Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
1997-12-02
2001-11-20
Davis, George B. (Department: 2122)
Data processing: artificial intelligence
Neural network
Learning task
C706S014000, C706S016000, C706S020000, C706S906000, C706S907000
Reexamination Certificate
active
06321216
ABSTRACT:
BACKGROUND OF THE INVENTION
Field of the Invention
The invention relates to a method for analysis and display of transient process events in a technical plant.
The method is suitable for analyzing and displaying process states and events in a power plant.
It is well known for individual process variables to be detected by measurement technology, observed and also evaluated as a function of a process state. However, better assessment of a technical process is possible by simultaneous, holistic observation of all relevant process variables.
SUMMARY OF THE INVENTION
It is accordingly an object of the invention to provide a method for analysis and display of transient process events, which overcomes the hereinafore-mentioned disadvantages of the heretofore-known methods of this general type and which enables simultaneous, coherent assessment and display of relevant process states and sequences of process states in a technical plant. In particular, it should become possible to analyze and display transient process events. An expansion of the method for diagnosing transient events is also to be disclosed.
With the foregoing and other objects in view there is provided, in accordance with the invention, a method for analyzing and displaying process variables, process states or sequences of process states in a technical plant, which comprises combining and evaluating all variables relevant to a process in relationship to one another by a neural analysis on a basis of Kohonen maps, by performing a topology-producing projection of data of relevant process variables on a two-dimensional Kohonen map.
In accordance with another mode of the invention, the technical plant is a power plant or part of a power plant.
In accordance with a further mode of the invention, there is provided a method which comprises additionally displaying sequences of process states detected as a reference course, to assess current sequences of process states on a previously statically visualized two-dimensional Kohonen map.
In accordance with an added mode of the invention, the detected process states are different load states.
In accordance with an additional mode of the invention, there is provided a method which comprises displaying a trajectory of process events by a graphic connection of visualized process states.
In accordance with yet another mode of the invention, there is provided a method which comprises displaying a plurality of trajectories one above the other, for permitting a comparison of process courses.
In accordance with yet a further mode of the invention, there is provided a method which comprises permitting a manual process analysis, once a user has selected map portions having the greatest deviation between two trajectories, by the following additional steps: calculating a standard deviation of a distribution of weight of selected neurons, and outputting weight indices having the greatest deviations from one another; and ascertaining process variables on which the weight indices of the neurons are based and which cause the most trajectory deviations on the map.
In accordance with yet an added mode of the invention, there is provided a method which comprises performing an automated diagnosis by the following additional steps: automatically detecting deviations between two trajectory courses on a process map with the aid of a supervisor map, by comparing trajectory portions to be examined on-line, including a predeterminable number of trajectory positions, with reference portions, and automatically performing following steps in the event of deviations by a predeterminable value; calculating standard deviations in weight distributions for neurons in which the trajectories differ by the predeterminable value, and ascertaining weight indices having the greatest deviations from one another; and ascertaining the process variables on which the weight indices of the neurons are based and which cause the most trajectory deviations on the map.
In accordance with yet an additional mode of the invention, there is provided a method which comprises: making a prognosis of process variables by the following additional steps: graphically extrapolating a trajectory course at a first time within a predeterminable time segment for ascertaining a process state to be expected at a second time, and making the extrapolation on the basis of a further course of a reference trajectory; ascertaining and outputting the process variables and process values belonging to the respective process state to be prognosticated; and furnishing the prognosticated process values for selected process variables for downstream closed-loop and open-loop control systems.
In accordance with again another mode of the invention, there is provided a method which comprises displaying a distribution of a selected process variable above the Kohonen map with height or color coding, by plotting a weight distribution corresponding to the process variable as a height or color above the map.
In accordance with a concomitant mode of the invention, there is provided a method which comprises displaying a distribution of a plurality of selected process variables above the Kohonen map, and ascertaining and displaying differences in weight between one neuron and its neighboring neurons of selected process variables in a color-coded manner.
Another principle for attaining the object would be a method based on primary component analysis, that is a mathematical method for dimensional reduction, which seeks to achieve the best possible description of a data distribution from a high-dimensional space within a low-dimensional space. In primary component analysis or variant analysis, that is achieved through the use of a linear projection in a space, which is defined by the intrinsic vectors of the data distribution. However, that linear principle involves restrictions, and primary component analysis is therefore not believed to be a satisfactory way to attain the stated object.
General feasibility of the method is assured even for the most difficult data distributions by using nonlinear, neural methods in accordance with the invention and the method is not subject to any linear restrictions.
Through the use of the holistic principle, not only the values of the individual process variables but also precisely their mutual effects on one another are taken into account.
In terms of the method, a projection onto nonlinear surfaces, so-called primary multiplicities, is performed. These nonlinear surfaces are defined by so-called topology-producing Kohonen maps in the state space of the plant.
In neural theory, such a map is understood to be a “self-organizing neural network”, in which all of the neurons are disposed side by side. The term self-organizing neural network is a coined term for a special class of neural networks, which structure themselves on the basis of input signals, as is seen in the publication by A. Zell, entitled “Simulation Neuronaler Netze” [Simulation of Neural Networks], Addison-Wesley Verlag, 1994, pp. 179-187. Unlike conventional neural networks, in Kohonen maps the spatial location of the individual neurons plays an important role.
The process state calculated by the method is plotted on this topology-producing map of potential process states and visualized and is also comparable directly with other process states, such as previous process states. Therefore, in principle the map is a topology-producing, two-dimensional window into the n-dimensional state space of the plant. In the context of this description, “topology-producing” means that the points which are located close together in the input space are also located close together in the output space, or in other words on the map.
In the method of the invention, after suitable data processing, the relevant n process variables are offered to a self-organizing network in a learning phase. The learning phase proceeds in two stages: First, the map forms in the state space of the plant, and afterward, the plant states are visualized by employing a mathematical method.
The formation of t
Goser Karl
Otte Ralf
Rappenecker Gerd
ABB Patent GmbH
Davis George B.
Greenberg Laurence A.
Lerner Herbert L.
Stemer Werner H.
LandOfFree
Method for analysis and display of transient process events... 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 analysis and display of transient process events..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for analysis and display of transient process events... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2598442