BACKGROUND OF THE INVENTION
FIELD OF THE INVENTION
[0001]The present invention relates to technology of real time feedback control using artificial intelligence such as a neural net.
DESCRIPTION OF THE RELATED ART
[0002]Conventionally, various plants have been controlled on the basis of various kinds of control theories so as to achieve desired control results.
[0003]In the case of controlling a rolling mill as one of the plants, fuzzy control or neuro-fuzzy control has been applied as control theory targeting the shape control for controlling the wavy state of the plate, for example. The fuzzy control is applied to the shape control using coolant, and the neuro-fuzzy control is applied to the shape control of a Sendzmir mill. As disclosed in Patent Literature 1, the shape control to which the neuro-fuzzy control has been applied is executed in the following manner. That is, similarity of the difference between the actual shape pattern detected by the shape detector and the target shape pattern to the preliminarily set reference shape pattern is obtained. In accordance with the control rule expressed by an operation amount of the control operation point to the predetermined reference shape pattern, the control output amount to the operation point is derived from the similarity. It is assumed that the related art herein is configured by applying the neuro-fuzzy control to the shape control for the Sendzmir mill.

[0004]
FIG. 1 shows the shape control for the Sendzmir mill as disclosed in FIG. 1 of Patent Literature 1. The shape control for the Sendzmir mill is performed by executing the neuro-fuzzy control. In this example, the pattern recognition mechanism 51 recognizes a shape pattern from the actual shape detected by the shape detector 52 so as to arithmetically select the preliminarily set reference shape pattern with the highest similarity to the actual shape. The control arithmetic mechanism 53 executes the control in accordance with the control rule constituted by the control operation point operation amount to the preliminarily set shape patterns as shown in FIG. 2. Referring to FIG. 2 in more detail, the pattern recognition mechanism 51 arithmetically selects one of the shape patterns (s) 1 to 8, to which the pattern of the difference (As) between the actual shape detected by the shape detector 52 and the target shape (sref) is the most approximate. The control arithmetic mechanism 53 selects one of the control methods from 1 to 8 for executing the control.
CITATION LIST [0005] Patent Literature 1: Japanese Patent No. 2804161
SUMMARY OF THE INVENTION [0006]
Various operations are performed in the rolling mill by executing the plate thickness control or tension control for making the plate thickness or tension of the rolled material uniform, and the control for accelerating-decelerating the rolling speed either automatically or manually in addition to the shape control of the rolled material. The plate thickness control or the tension control may be executed by

varying the rolling load, the rolling speed and the like of the rolling mill. Variations as described above may influence the shape of the rolled material.
[0007]
For example, if the rolling load is increased, deflection of the roll constituting the rolling mill is enlarged to cause the fringe elongation. The shape control (feedback control) then serves to restrain the fringe elongation so as to maintain the uniform shape. If the shape fluctuation owing to the increased rolling load is predictable, the shape control operation point is operated feedforwardly (hereinafter referred to as the shape FF control for short) without waiting execution of the corresponding shape control (feedback control, hereinafter referred to as the shape FB control for short). This makes it possible to prevent the shape fluctuation beforehand.
[0008]
In order to execute the shape FF control, it is necessary to predict how the shape is changed in accordance with the rolling load and the rolling speed. However, it is difficult to prepare the model for predicting the shape change in different rolling states. The problem of fluctuation in the shape change still exists by changing mechanical conditions such as temperature, rolling oil, and the roll of the rolling mill.
[0009]
In the case where a plurality of control systems are applied to a generally employed controlled plant, it is assumed that one of the control systems is influenced by operation amounts of the other control systems. In the feedforward control executed by predicting the influence as described above, if accuracy of predicting the influence is low, the resultant control effect may be lowered. This may cause the other control systems more influential.

