DE19527521C1 - Neural network training method - Google Patents

Abstract

Neural networks are trained using training data before they can obtain their generalisation function. The neural network is trained using continuously running on-line adaptation in combination with a cyclically repeated adaptation of network parameters. The on-line adaptation is limited to a selected parameter of the neural network. The on-line adaptation is provided for each new training data point, while the cyclically repeated adaptation occurs every time, depending on the amount of training data accumulated for the preceding data points.

Description

Neural networks need to be ability, first with learning or training data be trained. Collecting this training data is often lengthy and associated with great effort.

An example of this is that known from DE-OS 44 16 364 neural network that fed from a variety of it predicted input variables as a network response Process parameters calculated, which is used to preset a Sy serves to regulate a technical process. So will e.g. B. in a rolling process a predictive value for the rolling force depending on the rolling stock temperature, the thickness acceptance and other material- and system-specific gears are calculated. The one from the neural network correlation between the rolling force and the input size Eating is done online after each process sequence, i.e. after each Roll pass, adapted to the real process. To the measured during the process and then ß recalculated input variables and the rolling force in egg nem data point summarized, which then for the adaptation of Pa parameters of the neural network is used. The adap tion takes place with each newly determined data point, i.e. on- line. The adaptation must go through a special stability distinguish themselves because they are often direct and without monitoring carried out a specialist on the process-executing plant becomes. Therefore, online training only becomes non-critical Adapted parameters of the neural network, whereby adaptations algorithms are used that ensure the stability of the process ensure rens; e.g. B. Minimize the square mistake ler function between the network response and the post-calculation neten rolling force, the error function being only a global one Minimum but no local minima.  

DE 43 38 615 A1 discloses a method according to an analytical model of a process consisting of several sub-models using a neural network is corrected. From WO 95/14277 is known a neural network parallel to a conventional controller to use. From DE 43 01 130 A1 it is known that Parameters for a control model using a neural To influence the network. From DE 40 12 278 A1 known to use a neural network for diagnosis. IEEE Transactions on Systems, Man and Cybernetics, Vol. 24, No. 4, April 1994, pages 678-683, describes the basic Procedure for the adaptation of a neural network. From the DE 44 39 986 A1 is a Process optimization model implemented in a neural network known. The model of the process is changed in that the Weighting of the neural network in accordance with a learning algorithm of the neural network is changed. IEEE Transactions on Neural Networks, Vol. 4, No. 3, May 1993, pages 462-469 describes the implementation of a coherent neural network on a chip. IEEE transactions on Neural Networks, Vol. 6, No. 1, Jan. 1995, pages 144-156, describes an architecture based on neural networks of a controller, two neural networks, one for regulation and one for parameter identification.  

So that the well-known neural network already at the beginning of the line trainings at least approximately reasonable rolling force values predicts it can depend on a rolling force calculating randomly predetermined input variables Pre-trained rolling force model. Is there such a thing Model not available, so this can be for pre-training necessary prior knowledge by collecting training data acquired for example on comparable systems and in the neural network.

The invention has for its object, in particular the commissioning of a system or significant changes existing system using neural networks is controlled, or when converting an existing system on neuron-based control if on a previous one Collection of training data should be dispensed with, the new enable ronal network directly without pre-training a sensible on the system after just a few data points to show les behavior. Long-term drifts of the System can be recognized and compensated.

According to the invention the object is achieved by the in Patentan Proverb 1 specified learning method solved.

Advantageous further developments of the method according to the invention are specified in the subclaims.

To explain the invention, reference is made to the following Figures of the drawing referenced; show in detail

Fig. 1 is an example of a neural network with a continuously on-line running and a cyclically repeated adaptation algorithm,

Figure 2 is an example of the inventive combination of the on-line adaptation and the cyclically repeated adaptation of the neural network.,

Fig. 3 + 4 examples used in the cyclically repeated adaptation training data set,

Fig. 5 shows an example of the amount of training data when learning with exponential forgetting and

Fig. 6 shows an example for the typical error profile in the inventive learning method in comparison to a reference-learning method.

Fig. 1 shows a neural network, where a plurality of x summarized in egg nem input vector the input variables supplied is lead and the one, optionally also multi-dimensional, depending on response y _NN generated. The answer y _NN depends on adjustable parameters p of the neural network 1 . In the exemplary embodiment shown, the neural network 1 is used to emulate the relationship between influencing variables of a technical process, which are represented by the input vector x, and a process parameter y, represented by the response y _NN . An example of this is the prediction of the rolling force in a rolling process as a function of material and plant-specific influencing variables, such as the temperature of the rolling stock, the strength of the rolling stock, the thickness of the rolling stock and the decrease in thickness.

In order to learn the relationship to be simulated and to adapt the neural network 1 to the actual process, the parameters p of the neural network 1 are adapted with the aid of adaptation algorithms 2 and 3 in the sense of a reduction in the error between the response y supplied by the neural network 1 _NN and the actual value of the process parameter y changed.

After every nth (n 1) process sequence, in this case after every rolling pass, an adaptation is carried out on-line by means of the adaptation algorithm 2 , in that the influencing variables x measured and recalculated during the current process sequence take place _after the neural Network 1 are given up and the resulting response y _{NN is} compared with the process parameter y, which is also measured or recalculated. Implausible values or measurement errors are eliminated by the recalculation. Depending on the error yy _{NN determined} , selected parameters p 1 of the neural network 1 are changed in the sense of an error reduction. Non-critical parameters p 1 and such adaptation algorithms are selected that ensure the stability of the on-line adaptation and allow rapid changes in the process state to be followed. Non-critical parameters are e.g. B. the slopes of the regression levels; a possible adaptation procedure is the gradient descent.

