JPH0890020A - Device for setting draft schedule for tandem rolling mill - Google Patents

Abstract

PURPOSE: To quantify an operator's knowledge/operation concerning a draft schedule by outputting through an inference means the draft data of each stand, which are stored in a second storing means corresponding to an information processing part that is ignited when inputted in a neural network after learning. CONSTITUTION: The learning data for rolling conditions and rolling purpose of a tandem rolling mill and the draft data of each stand are preliminarily stored in a first storage part 3. By inputting the learning data in a self- organizing neural network, a self-learning is performed in a learning part 5 on each information processing part constituting the neural network. The draft data of each stand, which correspond to each data classified at the time of completion of learning in the learning part 5, are stored in a third storage part 8 for each information processing in which the learning data are classified. The draft data of the third storage part 8, which correspond to the information processing part that is ignited when inputted in the neural network after learning, are outputted from an inference part 9.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a draft schedule setting device for a tandem rolling mill, and more particularly to a device for determining a draft schedule for a tandem rolling mill based on rolling conditions and rolling objectives.

[0002]

2. Description of the Related Art It is a stable operation to optimize a draft schedule (reduction ratio distribution) in a continuous rolling mill in which a material to be rolled such as a steel plate is continuously passed through a plurality of stands arranged in tandem and rolled according to rolling conditions. ， It is important for ensuring quality accuracy. A method for automatically setting this draft schedule using a computer is disclosed in Japanese Patent Laid-Open No. 5-38511. This conventional draft schedule setting method is to quantify the knowledge of the operator's draft schedule in advance using a hierarchical neural network learned by a backpropagation algorithm, and use this to set the draft schedule. FIG. 5 shows a schematic processing procedure thereof, which will be described below. First, the construction and learning of a hierarchical neural network will be described. First, a hierarchical neural network whose structure is defined as follows is constructed (see FIG. 6) (S51). (1) The number of units in the input layer is the number of rolling conditions such as strip width, strip thickness, roll roughness, and rolled material required for the operator to determine the draft schedule. (2) The number of units in the output layer is the number of stands in the tandem rolling mill. (3) The number of intermediate layers is two, and the number of units is any number that is greater than or equal to the number of units in the input layer and greater than or equal to the number of units in the output layer. Next, a set of data pairs composed of the rolling conditions input to the input layer and the desired output of the output layer (reduction ratio distribution of each stand set by a skilled operator) is set in the above step S5.
The learning data of the hierarchical neural network constructed in 1 is collected (S52). The input data (rolling condition) of the collected learning data and the teacher data (reduction ratio set by the operator) are normalized so that each data falls within a range of ± 0.5 (S53). Next, using the learning data obtained in steps S52 and S53 above, when the normalized rolling conditions are input, the weights between the units are backpropagated so that the teacher data paired with them are output. Is corrected using (S54). Learning consists of neural network output and teacher data.
There are a method of performing until the time when the multiplicative mean error becomes smaller than an arbitrary set value α and a method of performing until the time when the number of learning times reaches an arbitrary set number N, but here, in order to avoid overlearning,
The method of setting the number of times of learning is adopted (S55). Next, a method for obtaining a draft schedule for the rolling conditions of the target rolled material will be described. First, the same normalization as during learning is performed on the rolling conditions of the target rolled material. Then, when the normalized rolling condition is input to the learned neural network, the normalized draft schedule estimated value can be obtained from the output layer (S56, S57).
Since the estimated value of the rolling reduction of each stand from the output layer is normalized to ± 0.5, the draft schedule for the rolling condition of the target rolled material can be obtained by performing non-normalization. (S58).

[0003]

In the conventional draft schedule setting device for the tandem type rolling mill as described above, when the input data is small (for example, 4 data of strip width, strip thickness, roll roughness, rolled material). ), The learning is repeated using the backpropagation algorithm, and a hierarchical network can be obtained so as to output the reduction rate within the required error. However, when the operator actually sets the draft schedule, not only the rolling conditions but also the rolling purpose (for example, emphasis on sheet thickness, sheet width, productivity, crown) is taken into consideration. If the number of input data is increased in order to more accurately reproduce the knowledge / operation of such an operator, it is possible to obtain a hierarchical neural network that outputs a reduction ratio within the required error due to a drop in the local solution. There was a problem peculiar to supervised learning that it disappeared. The present invention has been made in view of the above circumstances, and an object thereof is to quantify the knowledge / operation regarding the draft schedule of the operator even when the number of input data is large, and to reduce the reduction ratio within the required error. It is to provide a draft schedule setting device capable of outputting.

