AIAA
Publication September 1998
Paul Van Buskirk
Quality Monitoring and Control
Sensor
Validation
& System Verification
Abstract
Advances in AI modeling techniques, such as artificial neural networks, e-Model, genetic algorithms, VB-Model, etc., provide a robust tool for monitoring, control and optimization of a system. Monitoring is used for data verification, fault detection, forecasting and/or prediction. Total system or subsystem performance can be ranked when combined with statistical methods. When variables are adjustable, the system can be regulated to provide the “Very-Best” results through direct or fuzzy-logic control and optimization methods. With all of these methods, a bottom line criterion is the proper operation of sensors, such as temperature, pressure, flow, and etc. measurements. This paper provides a methodology utilizing data modeling techniques for the successful implementation of a sensor health management system.
Background
The key component in a sensor validation tool is the generation of a model that describes and predicts values of the measurement. By having a prediction of what the measurement should be and the actual data, a comparison can be made to determine possible errors. It should be agreed that if the model predicted value and the measured value are within accuracy requirements, then the potential of a sensor measurement fault is reduced. Concurrently, if a predicted value and a measured value do not agree, then at minimum, the sensor or the model should be evaluated for accuracy. Since modeling is a key component in this approach, modeling aspects using AI techniques are presented in the following section.
AI Model Development
Model development is an investment. Recent advances in PC and AI systems provide a significant reduction in this investment. And, there are numerous model applications that will produce a return-of-investment. This article gives an overview of model applications, available now, to improve the profitability, quality, safety or environmental controls of a process. Extensions of these applications can be made to any industry that has data.
This paper highlights neural networks as a model development tool, due to its widespread use and success. Other AI modeling methods can be used. The model developer should determine the best technique for a given application.
The advantage of AI modeling techniques, such as artificial neural networks (ANN), is that a considerable amount of time can be saved in model generation when data is available. Development of an equation based (fundamental) model for multivariable non-linear systems can be virtually impossible with certain processes. Furthermore, a fundamental model may require hundreds if not hundreds of thousands of equations. The number of calculations and resulting CPU times can exceed the actual system response in real time. Conversely, ANN models are computationally efficient. The code required typically does not exceed three pages, even for large systems.
ANN model development is data intensive. A database, or spreadsheet, contains the data that the ANN model uses. A column in the spreadsheet represents a single variable. A column can be used as an independent “input” or as a dependent “output” variable. How the output variable(s) change with changes in the input variables is what the neural net “learns”. A row of data across all used variables is an event the neural net will use to learn the cause-and-effect dependence. The ANN model then “trains” until the error between the predicted and actual value, across selected rows of data, is reduced to a satisfactory level. An introduction into how neural-networks work is provided in reference (1).
ANN model generation does not come without risk. The ANN model developer needs to follow certain data screening guidelines to ensure integrity in the data, as shown below:
General Recommendations for the AI Model Developer
-
All cause-and-effect variables are measurable and used.
-
Redundant variables are minimized or eliminated.
-
The variable data is uniformly distributed with sufficient points.
-
Data sampling times shows the actual system response.
-
The data is properly screened for outliers and redundancy.
-
True variability exists in the data and all data is valid.
-
Model accuracy requirements do not surpass measurement accuracy and variability.
-
Model prediction will not be extrapolated in any variable dimension.
The work involved with data screening can occupy the bulk of the time in ANN (or any other) model development. The ANN model software provider will also supply guidelines for software use, data screening, and model development. Numerous technical publications address issues related to ANN model data preparation.
When these guidelines are followed, an accurate and robust ANN model can be developed that can be used for multiple and strategic purposes. A general overview of these model applications is provided below:
AI Model Applications and Utilization
-
Sensor/data error and fault detection.
-
Prediction and forecasting.
-
Sensitivity or stability analysis.
-
Equipment or sub-system performance monitoring.
-
Total process or system indexing (comparison performance monitoring).
-
Control(s) monitoring.
-
Multivariable non-linear process/system control.
-
Multivariable non-linear process/system optimization.
The applications of a system model are not limited to the items as listed above. Endless possibilities exist. End use will be as varied as the personnel and industries that develop or utilize the model. The applications are available now, through numerous software providers. Implementation will provide a technology advantage, with benefits that improve profit, quality, environmental and safety controls.
AI Model Application and Utilization
Applications of an ANN model can range from sensor & data verification to forecasting to full process optimization. In most cases the required ANN model(s) for these applications can be developed from the same database. For a given process or system, these applications should be used together.
