Learning process for a neural network

Details Learning process for a neural network Learning process for a neural network

1996-07-26

2004-06-01

Davis, George B. (Department: 2121)

Data processing: artificial intelligence

Neural network

Learning task

C706S025000

Reexamination Certificate

active

06745169

ABSTRACT:

FIELD OF THE INVENTION
The present invention relates to a learning procedure for a neural network for open-loop or closed-loop control of a process with time-variable parameters.
BACKGROUND INFORMATION
Neural networks must be first trained with learning or training data before they become capable of generalizing. Acquiring the training data is often time-consuming and expensive.
An example is the neural network described in German Patent Application No. 44 16 364, which calculates a process parameter from a multiplicity of previously calculated input parameters supplied to it, which serves to preset a system controlling an industrial process. Thus, for example, in a rolling process, a predictive value for the rolling force is calculated as a function of the rolled stock temperature, thickness reduction, and other material-specific and system-specific input parameters. The relationship between the rolling force and the input parameters, simulated by the neural network, is adjusted on-line after each process run, i.e., after each pass of the rolled stock, to the actual process events. For this purpose, the input parameters measured during the process run and subsequently recalculated, and the rolling force are combined at a data point, which is then used for the adjustment of the neural network parameters. The adjustment is done with each new data point, i.e., on-line. The adjustment must be characterized by special stability, since it is often carried out directly and without supervision by a person skilled in the art on the system executing the process. Therefore, only non-critical parameters of the neural network are adjusted in on-line training, using adjustment algorithms and network structures, thus ensuring stability for the procedure; i.e., minimization of the square of the error function between the network response and the recalculated rolling force, with the error function preferably having only one global minimum, and no local minima.
In order for a conventional neural network to predict, at least approximately, reasonable rolling forces from the beginning of the on-line training, the neural network can be pre-trained using a rolling force model that predicts the rolling force as a function of randomly preselected input parameters. If such a model is not available, the preliminary knowledge required for pre-training can be acquired by collecting training data, for example, on comparable systems and inputting them into the neural network.
An object of the present invention is to provide an open-loop or close-loop control for a process with time-variable parameters to ensure high quality. It is desirable, especially in the case of starting up a new system or after radical changes in an existing system which is controlled using neural networks, to enable the neural network to exhibit reasonable behavior, after just a few data points, directly within the system without pre-training. The same is true when an existing system is remodeled and data cannot be collected beforehand. Furthermore, long-term system drifts must be recognized and compensated for.
SUMMARY OF THE INVENTION
This object is achieved according to the present invention by a learning process for a neural network for open-loop or closed-loop control of a (mainly industrial) process with time-variable parameters with the neural network being configured in at least two embodiments. The first embodiment uses an open-loop or closed-loop control network with which the process is controlled and which is trained with the current process data so that it builds a model of the current process. The second embodiment uses a background network, which is trained during the operation with representative process data so that it builds an averaged model of the process over a longer period of time, with the open-loop or closed-loop control network-being replaced, after a-certain learning time on the part of the background network or on the basis of an external event by the background network.
Another embodiment of the present invention includes a learning process in which the learning during the process run is subdivided into two phases: an initial phase and an operating learning procedure. The amount of training data with which the neural network is trained grows in the initial phase and remains basically constant during the operating phase.


REFERENCES:
patent: 5033006 (1991-07-01), Ishizuka
patent: 5175678 (1992-12-01), Frerichs
patent: 5353207 (1994-10-01), Keeler
patent: 5408586 (1995-04-01), Skeirik
patent: 5479573 (1995-12-01), Keeler
patent: 5486996 (1996-01-01), Samad
patent: 5513097 (1996-04-01), Gramckow
patent: 5566275 (1996-10-01), Kano
patent: 5600758 (1997-02-01), Broese
patent: 5608842 (1997-03-01), Broese
patent: 5625552 (1997-04-01), Mathur
patent: 44 16 364 (1994-11-01), None
Wang, “A self growing learning algorithm for determining the appropriate number of hidden units,” 1991 Intl conf on neural networks (IEEE), Nov. 1991.*
Hirose, “Back propagation algorithm which varies the number of hidden units,” Neural Networks v4 n1, Dec. 1991.*
Mozer, “Skeletonization: a technique for trimming the fat from a network via relevance assessment,” University of colorado department of computer science & institute of cognitive science, p107, Dec. 1991.

Affiliated with

Also associated with

Rating

Learning process for a 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 Learning process for a neural network, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Learning process for a neural network will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3326177