This is a planned extension of the existing Cybernetica CENIT that will add cognition to the application.
The goals of the extension are to:
Both generic functionality and application specific in the form of a new model interface will be added. The cognitive extension may either extend the current estimator (digital twin) or it may replace it entirely.
Cybernetica CENIT already implements adaption in the form of parameter estimation.
In addition we would like develop and implement methods for real-time and offline analysis of the estimator (digital twin) performance related to process data.
In this way it should be possible to automatically classify types of errors: sensor failure, input error or model error. Ultimately, the goal will be to suggest model improvements based on this analysis.
The performance of the MPC system will eventually degrade over time due to changing plant conditions, i.e. increased plant-model-mismatch. To prevent poor controller performance, the error/performance degradation first must be detected. There are currently no self-diagnosing capabilities in CENIT. An important part of Cognitive CENIT will be the ability to perform self-diagnosis and detect when the controller performance is unsatisfactory. In the case where unacceptable levels of control performance degradation has been found, further action (such as error classification and error correction) is needed.
Estimators are generally unable to distinguish between prediction deviations resulting from the following errors:
Being able to distinguish between these errors is important because the required response is very different:
In the case of input error, the appropriate response is some combination of correcting the faulty input signal and minimizing the faulty signal’s impact on the model-predictive control.
This can include:
In the case of model error, the appropriate response is to try to adapt the model to most accurately reproduce the process data.
An important goal for Cognitive CENIT will be to distinguish between these cases based on an offline training of a classification algorithm.
Situations where the model structure is incomplete or wrong may be identified using an automated analysis of the prediction error distributions. Currently Cybernetica CENIT estimators assume that the model structure is correct, and that the prediction error is normally distributed around a mean value, which the estimator tries to center at zero. In many cases this is not true, and significant deviation from normally distributed error may imply error in the model structure. Identifying this error is non-trivial and may be a well-suited task for an AI extension.
Cybernetica CENIT: TRL 9
Cybernetica Cognitive CENIT: TRL 1-2