JPH08318302A - Device for predicting abnormal behavior of looper - Google Patents

Abstract

PURPOSE: To learn the relationship between inputted data in a practical time and to output more accurate predicted results even when the number of the inputted data is larger by imparting load to every inputted data and learing Kohonen type neural network by a self-organization feature mapping algorithm. CONSTITUTION: The learning data consisting of components, rolling conditions and purpose of rolling of a rolling material which have large correlation to the presence of the abnormal behavior of looper and the presence of the abnormal behavior of looper are collected pairing them and they are stored in a 1st storing means 101. Using these learning data, the learning of the KOhonen type neural network is executed by the learning means 103 according to the self-organization feature mapping algorithm. First, the learning data are inputted by the learning means 103 from the 1st storing means 101 to the Kohonen type neural network. An information processing part where the difference between the load in an information processing part and the inputted data is min. is extracted and the load in the extracted information processing part is corrected so that the difference from the inputted data is made smaller.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a looper abnormal behavior prediction device for hot rolling equipment.

[0002]

2. Description of the Related Art In rolling equipment, for example, a hot finish rolling mill, the tension of rolled material (strip) between rolling mills (stands) has an influence on the accuracy of the product thickness, width and shape. It is large and it is required to control this tension constantly. For this reason, a looper device is installed between each stand, and as a control method of this looper device, there is a feedback control or a feedforward control which detects a state where the strip is passing through the stand and is based on the detection result.

The above control method cannot sufficiently cope with abnormalities such as sudden hunting of the looper that occurs during rolling and sudden rise of the looper when the rolling mill bites the material to be rolled. Therefore, it is determined whether or not an abnormal situation occurs by using a neural network (hierarchical model + BP (back propagation)
Patent application 1993, No. 333551 discloses a method of automatically predicting with a prediction device using an algorithm or a learning algorithm with a teacher such as a delta rule) and controlling based on the prediction. A flowchart of the process is shown in FIG. 5 and will be described below.

First, before finish rolling, rolling data (for example, target plate thickness, target plate width, target temperature, average temperature, cold strength, steel plate cross-sectional area) is input to the prediction device. This predicting device is composed of a hierarchical neural network shown in FIG. The prediction device uses information about the looper abnormal behavior that occurred during rolling and the rolling data before the rolling mill passes, and learns beforehand to output the presence or absence of the looper abnormal behavior during rolling when inputting the rolling data before rolling the strip. I'll do it. For the rolling data that is the target of prediction, this prediction device shows "abnormal:
If "Yes" is output, a warning device warns the operator so that abnormal looper behavior (hunting, sudden rise, etc.) does not occur, and a control intervention device resets the looper gain and selects the logic.

[0005]

When the hierarchical neural network used in the predicting device of the above method is trained by using a supervised learning algorithm such as BP algorithm or delta rule, the learning time is calculated with respect to the number of input data (n). Order O is (n ² ), and if the input data is small (for example, target plate thickness, target plate width, target temperature, average temperature,
It is possible to accurately learn the cold strength, the steel plate cross-sectional area (6 data), and the relationship between input and output in a short time.

However, in order to more accurately predict the presence or absence of abnormal behavior such as hunting or sudden rise of the looper, not only the above data as input data but also
It is necessary to add the rolling purpose (thickness of plate, width of plate, productivity, crown, etc.) and components of rolling material.
In that case, the prediction device is not practical because the learning time sharply increases as the number of input data increases.

In the present invention, in order to solve the above problems in the prior art, even when the number of input data is large,
An object of the present invention is to provide a looper abnormal behavior prediction device that can learn the relationship between input data within a practical time and output a more accurate prediction result.

[0008]