[0004]

In order to achieve the above object, the present invention stores learning data consisting of a rolling condition and a rolling purpose of a tandem type rolling mill and the rolling reduction data of each stand in association with each other. A first storage means for storing the first storage means;
By inputting the learning data stored in the memory means to the self-organizing neural network, the learning means for performing self-learning of each information processing unit constituting the neural network, and the learning means classified when the learning by the learning means are completed. The reduction data of each stand corresponding to the learning data
Second storing the learning data for each information processing unit into which the learning data is classified
The storage means and the newly input input data including the reduction condition and the reduction purpose are stored in the second storage means corresponding to the information processing unit that is fired when input to the neural network after learning. It is configured as a draft schedule setting device of a tandem rolling mill, which is provided with an inference unit that outputs rolling reduction data of each stand. Further, it is a draft schedule setting device for a tandem rolling mill in which all of the learning data, the rolling reduction data and the input data are normalized data. Further, it is a draft schedule setting device for a tandem rolling mill in which the ignition condition of the information processing unit in the inference means is that the inner product of the load of the information processing unit and the input data is maximum. Further, the second storage means is a draft schedule setting device for a tandem type rolling mill that averages and stores the rolling reduction data of each stand.

[0005]

According to the present invention, the learning data including the rolling condition and the rolling purpose of the tandem type rolling mill and the rolling reduction data of each stand are stored in advance by the first storage means. When the learning data stored in the first storage means is input to the self-organizing neural network, the learning means performs self-learning of each information processing unit that constitutes the neural network. The rolling reduction data of each stand corresponding to each data classified when the learning by the learning unit is completed is stored in the second storage unit for each information processing in which the learning data is classified. The reduction rate of each stand stored in the second storage means corresponding to the information processing unit that is fired when the newly input input data including the rolling condition and the rolling purpose is input to the learned neural network. The data is output by the inference means. As a result, even if the number of input data is large, the operator's knowledge / operation regarding the draft schedule can be quantified,
Since unsupervised learning is performed by the self-organizing neural network, the number of iterative calculations for learning hardly changes despite the increase in the number of input data and it is possible to estimate the draft schedule with less error.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments embodying the present invention will be described below with reference to the accompanying drawings for the understanding of the present invention. The following embodiments are examples of embodying the present invention and are not intended to limit the technical scope of the present invention. Here, FIG. 1 is a block diagram showing a schematic configuration around a draft schedule setting device A of a tandem rolling mill according to an embodiment of the present invention, FIG. 2 is a flow chart showing a schematic processing procedure by the device A, and FIG. Is a conceptual diagram of the Kohonen type neural network, and FIG. 4 is an explanatory diagram showing changes in the firing state of the Kohonen type neural network. As shown in FIG. 1, the draft schedule setting device A of the tandem rolling mill according to the present embodiment mainly associates learning data including rolling reduction conditions and rolling purposes of the tandem rolling mill with rolling reduction data of each stand. A first storage unit 3 (corresponding to a first storage unit) that is stored in advance and learning data stored in the first storage unit 3 are input to a self-organizing neural network to configure this neural network. The learning data is classified into the learning unit 5 (corresponding to learning means) that performs self-learning of each information processing unit and the rolling reduction data of each stand corresponding to the learning data classified when the learning by the learning unit 5 is completed. The third storage unit 8 (corresponding to the second storage unit) that stores each information processing unit, and the input data including the newly input reduction condition and rolling purpose It is composed of an estimation unit 9 (corresponding to an inference unit) that outputs the rolling reduction data of each stand stored in the third storage unit 8 corresponding to the information processing unit that fires when input to the subsequent neural network. .

