Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
2000-06-08
2004-07-06
Picard, Leo (Department: 2121)
Data processing: artificial intelligence
Neural network
Learning task
C706S014000
Reexamination Certificate
active
06760716
ABSTRACT:
BACKGROUND OF THE INVENTION
The present invention relates generally to the use of predictive models in process control systems and, more particularly, to adaptively adjusting the output of a predictive model using process parameter measurements.
DESCRIPTION OF RELATED ART
In many areas of the process control industry, it is common that a critical parameter such as the composition of a fluid or gas flow may be difficult to measure or may only be measurable using off-line techniques. Often times, such parameters are obtained by off-line lab analysis or by the use of an on-line analyzer that is difficult to maintain on a continuous basis or is unreliable. When an estimate of the critical parameter is needed for some reason, such as to perform process control, any of a number of various mathematical techniques have been used to create a process model that predicts the hard-to-measure process parameter based on process inputs or parameters that may be easily measured or ascertained during process operation.
More particularly, it is generally known to use one or more predictive models such as a non-parametric model (e.g., neural network models), a natural model (e.g., first principle models) or a parametric model (obtained by regression techniques) within process control systems to estimate process parameters. One use of a predictive model is as a virtual sensor which can be used to estimate or predict the output of a fictitious sensor at some point within a process control system. As indicated above, a virtual sensor is typically used when it is difficult or impossible to locate a real sensor at the desired process location. A virtual sensor can also be used to predict process parameters for which no real sensors exist, e.g., to measure parameters such as smell or taste that require a judgement on the part of an operator to determine. Another use of a predictive model is as a cross analyzer to check the correct operation of a sensor within a process. For example, if the cross analyzer and the sensor produce different enough measurements, it may be an indication that the sensor has failed in some manner.
As is known, neural network predictive models are constructed from sets of training data, wherein each set of training data has a value for each of a number of process inputs/outputs (which are the inputs to the model) and a value of the process parameter being predicted or estimated. After being developed from the training data, the neural network model is then used within the process control system environment to predict the process parameter, even though an actual sensor is taking no or only limited measurements of the parameter during process operation. First principle or natural models typically use one or more mathematical equations to model the behavior of the phenomena being predicted.
Predictive models when used to predict the value of a process control variable are subject to various types of errors, such as bias errors, drift errors, non-linear errors, etc. Bias errors are errors in which the output of the model has a bias or offset with respect to the value of the actual process parameter being predicted. Bias errors can be mathematically represented in the form of y=x+B, wherein y is the predicted or modeled process parameter value, x is the actual process parameter value and B is the bias. Drift errors are errors that are linear or are proportional to the value of the process parameter being modeled and can be represented in the form of y=Ax, wherein A is the drift error multiplier. Non-linear errors are errors which are non-linear with respect to the actual process parameter and can be mathematically represented in the form of y=f(x), meaning that the predicted value y is a non-linear function of the actual value x. Non-linear errors occur for any number of reasons but often result from the fact that the predictive model is inadvertently created to have some non-linear error component therein. For example, a natural model may not account for higher order (second, third, etc.) effects of the phenomena being modeled, a neural network model may be created from an incomplete or inadequate set of training data, etc., resulting in the output of the model having some non-linear error component. Besides inadequacies in the model, these errors may also be caused by changes in the process equipment or unmeasured process inputs, such as feed stock composition, that may directly impact the process output. Because unmeasured process inputs and changes in process operation are common within the process control industry, the application of neural networks, first principal models, or other commonly available predictive models have had limited commercial success in providing a reliable prediction of critical process outputs.
In the past, the only way to correct for linear or non-linear errors caused by mismatch between the developed model and the actual process was to reform the model using different or more accurate equations, using a different or more complete set of training data, etc. The process of recreating or reforming a predictive model is time consuming, usually requires an expert, can require enormous amounts of processing power and time and cannot typically be done in real time process operation. Still further, there is no guarantee that the prediction of the newly developed predictive model will be free of errors. Thus, instead of adaptively correcting the output of a predictive model used in, for example, a process control system, the predictive capability of a predictive model has been periodically checked by taking actual measurements of the process parameter being predicted and comparing these measurements with the output of the predictive model. One or more of such comparisons were typically used to determine if the model prediction had errors therein. However, the comparisons were not used to adapt the model. Instead, these comparisons were used to determine if the model produces an output that is tolerable and, if not, to reform the model using conventional model forming techniques.
SUMMARY OF THE INVENTION
The present invention is directed to adaptively modifying or adjusting the output of a typical predictive model, such as a neural network model or a natural model, using actual process measurements to thereby reduce or correct for non-linear as well as linear errors, without having to reform the predictive model itself. Such an adaptive predictive model can be easily implemented within a process on a continuous basis during operation of a process control system with only minor increases in the processing as compared with standard predictive models. Generally speaking, the adaptive predictive model uses measured process inputs in conjunction with process parameter measurements obtained by, for example, a lab sample or analyzer, on a continuous, a periodic or even a non-regular or non-periodic basis to adapt the output of a standard predictive model. Using this feedback mechanism, the predicted value of the process parameter may be automatically adapted to compensate for unmeasured disturbances and changes in process operation.
According to one aspect of the present invention, an adaptive predictive model for use in estimating a process parameter includes a predictive model configured to receive one or more model inputs and to produce a predicted value of the process parameter based on the one or more model inputs and a measurement input configured to receive a measurement of the process parameter. The adaptive predictive model also includes a combiner network coupled to the measurement input to receive the process parameter measurement and coupled to the predictive model to receive the predicted value of the process parameter. The combiner network combines the process parameter measurement with the predicted value of the process parameter to produce an adjusted predicted value of the process parameter.
The combiner network may include a correction block that determines a correction factor and a summer that sums the correction factor to
Blevins Terrence L.
Ganesamoorthi Sai
Mehta Ashish
Tzovla Vasiliki
Wojsznis Wilhelm K.
Fisher-Rosemount Systems Inc.
Holmes Michael B.
Marshall & Gerstein & Borun LLP
Picard Leo
LandOfFree
Adaptive predictive model in a process control system does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Adaptive predictive model in a process control system, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Adaptive predictive model in a process control system will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3223924