[0010]
In the above-described circumstances, the accuracy of predicting the influence has to be enhanced. The formula model of the controlled object is generally used for prediction. In most cases, however, the use of the formula model is insufficient for modeling the controlled object. Furthermore, it is difficult to reflect the change in the state of the controlled object owing to the influence of external factors. [0011]
The conventional feedforward control may cause the problem of deteriorating control accuracy owing to insufficient modelling of the controlled object. [0012]
It is an object of the present invention to provide technology that achieves the feedforward control with high prediction accuracy. [0013]
The disclosed control apparatus controls a controlled object subjected to a control operation and a disturbance to the control operation. The control apparatus includes a feedforward control method learning device that configures a first feedforward control neural network for predicting a variation of a control state variable of the controlled object from the disturbance to the controlled object by learning of learning data as a combination of the control state variable as a result of controlling the controlled object, and the disturbance to the controlled object, and a feedforward control execution device that predicts the variation of the control state variable of the controlled object from the disturbance to the controlled object based on the first feedforward control neural network, and directly or indirectly corrects a control operation amount as an operation amount of the control operation based on the variation of the control state variable to generate a control output to the controlled object.

[0014]
According to the disclosure, achievement of the feedforward control with high prediction accuracy is expectable.
BRIEF DESCRIPTION OF THE DRAWINGS
[0015]
FIG. 1 illustrates the shape control to be executed for the Sendzmir
mill as shown in FIG. 1 of Patent Literature 1.
[0016]
FIG. 2 shows a control rule specifying operation amounts of control
operation points corresponding to the prescribed shape patterns.
[0017]
FIG. 3 shows an outline of a structure of a controlled plant.
[0018]
FIG. 4 is an explanatory view of an FF control.
[0019]
FIG. 5 is an explanatory view of the FF control.
[0020]
FIG. 6 shows an outline of a plate control apparatus.
[0021]
FIG. 7 shows a control rule specifying operation amounts of control
operation points corresponding to disturbance patterns.
[0022]
FIG. 8 shows an example of a specific structure of a control rule
execution device.
[0023]
FIG. 9 shows an example of a specific structure of a control rule
leaning device.
[0024]

FIG. 10 shows a neural net structure used for the shape control for the Sendzmir mill.
[0025]
FIG. 11 shows an example of learning data stored in a learning data
database.
[0026]
FIG. 12 shows a structure of a neural net management table.
[0027]
FIG. 13 shows a structure of the learning data database.
[0028]
FIG. 14 is an explanatory view of an FF control according to a second
embodiment.
[0029]
FIG. 15 is an explanatory view of an initial setting for rolling mill
according an example 3.
[0030]
FIG. 16 is an explanatory view of an FF control according to a third
embodiment.
DESCRIPTION OF THE EMBODIMENT [0031]
Knowledge of the present invention, and circumstances leading thereto will be described, taking the shape control for the rolling mill as an example. [0032]
FIG. 3 shows an outline of structure of a controlled plant. In this case, control A and control B are executed to the controlled plant. A state variable x of the controlled plant fluctuates in accordance with a control output amount u from the control A to the controlled plant, and a control output amount d from the control B and/or through manual

operation to the controlled plant. A fluctuation amount of the state variable x of the controlled plant is set to X. [0033]
The control A controls a part of the state variables of the controlled plant, and the control B controls another part of the state variables. The control A controls a state variable y, and the control B controls a state variable y' . The state variable y to be controlled by the control A will fluctuate in accordance with not only the control output (operation amount) u of the control A, but also the control output (operation amount) d from the control B. Accordingly, the control output d from the control B becomes disturbance to the control A. Similarly, the control output u from the control A becomes disturbance to the control B. [0034]
In the embodiment of the present invention, it is assumed that the control A executes an automatic control, while generating the control output u, and processing the control output d from the control B as the disturbance, and vice versa. [0035]
In the embodiment of the present invention, the plate shape control executed in the rolling mill will be described as an example. It is assumed that the control A executes the feedback control of the plate shape (hereinafter referred to as "shape FB control"), and the control B executes the control of plate thickness and tension. The state variable y represents the shape, and the state variable y' represents the plate thickness and tension. The control output u represents operation amounts of AS-U and an intermediate roll, and the control output (disturbance) d represents operation amounts of the speed and the load. [0036]