SUMMARY OF THE INVENTION The gist of the present invention is to provide a looper abnormal behavior prediction device for predicting the looper abnormal behavior of a hot rolling facility for rolling based on the composition of rolling material, rolling conditions, and rolling purpose. There is a large correlation with the presence or absence of rolling material components, rolling conditions, and rolling purpose, input means, and learning data consisting of rolling material components, rolling conditions, and rolling objectives that have been stored in advance, and the learning data at the time of rolling. The first storage means for storing information relating to the presence or absence of the looper abnormal behavior in association with each other, a load is given to each input data composed of the components of the rolling material, the rolling conditions, and the rolling purpose, and the input data and the load are A Kohonen-type neural network composed of a plurality of information processing units conceptually arranged in a grid pattern so as to output a function relating to a difference, and rolling stored in the first storage means. The learning data consisting of the component of the material, the rolling condition, and the rolling purpose is input to the Kohonen type neural network, and the information processing unit having the smallest difference between the load of the information processing unit and the learning input data is extracted, and the information processing unit is extracted. And a learning unit that corrects the load of the information processing unit in the vicinity thereof so that the difference from the learning input data becomes small, and a second storage unit that stores the load of each information processing unit of the Kohonen neural network after learning, Learning data consisting of components of rolling material, rolling conditions, and rolling purpose is input to the Kohonen-type neural network that has a corrected load after learning to minimize the difference between the load of the information processing unit and the input data. The information processing unit corresponds to the data classification means for classifying the input data and the learning data consisting of the components of the rolled material, the rolling conditions and the rolling purpose stored in the first storage means. Third storage means for classifying information regarding presence / absence of looper abnormal behavior for each information processing section classified by the data classifying means, and storing occurrence frequency of looper abnormal behavior in each information processing section obtained from the classification result, Data consisting of components of rolling material to be predicted, rolling conditions, and rolling purpose are input from the input means to the data classification means, and stored in the third storage means corresponding to the classification result output from the data classification means. A looper abnormal behavior prediction apparatus comprising: an occurrence frequency of the looper abnormal behavior, which is the prediction target, and an estimation unit that estimates the presence or absence of the prediction target looper abnormal behavior. .

[0009]

In the present invention, learning data consisting of components of rolling material, rolling conditions, and rolling purpose, which have a strong correlation with the presence or absence of the looper abnormal behavior, and the presence or absence of the looper abnormal behavior are collected as a pair and are collected first. It is stored in the storage means (101).

Learning of the Kohonen type neural network is performed by the learning means (103) in the following manner according to the self-organizing feature mapping algorithm using the learning data consisting of the components of rolling material, rolling conditions and rolling purpose. . (1) First, the learning data is input to the learning means (103) from the first storage means (101) to the Kohonen type neural network. (2) Then, the information processing unit having the smallest difference between the load of the information processing unit of the Kohonen neural network and the input data is extracted. (3) Correct the loads of the extracted information processing unit and the information processing units in the vicinity thereof so that the difference from the input data becomes small.

By repeating the above processes (1) to (3), the load of each information processing unit is corrected to a representative value of the learning data classified into this information processing unit. When learning the Kohonen type neural network with the self-organizing feature mapping algorithm, the number of iterations of (1) to (3) above is 5000 times if sufficient learning data is prepared even if the number of input data increases. The degree is enough.

After the learning is completed, the weight of the information processing unit forming the Kohonen type neural network is stored in the second storage means (10
4) and input the learning data stored in the first storage means (101) to the Kohonen-type neural network that has a corrected weight after learning, and stores it in the data classification means (105). The input data is classified into an information processing unit in which the difference between the load of the information processing unit and the input data is minimized. Then, the information regarding the presence or absence of the looper abnormal behavior stored in the first storage means (101) is associated with each information processing section obtained from the classification result, and the occurrence of the looper abnormal behavior in each information processing section is generated. The frequency is stored in the third storage means (10
Store in 6).

When estimating the presence or absence of abnormal looper behavior during rolling of the rolled material to be predicted, first, input means (1
02) The data including the components of the rolled material to be predicted, the rolling conditions, and the rolling purpose is input to the data classification means (105), and the data is classified into one of the information processing units. In the estimation means (107), the occurrence frequency of the looper abnormal behavior stored in the third storage means (106) corresponding to the classification result is set as the occurrence frequency of the looper abnormal behavior to be predicted, and it exceeds the threshold value. If so, it is estimated that there is a possibility of abnormal looper behavior: Yes, and if not, there is a possibility of abnormal looper behavior.