Further, all of the learning data, the rolling reduction data, and the input data may be normalized data. Furthermore, the firing condition of the information processing unit in the estimation unit 9 may be that the inner product of the load of the information processing unit and the input data is maximum. Furthermore, the third storage unit 8 may average and store the rolling reduction data of each stand. Hereinafter, the processing procedure of the apparatus A will be described with reference to FIG. The overall structure of this processing procedure is roughly classified into a "learning process" (steps S1 to S3) and a "recall process" (steps S4 to S8). First, the "learning process" will be described. Rolling conditions (roll width, target thickness, rolling material, roll roughness, rough bar thickness) of rolled material used when learning operator's knowledge in advance
And rolling purpose (0 for emphasis on plate thickness, productivity, crown)
Learning data consisting of ~ 1) and the rolling reduction of each stand set by the operator (normalized with a real number of 0 ~ 1)
And are collected in pairs and stored in the first storage unit 3 (S1). In the above learning data, the rolling condition is composed of a plurality of data of different orders, and in order to prevent an error from occurring due to a cancellation of digits due to calculation, first, the normalizing unit 4 is arranged so that each data can be treated uniformly. Normalization is performed (S2).

Next, using the learning data after normalization, FIG.
Learning (self-organizing feature mapping) of the Kohonen network (self-organizing network) whose concept is shown in (a) is performed as follows. (1) The learning data after normalization is input to the Kohonen network, and the neuron (in the information processing unit) having the maximum inner product of the associated weight and the input data is extracted. In that case, both the load and the input data are standardized, and the inner product becomes maximum when the directions of both are the same. (2) The weights of the extracted neurons and the neurons in the range shown in FIG. 4 are corrected so that the inner product with the input data becomes large. (3) By repeating the above process, the weights associated with the representative values of the learning data classified into each neuron are learned. The self-organizing feature mapping as described above is unsupervised learning, and the number of iterations of (1) and (2) above is about 5000 if sufficient learning data is prepared even if the number of input data increases. Is enough. Here, the learning will be further explained. The set of normalized learning data representing rolling conditions and rolling purpose is expressed by the following equation. D = [d ₁ , d ₂ , ... D _n ] Next, the neural network is trained to learn a method of classifying the data of the set D into the following patterns. D ^p = [d ₁ ^p , d ₂ ^p , ... d _m ^p ], m <n Normally, in the Kohonen type neural network, neurons are arranged in a lattice as shown in FIG. Here, for convenience of description, FIG. 3B, which is a simplified version of FIG. 3A, will be described.

Now, let m = m ₁ × m ₂ . A pattern vector d _j ^p is associated with each neuron n _j (j = 1, 2, ..., M). Data vector d _i (i
, 1, 2, ..., N) is input to the neural network, each neuron n _j fires only the inner _integral of the data vector d _i and its own pattern vector d _j ^p . That is, the firing degree A _j of each neuron n _j is given by the following equation. A _j = (d _i , d _j ^p ) Now, A _k = max A _j (1 ≦ j ≦ m) (that is, n _k
Is the neuron with the highest firing degree), the data d is classified into the pattern d _k ^p . In the learning of the Kohonen type neural network, the data D is repeatedly input to the neural network, and each time, the pattern vector d _j ^p of the neuron having the highest firing degree and the neighboring neurons is updated as follows. To do.

[Equation 1] In this way, [d _j ^p ; j = 1, 2, ..., M] is obtained. That is, self-learning is performed.

After the learning is completed, the weights of the neurons forming the Kohonen type neural network are stored in the second storage section 6 (S3). Next, the "recall process" will be described. After normalizing all the learning data stored in the first storage unit 3 by the normalization unit 4, it is input to the Kohonen type neural network in which the weights after learning are set, and the data classification unit 7 accompanies the neurons. The input data is classified into the neuron having the maximum inner product of the weight and the input data (S4). The average of the rolling reduction of each stand corresponding to the classified input data is stored in the third storage unit 8 for each neuron into which the input data is classified (S5). Next, a case will be described in which the rolling reduction of each stand is estimated from the new rolling state and the rolling purpose input from the input unit 2. First, input data including a rolling condition (rolling condition) and a rolling purpose is input from the host computer 1 via the input unit 2 (S6). This input data is normalized by the normalization unit 4 (S7). After that, the learned weights stored in the second storage unit 6 are input to the Kohonen type neural network, and the data classification unit 7 outputs the inner product of the associated weights of the neurons and the input data to the maximum neuron. And the average value of the rolling reduction of each stand stored in the third storage unit 8 corresponding to the extracted neuron is estimated by the estimation unit 9 for each stand of the rolled material having the above-mentioned rolling state or rolling purpose. (S8).