The shape FB control according to the embodiment of the present invention performs learning of learning data as a combination of a reference shape pattern and a control operation corresponding to the shape pattern so as to perform control operations using learning results. [0037]
The neural network according to the embodiment of the present invention receives an input of the shape pattern generated in the rolling mill, and outputs a variation of the control operation to be corrected with respect to the shape pattern. The shape FB control varies the combination of the reference shape pattern and the control operation corresponding to the shape pattern so as to change the control operation for improving control results. [0038]
It is possible to apply results of learning the learning data as the combination of the reference shape pattern and the control operation corresponding to the shape pattern to the neural network. The learning allows the neural network to modify the control operation to the shape pattern to the preferred mode depending on the control result. [0039]
According to the embodiment of the present invention, in the shape FB control with the above-described neural network, correction may be made by another neural network which has been configured based on the same concept for configuring the neural network for the shape FB control. In the shape FB control, the correction may be made by executing a feedforward control (hereinafter referred to as "FF control"). [0040]
FIG. 4 is an explanatory view of the FF control. The FB control as shown in FIG. 4 corresponds to the shape FB control as described above. The FF control according to the embodiment of the present invention includes two FF controls, that is, FF control-1 and FF

control-2. The FF control-2 executes the feedforward control to the controlled plant, and the FF control-1 executes the feedforward control to the FB control. It is assumed that the FF control-1 and the FF control-2 are constituted by the separate neural networks, respectively for convenience of explanation. [0041]
As solid arrow of FIG. 4 shows, the neural network for the FF control-1 (first FF control neural network) receives an input of disturbance, and outputs a variation of a control deviation to be input to the FB control. The neural network for the FF control-1 predicts the change in the controlled object state variable (control deviation) based on the disturbance. Based on the predicted change in the controlled object state variable (control deviation), the controlled object state variable (control deviation) is corrected. [0042]
As solid arrow of FIG. 4 shows, the neural network of the FF control-2 (second FF control neural network) receives an input of disturbance, and outputs a variation of the control operation amount that has been output from the FB control. Based on the disturbance, the neural network of the FF control-2 predicts the variation of the control operation amount to be processed (corrected) by the shape FB control. Based on the predicted variation of the control operation amount, the control operation amount is corrected. [0043]
As described above, the neural network of the FF control-1 receives the input of disturbance (speed, load), and outputs the variation of the controlled object state variable (shape). Accordingly, as the dashed arrow of FIG. 4 shows, teacher data used for learning of the neural network of the FF control-1 correspond to the controlled object state variable (shape) to be output from the neural network.

That is, the relation between the controlled object state variable and
the disturbance is learned.
[0044]
Similarly, the neural network of the FF control-2 receives the input of disturbance (speed, load), and outputs the variation of the control operation amount (AS-U, intermediate roll). Accordingly, as the dashed arrow of FIG. 4 shows, the teacher data used for learning of the neural network of the FF control-2 correspond to the control operation amount (AS-U, intermediate roll) to be output from the neural network. That is, the relation between the control operation amount and the disturbance is learned. [0045]
In the embodiment of the present invention, the FF control-2 corrects the variation of the control operation amount that has been output from the FB control so as to improve accuracy of control to the controlled plant. Furthermore, the FF control-1 corrects deviation of the residual controlled object state variable caused by excess or deficiency of the correction amount output from the FF control-2 from the target value so as to accelerate execution of the FB control. The correction may reduce the deviation of the controlled object state variable of the controlled plant from the target value. This makes it possible to improve learning accuracy in accordance with the results of controlling the neural network of the FF-control-2. [0046]
FIG. 5 is an explanatory view of the FF control according to the embodiment of the present invention. [0047]
The FF control system according to the embodiment of the present invention executes the FF control by sampling the disturbance, the state variable variation, the operation amount variation and the like in a