The input variables of the input vector x determined after every nth rolling stock pass together with the measured or recalculated process parameter y, which serves as a reference variable for comparison with the answer y _NN , form a data point which is stored in a memory device 4 . Training of the neural network 1 is carried out at cyclical intervals on the basis of a training data set, possibly formed by clustering, selecting or averaging, the parameters p 2 of the neural network 1 being adaptively changed by means of the adaptation algorithm 3 . The parameters ter p₁ and p₂ can be the same or partially or completely different. If the cyclically repeated training is carried out as background training, it is not subject to online real-time conditions and can then work on the basis of training data of any size and with time-consuming globally optimizing learning algorithms that are not online-capable.

After completion of a cyclically repeated training, the neural network 1 is first subjected to an on-line adaptation with at least part of the data points on which the training data set is based, before the neural network 1 is reactivated for controlling the system and scores with new data on-line is trained. This ensures that the neural network immediately adapts to the current form of the system to be controlled after the global background training.

Fig. 2 illustrates the learning method according to the invention with the aid of two blocks 5 and 6 , of which block 5 denotes the on-line adaptation that is constantly running and block 6 denotes the adaptation which takes place cyclically. The cyclical adaptation 6 consists of an initial initial learning phase 7 and a later re-learning phase 8 .

These two different learning phases 7 and 8 are illustrated in FIG. 3, in which the number of data points N used for the cyclic training of the neural network 1 are plotted as a function of the contribution K of these data points N to the training. During the initial learning phase 7 , the cyclically repeated training takes place with a steadily increasing amount of training data 9 , all data points stored from the beginning of the learning process being used each time. The frequency of the cyclically repeated training is a parameter to be optimized for the given application. B. after every new data point, after a predetermined percentage increase in the amount of training data, or if the error yy _NN exceeds a certain value, a new adaptation takes place. In addition, the size of the neural network 1 can be changed as a function of the size of the amount of training data available, starting with a small neural network that slowly increases over time. The definition of the network size z. B. by methods of "cross-validation" or other heuristics such as residual errors in the entrance space.

After a predetermined time or when the training data amount N reaches a predetermined value N _F , the learning phase begins, in which the training data amount 10 is kept constant. The size N _F can be static or dynamic, for example by means of cross-validation techniques. The frequency of the cyclical adaptation and the size of the neural network 1 is either constant or is determined in a similar manner to that in the initial learning phase.

When a neuron-controlled system is started up, a large number of system parameters are usually changed in order to optimize the system. The suboptimal behavior immediately after start-up should be forgotten by the neural network 1 if possible. Therefore, according to the representation in FIG. 4, it is provided that the amount of training data 11 , which is constantly growing in the initial learning phase 7, is not used completely, but rather successively forgetting the oldest data points. The growth of the amount of training data 11 must of course take place faster than forgetting. The speed of forgetting can be determined in the form of a constant fraction of the growth rate of the training data set 11 , depending on the error yy _NN or depending on the expert knowledge of the commissioning engineer.

Fig. 5 shows an example of an exponentially decreasing "natural" forgetting function of a training data set 12 , which is caused by the fact that the data points in the training data set 12 are weighted with increasing age with an ever smaller weight factor.

In Fig. 6 is for the inventive learning method of the ty european error profile 13 of the residual error F of the neural network 1 in accordance with the number of available data points N in comparison with the error profile 14 of a reference-learning method, here inheritance file in a curved state, shows.

Claims (8)

1. Learning method for a neural network ( 1 ) consisting of a combination of a constantly running on-line adaptation ( 5 ) and a cyclically repeated adaptation ( 6 ) of parameters (p) of the neural network ( 1 ), the on- line adaptation (5) in each case takes place in response to a current data point from your neural network (1) currently supplied input variables (x) and to the performance thereon response (y _NN) of the neural network (1) to be compared reference variables (y) there, and the cyclically repeated adaptation ( 6 ) takes place as a function of a training data volume ( 9 to 12 ) formed from the data points (x, y) that have occurred up to that point. 2. Learning method according to claim 1, characterized in that the on-line adaptation ( 5 ) to selected parameters (p₁) of the neural network ( 1 ) is limited. 3. Learning method according to claim 1 or 2, characterized in that in an initial initial learning phase ( 7 ) the cyclically repeated adaptation ( 6 ) each time on the basis of the training data amount ( 9 , 11 ) which grows with each interim online adaptation ( 5 ). is performed and that at a later Nachlernphase (8), the training data set (10) for cyclically repeated adaptation (6) bounded by the respectively oldest in the training data set of data points contained (10) to the extent forget how new data points to the training data set ( 10 ). 4. The learning method according to claim 3, characterized in that during the initial learning phase (7), the oldest data points of the training data set (11) are not being a lesser extent, amount as the new data points to the training data (11) added. 5. Learning method according to claim 3 or 4, characterized in that the data points are forgotten by being removed from the training data set ( 11 ). 6. learning method according to claim 3 or 4, characterized, that the data points are forgotten by increasing with age with an ever lower weight factor be weighted. 7. Learning method according to one of the preceding claims, characterized in that the size of the neural network ( 1 ) is changed depending on the size of the amount of training data present ( 9 to 12 ). 8. Learning method according to one of the preceding claims, characterized in that a cyclically successful adaptation is first followed by an on-line adaptation with at least part of the data points on which the training data set ( 9 to 12 ) is based before the neural network ( 1 ) is trained with new data points.