of the noise present in the plant data.

Regarding the dynamic data reconciliation module, the error-in-variable method (EVM) was implemented as an important contribution to the environment. In contrast to the ?traditional? data reconciliation, in which errors are considered to be present only in the output variables, in the EVM errors are present in both input and output variables. In this work, a sequential solution method was used in the gPROMS software, so that the errors and parameters are calculated at the same time (i.e. simultaneous parameter estimation and data reconciliation).

The ?traditional? data reconciliation is performed in gPROMS by means of the ESTIMATION entity. In this entity, the objective function is automatically set by the software, and the experimental data is taken from an EXPERIMENT entity. On the contrary, the OPTIMIZATION entity does not permit the use of the EXPERIMENT entity, and the objective function is not automatically defined by the software. This work proposes a new approach based on the creation of an OPTIMIZATION entity, instead of the usual ESTIMATION entity for the data reconciliation activity. Therefore, a different way to incorporate the actual values of the measured variables into the least square objective function was developed. It was also necessary to define the objective function in the system model; moreover, the model had to be adjusted and re-defined so that it contained all the structures for a dynamic EVM, and was still useful for all the other model-based activities involved in the environment: simulation, parameter estimation, and optimization.

Finally, the pre-processed data was use to evaluate the performance of the different outlier detection/cleaning methods in the dynamic EVM data reconciliation.

References Liu, H.; Shah, S. and Jiang, W. On-line outlier detection and data cleaning. Comp. Chem. Eng. 28 (2004), 1635-1647.

Martin, D.R. and Thomson, D.J. Robust resistant spectrum estimation. Proceedings of the IEEE, 70 (1982), 1097-1114.

Rolandi, P.A. Model-based framework for integrated simulation, optimization and control of process systems. Doctoral Thesis. Department of Chemical Engineering, The Univeristy of Sydney, Australia, 2004.