predetermined sampling period. As the upper section of FIG. 5 shows, under the shape FB control to the rolling mill, the rolling speed fluctuation as the disturbance to the plate shape control is continued for a long time relative to the control period in which the control operation is updated under the shape control. For example, the time period taken for accelerating or decelerating the rolling speed is longer than the control period under the shape control. Therefore, for learning of the neural network, it is not appropriate to use the actual data to be acquired in the timing when the rolling speed is kept constant to make the plate shape similar to the target shape before and after occurrence of the disturbance, that is, before and after fluctuation of the rolling speed. It is appropriate to use the actual data to be acquired in the timing when the rolling speed fluctuates for learning of the neural network. [0048]
In the embodiment of the present invention, it is assumed that the deviation of the shape from the target shape (hereinafter referred to as "shape deviation") does not fluctuate even in the fluctuation of rolling speed so long as the FF control (FF control-1 and FF control-2) is optimally executed. The variation As(i) of the shape deviation between sampling points in the control period is considered to have been generated because of insufficient FF control. It is also assumed that the neural network of the FF control-2 outputs the variation Au(i) between the sampling points for the controlled operation amounts to be input to the controlled plant as indicated by the lower section of FIG. 5. The disturbance is assumed to be the disturbance amount d(i), and the disturbance variation Ad(i) as the variation of the disturbance amount d(i). The disturbance variation Ad(i) is considered as the deviation of the disturbance amount d(i) from the disturbance amount d(i-l) which has been acquired in the previous sampling of the

disturbance amount d(i). It is considered that the measurement "i+1" is detected by the detector as a result of change in the disturbance from "i-1" to "i". [0049]
The FF control-2 outputs the variation Au(i) of the controlled operation amount at the next time. Unless the FF control-2 is in the optimum state, as the middle section of FIG. 5 shows, the variation As(i) remains in the shape deviation of the controlled object state variable as a result of correcting the control operation amount u(i) by the variation Au(i) under the FF control-2. The FF control-1 predicts the variation As(i) of the shape deviation, and adds (corrects) the predicated value to the controlled object state variable (control deviation) to be input to the FB control. Repetitive execution of the FF control-2 and the FF control-1 may cause the FF control-2 to correct the FB control output by the variation Au(i). The variation As(i) of the shape deviation is expected to become 0 (zero) in the end. [0050]
In the above-described case, the FF control-1 and the FF contro-2 have been described for convenience of explanation. However, it is possible to use the single neural net for executing the FF control. The following examples will be described on the assumption of using the single neural net. [0051]
The specific examples will be described in detail referring to the drawings. [Example 1] [0052]
FIG. 6 shows an outline of a plant control apparatus according to the example. The plant control apparatus as shown in FIG. 6 is constituted by a controlled plant 1, an FB control device 2 for

executing feedback control to the controlled plant 1, a control execution device 20 which inputs a disturbance Si to the controlled plant 1 for supplying a control operation correction amount S02, and a state variable correction amount SOI, both of which have been determined in accordance with the control rule as shown in FIG. 7 for controlling the controlled plant 1, a control method learning device 21 which inputs the disturbance Si to the controlled plant 1 for learning so as to reflect the learned control rule in the control rule of the control execution device 20, a plurality of databases DB (12, 13), and a database management table 15.
[0053]
The FB control device 2, the control execution device 20, and the control method learning device 21 may be implemented by a computer that allows the processor to execute the software program specified with respect to the processing to be performed by the respective devices. The respective devices may be implemented by multiple computers. Alternatively, two or all of those devices may be implemented by the same computer.
[0054]
The FB control device 2 executes the shape FB control as described above using the neural network for feedback control. The control rule is enabled by the neural network (hereinafter referred to as "neural net") . The control operation correction amount represents the value indicating the specific amount of correction to the control operation amount, corresponding to the variation of the control operation amount. The state variable correction amount represents the value indicating the specific amount of correction to the state variable, corresponding to the variation of the state variable. The control method learning device 21 serves to configure the first FF control neural net and the second FF control neural net as described above. The control execution device 20