In the above "recall process", new data d is input to the learned neural network to search for a neuron having the highest firing degree. At this time,
Data that fires the same neuron is considered "similar". However, the "recall process" has the following processes. (1) First, with respect to the data used for learning, the learning data is classified as follows according to which neuron fires the maximum. D = D ₁ ∪D ₂ ∪ ... ∪D _m (2) Calculate the average rolling reduction pattern for the classification data D _i . That is, the average rolling reduction pattern is associated with each neuron n _j . (3) When a certain data d is given, the neuron with the highest firing degree is searched for, and the average rolling reduction pattern corresponding to the neuron is called. The step S4 executes the step (1), the step S5 executes the step (2), and the steps S6 to S8 execute the step (3). Information is transmitted to the measuring device 10, the drive motor control device 11, and the reduction control device 12 in order to realize the reduction ratio of each stand. Therefore, according to this draft schedule setting device capable of automatically quantifying the operator's knowledge / operation, a load is given to each input data composed of rolling condition and rolling purpose, and a function related to the sum of the input data and the load is given. Since the Kohonen-type neural network part consisting of a plurality of neurons conceptually arranged in a grid pattern to output is unsupervised learning by self-organizing feature mapping, iterative calculation for learning is performed even if the number of input data increases. It is possible to infer a draft schedule with almost no change in the number of times and with little error. still,
In the above-described embodiment, the average value of the rolling reduction ratio of each stand is stored in the third storage unit 8. However, in actual use, individual data may be stored and averaged at the time of estimation by the estimation unit 9. Good. Further, instead of the average value of the rolling reduction, its specific value may be used. For example, the maximum value or the minimum value may be used.

[0012]

Since the draft schedule setting device for the tandem rolling mill according to the present invention is configured as described above, even if the number of input data is large, the operator's knowledge / operation regarding the draft schedule can be quantified.
Since unsupervised learning is performed by the self-organizing neural network, the number of iterative calculations for learning hardly changes despite the increase in the number of input data and it is possible to estimate the draft schedule with less error.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration around a draft schedule setting device A of a tandem rolling mill according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a schematic processing procedure by the apparatus A.

FIG. 3 is a conceptual diagram of a Kohonen neural network.

FIG. 4 is an explanatory diagram showing changes in the firing state of a Kohonen neural network.

FIG. 5 is a flowchart showing a schematic processing procedure in an example of a draft schedule setting method for a conventional tandem rolling mill.

FIG. 6 is a conceptual diagram of a conventional hierarchical neural network.

[Explanation of symbols]

 A ... Draft schedule setting device 3 ... First storage unit (corresponding to first storage unit) 5 ... Learning unit (corresponding to learning unit) 8 ... Third storage unit (corresponding to second storage unit) 9 ... Estimating unit (Equivalent to inference means)

Continuation of front page (51) Int.Cl. ⁶ Identification number Office reference number FI Technical display location G05B 13/02 L 9131-3H G06F 15/18 550 E 8837-5L G06G 7/60

Claims (4)

[Claims] 1. A first storage means for preliminarily storing, in association with each other, learning data consisting of a rolling condition of a tandem type rolling mill and a rolling purpose and rolling reduction data of each stand, and the first storing means. By inputting the stored learning data to the self-organizing neural network, a learning means for performing self-learning of each information processing unit constituting the neural network, and learning data classified when the learning by the learning means is completed Second learning means for storing the corresponding rolling reduction data of each stand for each information processing unit into which the learning data is classified, and the newly input neural network after the learning, with the input data including the rolling reduction condition and the rolling reduction purpose. It is recommended to output the rolling reduction data of each stand stored in the second storage means corresponding to the information processing unit that fires when input to the network. And a draft schedule setting device for a tandem type rolling mill. 2. The draft schedule setting device for a tandem rolling mill according to claim 1, wherein the learning data, the reduction ratio data, and the input data are all normalized data. 3. The draft schedule setting device for a tandem rolling mill according to claim 1, wherein the ignition condition of the information processing unit in the inference means is that the inner product of the load of the information processing unit and the input data is maximum. 4. The draft schedule setting device for a tandem rolling mill according to claim 1, wherein the second storage means averages and stores the rolling reduction data of each stand.