This approach provides constant monitoring of the sensors, equipment, controls, and total system performance. Deterioration in the performance of any component in a system can then be quickly pinpointed and corrected. As a result, the stability of the process will be improved, providing an increased utilization of the control and optimization applications. Maximum utilization of the control and optimization applications will increase the process profit and quality. Additionally, safety and environmental goals will be maintained and improved.
Utilization of ANN models for sensor validation and fault detection is in common use. A reliable fault detection technique that uses several AI modeling tools is described in reference (2). This application is most effective when all sensors are monitored as part of a preventative maintenance program. The keys to successful process control and optimization implementation are reliable and accurate sensors. An example of a sensor validation and error detection application is shown as follows:
Sensor Error Meter
This application shows a typical sensor monitoring system. The lists on the left are the various sensor names. The box graphs show the sensor errors, in percentage. The top graph shows the current error, the bottom graph shows the average error. Only the top four sensors with the maximum errors are given in this example.
Sensor Error Meter Technologies
There are numerous approaches in detecting faulty sensor measurements. These include common sense, calibration, statistical and the modeling approach as previously provided. The approaches that are the most rigorous include a combination of statistical and modeling procedures.
From the model there is an associated model prediction error. Simultaneously, there is an associated measurement error. As a sensor declines in accuracy, there is an associated increase in the predicted and actual data value. Using the one, two and three sigma limits of the accuracy of the model and simultaneously, the percent error in measurement as listed by the vendor, a comparison can be made as to the probably accuracy of the sensor. This includes both a biased error and random error. This approach is both intuitive and has been used with success within the process industries. The procedure is given below:
-
Develop an AI model of the sensor that utilizes cause and effect relationships within this system. In general, it is best to not include the sensor under consideration as an input to the model. The model input should be independent of the actual sensor under consideration.
-
Using actual operating conditions or new data, compare the predicted value from the AI model with the actual measured value. The data error produced will show a bias or noise in the difference calculation. In other words, a constant negative or positive value will indicate a bias. A fluctuating sign of the difference indicates a noisy sensor.
-
Delta errors that exceed the vendor or calibrated accuracy of the sensor indicate a potential problem. Model accuracy then comes into play.
-
Errors between the predicted and actual values that are below the one sigma error band are considered to be the normal variation of a system including both sensor accuracy and model accuracy.
-
Errors that exceed a three-sigma band indicate a concern with 99% probability that either the model or the sensor itself is in error. Model error can be investigated based on extrapolation or interpolation considerations as well as utilization of student t-test analysis based on localized data probabilities.
In other words, if the model has shown to be suspect with similar operating conditions or the model was not developed with similar data, then the probability of model error is high. However, if the model has shown robust prediction with similar data, then the actual sensor measurement is then suspect.
-
Errors that are within the one and three sigma limits require additional analysis that have been developed but are beyond the scope of this article. The basic methodology with these approaches includes a technique termed error detection and data reconciliation as given in reference (6) and (7).
Conclusions
A method for determining gross errors and sensor accuracy has been presented. The method is simple, understandable yet robust in detecting significant or minimal errors in measurement systems. Both statistical and sensor modeling techniques are combined to provide indications in both sensor bias and / or noise levels. This method has been used in industrial applications with success.
Paul Van Buskirk is a Principal with Quality Monitoring & Control. He may be reached at 281.359.4471 or via email at QMC_Plus@MSN.com.
References
1. Bhagat, P., “An Introduction to Neural Nets.”, Chem. Eng. Progress, p 55 (Aug 1990)
2. Smith, S., “SDI’s e: Real Time Prediction in the Chemical Industry”, PC AI, p 18 (Jan/Feb 1998)
3. Chitra, S., “Use Neural Networks for Problem Solving”, Chem. Eng. Progress, p 45 (April 1993)
4. Van Buskirk, J., “Modeling of a High-Density Polyethylene (HDPE) Process”, Texas A&M University, Masters Thesis (Dec 1996)
5. Ogunnaike, B. and W. H. Ray, Process Dynamics, Modeling, and Control , Oxford University Press (1994)
6. Serth, R.W., Heenan, W.A., "Gross Error Detection and Data Reconciliation in Steam-Metering Systems", AIChE Journal, p 733 (May 1986)
7. Heenan, W.A., Serth, R.W., "Detecting Errors in Process Data", Chemical Engineering, p 99 (Nov 86)
8. Van Buskirk, P. D., "Neural Networks for Monitoring, Control and Optimization", PC AI, p 28 (May/June 1998)