executes the above-described FF control using the first FF control neural net and the second FF control neural net as described above. [0055]
The control execution device 20 is mainly constituted by a control rule execution device 10 as a principal element. [0056]
The control execution device 20 generates input data SI to be input to the control rule execution device 10 in response to the disturbance Si of the rolling mill as the controlled plant 1. The control rule execution device 10 generates a control operation point correction amount S02 from the disturbance Si of the controlled object using the neural net (control rule) expressing the relation between the disturbance Si of the controlled object and a control output S2. The device also generates a control state variable correction amount SOI from the disturbance Si of the controlled object using the neural net (control rule) expressing the relation between the disturbance Si of the controlled object and a control state variable S3. [0057]
The above-structured control execution device 20 refers to the control rule database 12 for executing the process to be described later. The control rule database 12 is accessibly connected to both the control rule execution device 10 in the control execution device 20, and a control rule learning device 11 in the control method learning device 21 to be described later. The control rule (neural net) as the learning result of the control rule learning device 11 is stored in the control rule database 12. The control rule execution device 10 refers to the control rules stored in the control rule database 12. [0058]
FIG. 8 shows an example of a specific structure of the control rule execution device 10 according to the example. Upon reception of

the input data SI, the control rule execution device 10 outputs the control state variable correction amount SOI and the control operation point correction amount S02. The control rule execution device 10 includes a neural net 101 which is derived from integrating the first FF control neural net and the second FF control neural net as described above. The neural net 101 specifies the control state variable correction amount SOI and the control operation point correction amount S02 basically by executing the process as shown in FIG. 7. In the example, the control rule execution device 10 further includes a neural net selector 102 that selects an optimum control rule to be applied to the neural net 101 by referring to those stored in the control rule database 12 for executing the control. The control rule execution device 10 as shown in FIG. 8 is configured to select the necessary neural net from those classified by operator groups or the control purpose for use. The control rule database 12 may contain actual data Si (data for the operation group), which allows selection of the neural net as data from the controlled plant 1. As the control rule is enabled by executing the neural net, the neural net and the control rule are regarded as being substantially synonymous with each other rather than distinguished from each other. [0059]
Referring back to FIG. 6, the control method learning device 21 executes learning of the neural net 101 to be used for the control execution device 20. If the control execution device 20 executes the control to the controlled plant 1, the deviation from the previous value is necessary for acquiring the actual data in accordance with the control sampling under the shape control as FIG. 5 shows. The deviation from the previous value is acquired using data time delayed by the control sampling time. Referring to FIG. 6, Z"1 denotes a function unit

which causes the time delay to the respective data in accordance with
the control sampling period.
[0060]
The control method learning device 21 is constituted by a learning data generator 7, the control rule learning device 11, the control rule database 12, and the learning data database 13 as principal elements. [0061]
Upon reception of inputs of the control output S2 to the controlled plant 1, and the deviation of the state variable S3 of the controlled plant 1 from the value in the previous control sampling, the learning data generator 7 in the control method learning device 21 generates new teacher data S7a to be used for learning of the neural net, and supplies the data to the control rule learning device 11. The teacher data S7a correspond to the control state variable correction amount SOI and the control operation point correction amount S02 to be output from the control rule execution device 10. [0062]
FIG. 9 shows an example of a specific structure of the control rule learning device 11 according to the example. The control rule learning device 11 is constituted by an input data generator 114, a teacher data generator 115, a neural net processor 110, and a neural net selector 113 as principal elements. The control rule learning device 11 receives an externally input data S8a, and new teacher data S7a from the learning data generator 7, and refers to the data accumulated in the control rule database 12 and the learning data database 13. [0063]
In the control rule learning device 11, the teacher data generator 115 processes the new teacher data S7a from the learning data generator 7 to generate total teacher data S7c including the past teacher data S7b stored in the learning data database 13, and supplies the teacher data