According to the above looper abnormal behavior prediction device,
A plurality of information processing units that are conceptually arranged in a grid pattern to give a load to each input data composed of components of rolling material, rolling conditions and rolling purpose, and output a function related to a difference between the input data and the load. Since the Kohonen-type neural network consisting of is learned by the self-organizing feature mapping algorithm, the learning time order O is (n) for the number of input data (n), and practical learning is possible even if the input data increases. You can finish learning in time.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. 1 is a block diagram showing the principle of the looper abnormal behavior prediction apparatus of the present invention, FIG. 2 is a system configuration diagram when implementing an embodiment of the present invention, and FIG. 3 is provided in the looper abnormal behavior prediction apparatus. FIG. 4 is a conceptual diagram showing a Kohonen type neural network composed of a plurality of information processing sections, and FIG. 4 shows the relationship between the distance from a specific information processing section that has fired and the number of times of learning when learning the Kohonen type neural network. It is a figure. The embodiments described below are merely examples embodying the present invention, and do not limit the technical scope of the present invention.

The rolling equipment shown in FIG. 2 includes a hot finish rolling mill (201) arranged in tandem and a looper device (202) arranged between the rolling mill (201), and the rolling mills (201) and the looper. The control device (203) for controlling the device (202), the prediction device (204) for predicting the looper abnormal behavior, and the prediction device (2
04) and the control device (203) to connect the control intervention device (205), and the prediction device (204) connected to the alarm device
(206) is provided.

Each of the rolling mills (201) is driven by a mill motor (207). The looper device (202) is a rolling mill (2
The looper (209) comes into contact with the rolled material (208) running between (1) and (1) so as to give an appropriate tension to the rolled material (208), and the looper (209) is appropriately adjusted by the looper motor (210). Torque is given. The position of the looper (209) is detected by the looper angle detection device (211). The control device (203) includes a looper position setting device (212), a controller (213), and a mill motor speed control device (214).

That is, the looper position detected by the looper angle detection device (211) and the looper position setting device (212)
By controlling the speed of the mill motor (207) via the controller (213) and the mill motor speed controller (214) based on the deviation from the set position set by
The bending length of the rolled material (208) between (201) is controlled to perform feedback control so as to keep the looper position at a set position, and the tension of the rolled material is maintained at a predetermined value.

The predicting device (204), based on the information obtained by learning, provides data (a component having a large correlation with the looper abnormal behavior related to the rolled material before strip running) given by the host computer (215). , Rolling condition, purpose of rolling) ", the frequency of occurrence of looper abnormal behavior (hunting, sudden rise, etc.) during the passage of this rolled material is calculated, and if this frequency is greater than the threshold value, the possibility of looper abnormal behavior is: If the presence signal is smaller than the threshold value, there is a possibility that the looper has abnormal behavior: a absence signal is output to the control intervention device (205) and the alarm device (206).

When the control intervention device (205) receives a signal indicating the possibility of abnormal looper behavior: existence, it resets the control gain and selects the control logic, and outputs it to the control device (203). The control device (203) controls the rolling mill (201) based on the reset control gain and control logic.

The Kohonen type neural network used in the predicting device (204) is composed of a plurality of information processing units which are conceptually arranged in a lattice as shown in FIG. 3, and input data X = ( X ₁ , X ₂ , ..., X _k ) has the same dimension load W = (W _ij1, W _ij2, ..., W _ijk ). W conceptually indicates the load of the information processing unit arranged in the i-th row and the j-th column. The input data is input to all the information processing units, and the information processing unit calculates the difference Y _ij between the input data and the load by the following equation (1).

[0022]

[Equation 1]

The self-organizing feature mapping, which is a learning algorithm for the Kohonen neural network, will be described below.

In advance, the components of the rolled material, the rolling conditions,
The learning data for the purpose of rolling and the presence / absence of abnormal looper behavior are collected as a pair and stored in the first storage means (101). The learning data is input to each information processing unit of the Kohonen type neural network, and the information processing unit having the smallest difference between the load defined by the equation (1) and the input data is extracted from the information processing units. The weights of the extracted information processing units and the information processing units in the vicinity thereof are corrected according to the following equation (2) so that the difference with the input data becomes small. Note that, as shown in FIG. 4, as the learning number T increases, the learning coefficient α becomes smaller and the range for correcting the load becomes narrower.

W _ijk ^new = W _ijk ^old + α (X _k −W _ijk ^old ) ... (2) where W _ijk ^new : corrected load W _ijk ^old : uncorrected load α = 0.5 (T <500) 0.3 (500 ≦ T <1000) 0.1 (1000 ≦ T <3000) 0.05 (3000 ≦ T) where T is the number of learning

By repeating the above process,
The load of each information processing unit changes to a representative value of the learning data classified into this information processing unit.

