Data processing: artificial intelligence – Neural network
Reexamination Certificate
1998-08-13
2001-11-06
Davis, George B. (Department: 2122)
Data processing: artificial intelligence
Neural network
C706S023000, C706S907000
Reexamination Certificate
active
06314413
ABSTRACT:
BACKGROUND OF THE INVENTION
Field of the Invention
The invention relates to a method for controlling process events of a technical plant.
The method is suitable for the optimization and analysis of process events of a power station or a power plant.
It is generally known to detect individual process variables by measurement, to consider them and also to evaluate them as a function of the process state. In addition, it is known to model and to predict individual process variables by applying mathematical, statistical or neural algorithms. One disadvantage of these signal-supported methods is that in regions in which a large number of process signals are observed, the interpretability and comprehensibility—and therefore the current knowledge about the process state—is lost. When hundreds of process variables change simultaneously during transient process events, no estimate of the current process state can be obtained, and in particular it is impossible to assess the course of the transient process event.
With conventional methods, when non-linear relationships exist between the process variables, it is not possible to make any determination as to which process variables have to be changed simultaneously or what percentage change is suitable, in order to transfer from a current process state into a desired process state. A well-known technique for solving this problem is to carry out a what-if simulation for previously firmly defined steady-state operating regions, during which the influence of each individual process variable on the desired target variable is ascertained. One or more input signals can be changed, and the resulting behavior of the target variable can be calculated. A disadvantage of this method is that lengthy trial and error is required to obtain information as to which process variables have to be set and to which value a particular process variable must be set in order to move the process in a required or desired direction.
The transient process regions present a major problem because there is a lack of information about what can actually be viewed as a desired event in these regions. A combination of, for example, 200 measurement signals cannot readily be viewed to see whether it represents an optimum or a faulty process state.
SUMMARY OF THE INVENTION
It is accordingly an object of the invention to provide a method of optimizing and analyzing process events of a power station plant, which overcomes the hereinafore-mentioned disadvantages of the heretofore-known methods.
With the foregoing and other objects in view there is provided, in accordance with the invention, a method that permits a simultaneous and coherent assessment and indication of the relevant process variables of a technical plant. In accordance with the method, for each operating point at any particular time in time in the plant, information about the basic influencing variables is obtained, and process variable changes which are necessary in order to move from the current state into a desired operating state are prescribed. In addition, a visualization method enables an optimal process state to be distinguished from a non-optimal process state.
With the foregoing and other objects in view there is provided, in accordance with the invention, a method for controlling process events of a technical plant, which comprises: collecting or combining the variables relevant to a process; realizing a topology-maintaining nonlinear projection of data from the variables onto a multidimensional self-organizing neural map (SOM); evaluating the variables in relation to one another using the neural map; and controlling process events in accordance with the evaluation step.
In accordance with an added feature of the invention, the collecting step comprises collecting the variables into a vector.
In accordance with an additional feature of the invention, the technical plant is a power station plant or a portion of a power plant.
In accordance with another feature of the invention, a trajectory of process events displayed by connecting visualized process states graphically; and reference channels with adjustable tolerance widths are displayed on the map.
In accordance with a further feature of the invention, sequences of process states are evaluated using the reference channels.
In accordance with again an added feature of the invention, the process is steered through one of the reference channels.
In accordance with again an additional feature of the invention, the process is steered through one of the reference channels using a process steering means.
In accordance with again another feature of the invention, the method further comprises the steps of detecting if a trajectory deviates from the reference channel on the neural map; evaluating a winner rate neuron representing the state; and bringing the process back into the reference channel.
With the above and other objects in view there is also provided, in accordance with the invention, a method of controlling process events of a technical plant. The method comprises the steps of:
selecting process variables;
developing a self-organizing, neural network in a state space of a plant, on a basis of a self-organizing neural algorithm using process values of the process variables;
representing the self-organizing neural network as a neural map;
determining plant states from the process variables;
producing a reference channel of permissible process events by projecting the plant states onto the neural map;
recording the process events by linking the neural map to the process;
displaying a trajectory on the neural map;
monitoring whether the trajectory remains within or deviates from the reference channel; and
steering the process of the technical plant over the neural map in real terms in the state space.
In accordance with yet another feature of the invention, the plant states are projected onto the neural map with a winner-takes-all algorithm or with a winner-takes-most algorithm.
In accordance with a concomitant feature of the invention, the process values of the process variables are conditioned prior to the developing step.
The current process state is represented on this SOM and can thus be compared with other process states in context. If the current process state is remote from a desired process state, it is possible for the necessary process variable combination to be output in order to get from the current to the desired process state.
As a result of an integrated approach, values of the individual process variables and their mutual influences on one another are taken into account.
In neural theory, a self-organizing map (SOM) is understood to be a “self-organizing neural network”, in which all the neurons are arranged alongside one another. The self-organizing neural network is a term which has been introduced for a specific class of neural networks which structure themselves using input signals. See, A. Zell “Simulation Neuronaler Netze” [Simulation of Neural Networks], Addison-Wesley, 1994, pp. 179 to 187, which is herein fully incorporated by reference. As distinct from conventional neural networks, the spatial location of the individual neurons and their neighborhood relationships play an important role in the SOMs.
In system theory, the state space of a plant denotes the n-dimensional vector space in which the process data can be plotted one above another. For example, a data example or sample, that is to say a vector having the values of the n process variables at a time t
0
, corresponds to a point in this state space. In the following text, an input space will be understood to mean precisely this state space. The output space is the 2-dimensional space which is depicted on the map.
With the aid of the SOMs, it is possible for socalled topology-maintaining projections to be realized. In this context, topology-maintaining means that the points (data points) which are located close beside one another in the input space are also located close beside one another in the output space, that is to say on the map. The map therefore represents
ABB Patent GmbH
Davis George B.
Greenberg Laurence A.
Lerner Herbert L.
Stemer Werner H.
LandOfFree
Method for controlling process events using neural network does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for controlling process events using neural network, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for controlling process events using neural network will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2577435