S7c to the neural net processor 110. Those teacher data S7a, S7b are
suitably stored in the learning data database 13, and used.
[0064]
Similarly, the input data generator 114 processes the input data S8a from the neural net learning controller 112 to generate total input data S8c including the past input data S8b stored in the learning data database 13, and supplies the input data S8c to the neural net processor 110. Those input data S8a, S8b are suitably stored in the learning data database 13, and used. [0065]
The neural net processor 110 is constituted by a neural net 111 and a neural net learning controller 112. The neural net 111 acquires the input data S8c from the input data generator 114, the teacher data S7c from the teacher data generator 115, and the control rule (neural net) selected by the neural net selector 113, and stores the finally determined neural net in the control rule database 12. [0066]
The neural net learning controller 112 timely controls the input data generator 114, the teacher data generator 115, and the neural net selector 113 so as to acquire an input from the neural net 111, and stores the processing result in the control rule database 12. [0067]
The neural net 101 in the control execution device 20 as shown in FIG. 8, and the neural net 111 in the control method learning device 21 as shown in FIG. 9 are configured based on the same concept. The basic concept for those neural nets, however, will be differently transformed as they are used in the manner to be described below. The neural net 101 in the control execution device 20 is configured to have preliminarily set contents, and to obtain the control state variable correction amount SOI and the control operation point correction amount

S02 as outputs corresponding to the input data SI. In other words, this neural net is used for the unidirectional processing. Meanwhile, the neural net 111 in the control method learning device 21 is configured to learn and obtain the neural net that satisfies the input-output relation between the input data SI and the learning data constituted by the input data S8c and the teacher data S7c with respect to the control state variable correction amount SOI and the control operation point correction amount S02. [0068]
The control rule learning device 11 as shown in FIG. 9 processes results of data processing executed by the neural net learning controller 112 as follows. Using the learning data as a combination of the data S8c obtained by time delaying the input data SI to the control execution device 20, and the teacher data S7c generated by the teacher data generator 115, learning of the neural net 101 used in the control rule execution device 10 is performed. Actually, the neural net 111 that is the same as the neural net 101 in the control rule execution device 10 is installed in the control rule learning device 11, and subjected to the operation test under various conditions so that the resultant response is learned. The control rule which has been confirmed to provide better learning results is acquired. As the learning requires the use of a plurality of learning data sets, some of the past learning data sets are taken from the learning data database 13 which accumulates the learning data generated in the past. Those data are learned and processed, and the resultant learning data are stored in the learning data database 13. The learned neural net is stored in the control rule database 12 for the use in the control rule execution device 10. [0069]

Learning of the neural net may be performed in the timing of each generation of new learning data by using the new learning data together with the past learning data. Alternatively, learning may be performed in the time when certain sets of learning data sets (for example, 100 data sets) are stored by using those learning data together with the past learning data.
[0070]
An explanation will be made with respect to the plant control method, taking the shape control to the Sendzmir mill as shown in FIG. 1 and FIG. 2 as an example.
[0071]
It is assumed that in the example, any one of the neural nets is selected in accordance with the operation conditions so as to be used for executing the feedforward control. In this case, the operation conditions may be distinguished based on the control priority and the preliminarily determined condition. Specifically, the operation condition is determined based on a specification A indicating the control priority, and a specification B indicating the preliminarily determined condition.
[0072]
The specification A prioritizes the control in the plate thickness direction. For example, the specification Al prioritizes the control to the end portion, and the specification A2 prioritizes the control to the center portion. The specification B corresponds to the preliminarily determined condition. For example, as the relation between the shape pattern and the control method varies with various conditions, the specification needs to be classified into specification Bl corresponding to the plate thickness, and specification B2 corresponding to the steel type, respectively. As the conditions are changed, the degree of the resultant influence to the shape of the shape operation point varies.

