Patent Application
20050137995
The present invention relates to a method for regulating a thermodynamic process, in which process variables in the system are measured, predictions are calculated in a neural network on the basis of a trained, current process model and compared with optimization objectives, and actions suitable for regulating the process are carried out in the system.
In the case of a known method of the type described above in the Technical Field section, process variables that are difficult or expensive to measure are predicted by means of the process model in the neural network. To be able to follow changes of the process, three steps are carried out in a cycle, that is a process analysis to find a starting point for the process model, training of the neural network, and application of the process model for the prediction. This procedure is time-consuming and labor-intensive.
The present invention is based on the object of providing improvements with regard to regulating a thermodynamic process, with the regulating being of the type in which process variables in the system are measured, predictions are calculated in a neural network on the basis of a trained, current process model and compared with optimization objectives, and actions suitable for regulating the process are carried out in the system. In accordance with one aspect of the present invention, at the same time as the regulating described in the immediately preceding sentence, the process is automatically analyzed and at least one new process model is formed, trained and compared with the current process model with respect to the predictions.
The fact that the process is automatically analyzed and at least one new process model is formed, trained and compared with the current process model with respect to predictions at the same time as normal regulating operation is in progress allows an adaptation of the model to a changed process to be achieved without increased expenditure on personnel. This completely automatic model adaptation preferably runs in the background, i.e. as a so-called batch job on the data-processing system, as opposed to running in the foreground, so that the expenditure of time is also no greater. A number of new process models with, for example, different topologies of the neural network and different numbers of training cycles allow an adaptation even to great changes of the process to be achieved.
Taking place in a cement kiln, as an example of a thermodynamic process, is a combustion process which is to be regulated in such a way that it has, on the one hand, a certain stability and, on the other hand, a certain plasticity, i.e. it adapts itself to the conditions, with certain optimization objectives having been set. The state in the cement kiln is described by various process variables, such as for example lime mass flow, air mass flow, or the like, some of which at the same time form manipulated variables. The state in the cement kiln is changed by actions, i.e. changes of manipulated variables. For online monitoring and regulation and predictions of future states of the cement kiln, a neural network is implemented on a data-processing system. The neural network defines a process model which indicates the change in the state as a reaction to actions and is independent of the optimization objectives. A quality function is used to perform a situation assessment, which assesses a specific, current state while taking the optimization objectives into consideration.
To be able to predict specific process variables, for example the FCaO value (which is also known as the clinker index and is a conventional measure of the quality of cement), to define the quality of the cement, in the case of a known method: first a process analysis is carried out in order to identify a function to determine the desired process variable, and then training of the neural network is performed with the process model based on the data obtained and finally the neural network is applied.
According to the present invention, on the other hand, a model adaptation is performed fully automatically in the background. For this purpose, first an automatic process analysis is carried out, providing a list of all the relevant process variables by means of methods of process identification (e.g., preferably various methods of process identification) in defined time cycles.
On this basis, various types of neural networks with various parameter constellations, such as learning rates and training cycles, number of layers, size of layers and other aspects of topology, parameters of the data processing (low-pass filter sizes or the like) are trained in automatic modeling and are verified on the respectively available database. The search for suitable network parameters can be realized in the high-dimensional parameter space by suitable optimization methods and search strategies (for example evolutionary methods).
If a process model which is better, i.e. works more accurately, than the model currently being used is found by the analysis and modeling, this new process model is used from then on.
This model adaptation provides an automatic adaptation to changing process properties of the respective plant, including major interventions, such as alterations or conversions, so that an adequate process model is ensured. Previously unconsidered process variables are also included if need be in the modeling.
| Number | Date | Country | Kind |
|---|---|---|---|
| 02 018 426.3 | Aug 2002 | EP | regional |
The present application is a continuation of International Application PCT/EP2003/008599, which was filed Aug. 2, 2003, designates the U.S., and is incorporated herein by reference, in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/EP03/08599 | Aug 2003 | US |
| Child | 11058111 | Feb 2005 | US |