As described above, when the Kohonen type neural network is trained by the self-organizing feature mapping algorithm, even if the number of input data (n) is increased, the order O of the above iterative times is (n). Learning can be finished in a practical learning time.

Further, since the input data is composed of "a plurality of data (composition of rolling material, rolling condition, purpose of rolling) having a large correlation with the presence or absence of abnormal looper behavior", the abnormal looper behavior (hunting during striping to be predicted) , Sharp rise, etc.) prediction accuracy is improved.

[0029]

As is apparent from the above description,
According to the present invention, a load is applied to each input data composed of components of rolling material, rolling conditions, and rolling purpose, and conceptually arranged in a grid pattern so as to output a function relating to a difference between the input data and the load. By learning a Kohonen-type neural network consisting of multiple information processing units by self-organizing feature mapping, the number of iterations for learning hardly changes even when the number of input data increases, and it can be achieved in a practical time. Can learn. In addition, since the input data is composed of "a plurality of data (composition of rolling material, rolling condition, rolling purpose) having a strong correlation with the presence or absence of abnormal looper behavior", the abnormal looper behavior (hunting, sudden rise, etc.) during the target strip running )
The prediction accuracy of is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the principle of a looper abnormal behavior prediction device of the present invention.

FIG. 2 is a system configuration diagram when implementing an embodiment of the present invention.

FIG. 3 is a conceptual diagram showing a Kohonen type neural network configured by a plurality of information processing units included in the looper abnormal behavior prediction device of the present invention.

[Fig. 4] When learning a Kohonen neural network,
It is a figure which shows the relationship between the distance from the specific information processing part which ignited, and the number of times of learning.

FIG. 5 is a diagram showing a flowchart of a conventional processing process.

FIG. 6 is an explanatory diagram of a hierarchical neural network.

[Explanation of symbols]

101 ... First storage means, 102 ... Input means, 103 ... Learning means, 104 ... Second storage means, 105 ... Data classification means, 106
... Third storage means, 107 ... Estimating means, 201 ... Rolling mill, 202
… Looper device, 203… Control device, 204… Prediction device, 205
… Control intervention device, 206… Alarm device, 207… Mill motor,
208… Rolled material, 209… looper, 210… looper motor,
211 ... Looper angle detection device, 212 ... Looper position setting device, 213 ... Controller, 214 ... Mill motor speed control device, 215 ... Host computer.

Claims (1)

[Claims] 1. A looper abnormal behavior prediction device for predicting abnormal looper behavior of hot rolling equipment for rolling based on the composition of rolling material, rolling conditions, and rolling purpose, and a component of rolling material having a large correlation with the presence or absence of abnormal looper behavior. ,
Input means for capturing the rolling conditions and rolling purpose, and the components of rolling materials, rolling conditions, and
First storage means for storing learning data consisting of rolling objectives and information relating to presence / absence of abnormal looper behavior during rolling of these learning data, and input data composed of components of rolling material, rolling conditions, rolling objectives A Kohonen-type neural network composed of a plurality of information processing units conceptually arranged in a grid to output a function related to the difference between the input data and the load, which is stored in the first storage means. Rolling material composition, rolling conditions,
Learning data consisting of rolling purpose is input to the Kohonen neural network, and an information processing unit having a minimum difference between the weight of the information processing unit and the learning input data is extracted, and the information processing unit and its neighboring information processing units. Learning means for modifying the load of the so that the difference from the learning input data becomes small, the second storage means for storing the load of each information processing unit of the Kohonen neural network after learning, the components of the rolling material, the rolling conditions , Input the learning data consisting of the rolling purpose to the Kohonen type neural network with the corrected weight after learning, and input to the information processing unit where the difference between the load of the information processing unit and the input data is the smallest. Data classification means for classifying data, and presence of looper abnormal behavior corresponding to learning data composed of components of rolling material, rolling conditions, and rolling purpose stored in the first storage means. Third storage means for classifying the information about the information processing units classified by the data classifying unit and storing the occurrence frequency of the looper abnormal behavior in each information processing unit obtained from the classification result; The looper abnormality stored in the third storage means corresponding to the classification result output by inputting data consisting of material components, rolling conditions, and rolling purpose from the input means to the data classification means. A looper abnormal behavior prediction apparatus comprising: an occurrence frequency of the behavior as a frequency of occurrence of the prediction target looper abnormal behavior, and estimation means for estimating the presence or absence of the prediction target looper abnormal behavior.