[0073]
In this case, the controlled plant 1 corresponds to the Sendzmir mill, and the actual data correspond to the actual shape. The Sendzmir mill includes a cluster roll for cold rolling of the hard material such as stainless steel. The Sendzmir mill employs a small-diameter work roll for the purpose of bringing the hard material into the high pressure state. It is therefore difficult to obtain the flat steel plate. In order to cope with the difficulty, the cluster roll structure and various kinds of shape controllers are employed. Generally, the Sendzmir mill includes upper and lower first intermediate rolls, each of which is one-side tapered so as to be shiftable, and further includes six divided rolls and two rolls called AS-U at the upper and the lower sides. In the example to be described below, the rolling speed and the rolling load are used as the shape disturbance Si, and a deviation ASi between the disturbance Si and the previous value is used as the input data SI. The AS-Us with #1 to #n, and roll shift amounts of the upper and the lower first intermediate rolls are used as data of the control output S2.
[0074]
FIG. 10 shows a structure of the neural net used for executing the shape control to the Sendzmir mill. The neural net indicates the neural net 101 for the control rule execution device 10, or the neural net 111 for the control rule learning device 11. Those neural nets have the same structures.
[0075]
Referring to the example of the shape control executed to the Sendzmir mill as shown in FIG. 10, the disturbance data SI from the controlled plant 1 correspond to actual data of the Sendzmir mill, which include the rolling speed and the rolling load. Each input layer of the

neural nets 101, 111 is constituted by the deviation ASi between the present disturbance value Si and the previous value.
[0076]
Each input layer of the neural nets 101, 111 is constituted by a shape deviation variation 301 of the Sendzmir mill, and an operation point operation variation 302 of the AS-U and the first intermediate roll collectively as the shape control operation point of the Sendzmir mill. On the assumption that the AS-U has seven saddles, and the upper and the lower first intermediate rolls are shiftable in the plate thickness direction, the output layer has the shape deviation variations 301 by the number corresponding to the number of zones of the shape detector, and nine operation point operation variations 302 for seven AS-Us and two first intermediate rolls. The number of the intermediate layers, and the number of neurons for each layer may be set appropriately.
[0077]
Learning of the above-structured neural nets 101, 111 as shown in FIG. 10 is performed with respect to the method of correcting the disturbance so that the shape FF control is executed using the learned neural nets. Although the neural nets have the same structures, they may exhibit different characteristics depending on learning conditions. This makes it possible to cause the neural nets to generate different control outputs to the same disturbance.
[0078]
A plurality of neural nets may be selectively used in accordance with other conditions of rolling mills so as to execute the optimum FF control under various conditions. This applies to the specification B. The above-described structure as shown in FIG. 8 is the specific example in accordance with the specification. In the example shown in FIG. 8, different neural nets 101 to be used in the control rule execution

device 10 are prepared in accordance with the rolling results, the name of the rolling mill operator, the steel type or plate thickness of the rolled material or the like, and registered in the control rule database 12. The neural net selector 102 selects the neural net which satisfies the current condition as the neural net 101 to be used in the control rule execution device 10. The neural net may be selected by the neural net selector 102 in accordance with the current condition, for example, the plate width data which have been acquired from the actual data Si of the controlled plant 1. The multiple neural nets to be used may have different numbers of the intermediate layers, and units of the respective layers so long as the input and output layers are provided as shown in FIG. 10. [0079]
An outline of the learning data generator 7 will be described. As FIG. 6 shows, the learning data generator 7 generates teacher data S7a to the neural net 111 to be used in the control rule learning device 11 based on variations AS3 and AS2 of the control state variable S3 and the control output S2, respectively. [0080]
The teacher data S7a in this case will become the shape deviation variation 301 and the operation point operation variation 302 to be output from the output layer of the neural net 111 as shown in FIG. 10. Using the shape deviation variation 301 and the operation point operation variation 302, the teacher data generator 7 generates the teacher data S7a to the neural net 111 to be used in the control rule learning device 11. [0081]
FIG. 11 shows an example of data stored in the learning data database 13. Learning of the neural net 111 requires the learning data as a combination of many input data S8a and the teacher data S7a. The

