Quality Monitoring & Control
"What do
you want to solve today?"
e Model for Excellence in Prediction Accuracy!
This amazing tool will provide actual equations of your process! e Model, developed by System Dynamics International (SDI), is a highly advanced Evolutionary Algorithm. Extensive evaluations were performed by QMC to find the highest quality program available. Our summary reveals why this is an extraordinary program!
Utilization of real-time prediction techniques of critical measurements in the Chemical Industry is sometimes suspect for accuracy. This is due to the multitude of operating conditions that may exist, limitations in mathematical modeling techniques, and missing or bad independent variable measurements. Procedures for sensor validation and flow meter error detection have been developed that address the concern for bad or faulty measurements. However, the ability to verify model accuracy and robustness over a wide range of operating conditions remains a significant concern.
A comparison has been made for the prediction accuracy of three modeling approaches. These modeling approaches include neural net, non-linear regression, and e Model. Data was based on an actual polymerization process. Polymerization processes are known to be non-linear. The modeling basis included a multi-variable approach with seven independent variables (such as, flows, pressures and temperatures) and one dependent variable, a polymer property. Over 2,000 data rows were used in this evaluation with proper outlier screening and applied time response delays. This data is noisy and the dependent variable has an estimated measurement error of five to ten percent. Steady state results are given in the following table:
| e Model | Neural Net | Non-Linear | |
Average Relative Error % |
8.8 |
15.2 |
11.9 |
Standard Deviation |
6.8 |
19.6 |
11.2 |
The results for this industrial application show the e Model to provide the minimum relative error and error variation when compared to the neural net and non-linear regression methods. Additionally, the lowered variation (standard deviation) implies improved noise smoothing when compared to the other modeling approaches.
The following shows a graphical display of the e Model interface, where e is used to model a polymer process. This data is based on a design of experiments statistical approach, where there are seven independent process variables and the truth (dependent) variable is a polymer property termed the melt index (MI).
To meet the needs in the Process Industries for critical measurement verification, a redundant approach was utilized. This approach uses sensor and flow meter verification for the independent variables, and multiple modeling techniques to validate critical or key process measurements. The multiple modeling techniques include neural nets, multivariable parameter estimation, and e Model. A voting protocol can then be established that compares the relative predictions of all model estimates to validate the actual critical measurement. Predicted and actual measurement accuracy’s provide statistical ranges as part of the voting protocol algorithm for critical measurement fault detection.
The following shows the sensor fault detection screen, emulating a real-time application:
In this application, the critical measurement is a polymer property. The measurement is provided by laboratory analysis, and subject to analytical or calibration errors. The models used include neural net, a regression technique, and e Model. The model predictions are based on several independent value measurements (i.e., inputs), filtered through a sensor validation technique.
Operating and polymer sales guidelines require narrow limits on polymer property variations. Since all models show that the polymer property did not significantly change, the polymer plant is within compliance with sales requirements, and the laboratory measurement should be reevaluated for accuracy.
Key to this approach is selection of modeling techniques that provide independent estimates of these critical measurements. Evaluations of final model accuracy and error deviation have shown the e Model method to be superior when the fundamental equations of a process are unknown, which is normally the case in the Chemical Industries. The e Model provides a fundamental class of operators and functions which reflect the actual process. Both accuracy and robustness improvements are a consequence of this flexibility, when compared to other techniques, which are usually fixed in their mathematical form.
Visit the web site of e at System Dynamics International (SDI). This high technology engineering services firm has designed this advanced Evolutionary Algorithm Tool to solve very difficult problems with software for the engineering and scientific industries.