Data processing: generic control systems or specific application – Generic control system – apparatus or process – Optimization or adaptive control
Reexamination Certificate
2001-06-15
2004-03-02
Patel, Ramesh (Department: 2121)
Data processing: generic control systems or specific application
Generic control system, apparatus or process
Optimization or adaptive control
C700S028000, C700S048000, C700S045000, C706S047000, C706S048000
Reexamination Certificate
active
06701195
ABSTRACT:
BACKGROUND OF THE INVENTION
The invention concerns a method and an apparatus for a case-based reasoning (CBR) system, especially developed for the task of sensor value prediction within a cement kiln control system.
Providing accurate predictions of the cement kiln behavior for a limited period into the future, e.g. approx. 1 hour, can enable a human controller of the cement kiln to make more informed decisions, as well as providing a basis for more automated control within a cement making plant. The invention provides an alternative to existing technologies, such as rule-based control systems, that require prohibitively high installation and maintenance costs.
As part of an existing system which provides extensive support facilities for the control of a cement production plant, all sensor data for the cement kiln and related machinery are routinely stored within a database. The data is represented as time-stamped floating-point numbers.
As an example of the amount of data that needs to be processed in a sensor-based technical process; the sensor sampling rate inside a cement kiln is typically once a minute or more frequent, there are typically over 400 sensors in the cement kiln and related apparatus, and the data archive can contain in access of 1 year's storage of data. This means that the raw data can be of the order of 10
8
→10
9
floating point numbers. Therefore, any automated method that exploits this data to perform sensor value prediction needs to be able to cope with a large amount of unstructured sensor data.
The invention is most suited to technical processes that involve human intervention. Typically, the cement kiln in an active cement production plant is monitored and controlled by a human expert, roughly every 0-15 minutes. Due to the high numbers of sensors involved, it is difficult for a human expert to get an adequate over-view of the status of the kiln and, therefore, there is a need for automated support in the analysis task. In particular, when exceptional behavior occurs, e.g. sensor values going out of predefined ranges, or an abrupt change in sensor values, support is required to determine both what the likely/possible consequences are of the exceptional behavior and what corrective actions the human expert should carry out. In order to support this, an automated system is required that can accurately project the values of all sensors for a significant time period, e.g. >1 hour, into the future.
Nevertheless, the user may, at any one time, dynamically select a reduced subset of signature sensors that are considered to contain the most salient information to characterize the current state of the technical process. Hence, the automated prediction system must be flexible enough to react to this dynamic user selection.
The sensor data collected for a technical process can often be problematic. For example, due to the relative close proximity of many of the sensors within the cement kiln, there is significant redundancy in the information that is represented in the data stored for different sensors. Some level of random noise in the recorded data must also be tolerated. Perhaps more significantly, it cannot be guaranteed that all values for all sensors are always available. There are some periods of time where no sensor values are recorded, e.g. due to a failure in the database. More commonly, missing values will occur for a single sensor for a period of time, e.g. due to a failure in the sensor itself. These imperfections in the raw data must be tolerated by the prediction system.
The final complexity of the problem is that each application of the prediction system to a new technical process or feature thereof will require some recalibration. For example, each cement kiln has its own characteristics. Indeed, the set of sensors contained is likely to change from cement kiln to cement kiln. Hence, the sensor-value prediction system must be newly adapted to each cement plant in which it is installed; a costly procedure for any technique that is model-based. Furthermore, as for many other types of manufacturing apparatus, a single individual cement kiln is subject to aging. In other words, the behavioral characteristics of a single cement kiln are known to drift gradually over time. Hence, any behavioral model developed for an individual cement kiln must be periodically refitted to adapt to these changes; which is also a potentially costly maintenance problem.
Model-based techniques, in conjunction with Artificial Intelligence technology, such as Neural Networks and Fuzzy Logic, represent the state-of-the-art for automated control systems for cement plants. The main problem with this type of approach is that the general model of the technical system embedded in the prediction system must be adapted and parameterized by highly-skilled experts in order to be applied within a particular cement plant. In addition, due to drift in the behavior of a single cement kiln over time, the model needs to be periodically maintained, e.g. re-parameterized, so as to remain reliable over time. The disadvantages of high application and maintenance costs are likely to be encountered by an model-based technique.
A general alternative to hand-constructed and adapted models are machine learning techniques that can be trained on existing data. The most popular of such machine learning techniques that can be trained on existing data. The most popular of such machine learning approaches are artificial neural networks that have been successfully employed to perform diagnostic tasks based on sensor data in similar application fields to that of this invention. Nevertheless, some fundamental problems remain with artificial neural networks that serious prohibit their use for the cement kiln control application; including:
a Ability to deal with missing data: Some techniques exist for generation of missing sensor values, such as linear interporlation. Nevertheless, the degree of noise in the application data may hinder the training of artificial neural networks. Furthermore, it is not clear how an artificial neural network can deal with the dynamic selection of a subset of relevant sensors.
b Interpretation of results: The basis behind the predicted by a human controller results generated by an artificial neural network are not easily open to human inspection by a human expert. Hence, a control expert is unable to assess the reliability of the prediction. For this reason, neural networks are better suited to completely automated applications where human. inspection of the predictions is not required.
c Ability to predict exceptional behavior: A trained artificial neural network is generally good at recognizing the general trends that frequently re-occur within the training data but poor at reproducing rarely occurring, exceptional circumstances. Nevertheless, rare behavior is often the most important to predict with respect to the state of the art, the objects of the invention are; a new method and a new apparatus for process optimisation, especially in a cement kiln, based on the data produced by sensors.
EP 0 582 069 A2 discloses a method for control of a process having manipulated and controlled variables with the controlled variables having target values which depend on the adjusted value of said manipulated variables. The process is controlled in real time through a process controller under the operation of a computer. The method of control comprising the steps of establishing a first performance index to compute the absolute value of the deviation for each control variable in the process from its target value over a specified time horizon; generating a first linear programming model the solution of which minimizes said first performance index; solving the first linear programming model; establishing a second performance index to compute the absolute change in the value of each manipulated variable from its previous value for each control variable over a specified time interval; generating a second linear programming model the solution of which minimizes said second perf
Brown Michael
Heidemann Lueder
Schneider Karsten
Baker & Botts LLP
Hartman Jr. Ronald D
Patel Ramesh
Siemens Aktiengesellschaft
LandOfFree
Sensor prediction system utilizing case based reasoning does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Sensor prediction system utilizing case based reasoning, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Sensor prediction system utilizing case based reasoning will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3234971