teacher data S7a (shape deviation variation 301, operation point operation variation 302) generated by the learning data generator 7 are combined with time delay data S8a of the input data SI (disturbance 201 and disturbance variation 202) which have been input to the control rule execution device 10 in the control execution device 20. The combined data as a set of learning data are stored in the learning data database 13.
[0082]
The plant control apparatus as shown in FIG. 6 employs the respective databases 12, 13. FIG. 12 shows a structure of a neural net management table 15 for linkedly managing and operating the databases 12, 13. The management table 15 includes a specification management table. Specifically, the management table 15 includes specifications classified by the plate width (Bl) and the steel type (B2), and control priorities Al, A2. The plate width (Bl) is classified into four, for example, 3-feet width, meter width, 4-feet width, and 5-feet width. The steel type is classified into 10 from the steel type (1) to steel type (10). The specification A of the control priority is classified into two types, that is, Al and A2. In the above-described case, there are 80 classifications in total, and accordingly, the suitable neural net selected from 80 neural nets may be used in accordance with the rolling condition.
[0083]
The neural net learning controller 112 associates the learning data as the combination of the input data and the teacher data as shown in FIG. 11 with the corresponding neural net No. in accordance with the neural net management table 15 as shown in FIG. 12 so that the paired data are stored in the learning data database 13 as shown in FIG. 13.
[0084]

The learning data are generated in each timing when the control execution device 20 executes the shape control to the controlled plant 1. When the certain size of the teacher data (for example, 200 pairs) have been accumulated, or newly accumulated in the learning data database 13, the neural net learning controller 112 instructs to execute learning of the neural net 111. [0085]
The control rule database 12 stores multiple neural nets in accordance with the management table 15 as shown in FIG. 12. The neural net learning controller 112 designates the neural net No. requiring learning. Then the neural net selector 113 takes the designated neural net from the control rule database 12, and sets the thus taken neural net to the neural net 111. The neural net learning controller 112 instructs the input data generator 114 and the teacher data generator 115 to take the input data and the teacher data corresponding to the neural net from the learning data database 13. Using the acquired data, learning of the neural net 111 is executed. It is possible to use any one of neural net learning methods which have been proposed. [0086]
Upon completion of learning of the neural net 111, the neural net learning controller 112 rewrites the learned neural net 111 to the corresponding neural net No. in the control rule database 12. The learning is then completed. [0087]
Learning may be executed to all the neural nets as defined in FIG. 12 simultaneously at fixed time intervals (for example, daily), or may be timely executed to only those with neural net Nos. in the timing when the certain size of new learning data (for example, 100 pairs) have been accumulated. [0088]

In the example, automatic learning may be performed with respect to the disturbance to the rolling mill as the controlled plant 1 (fluctuation in the rolling speed, fluctuation in the rolling load), the corresponding control operation point operation amount, and the resultant shape control deviation. Execution of the feedforward control using the learned neural nets 101, 111 may reduce the control operation point operation amount and the variation of the resultant shape control deviation, resulting in improved control accuracy. [0089]
The control rule database 12 stores the neural nets to be used in the control execution device 20. If the neural net to be stored is configured only by executing an initial processing with random numbers, the long time may be taken for learning of the neural net until the permissible control is enabled. For the purpose of reducing the time as described above, upon configuration of the control device for the controlled plant 1, the control rule is preliminarily learned simulatingly based on the currently determined control model of the controlled plant 1. The neural net having simulating learning completed is then stored in the database so that the control is executable while exhibiting the performance to a certain degree at the start of the controlled plant. [Example 2] [0090]
The example 1 has been described, taking execution of the FB control to the controlled plant as an example. There may be the case where the plant is dependent on the operator's manual operation rather than execution of the FB control. The above-described feedforward control is applicable to the case where the FB control is not executed. This case will be described in the example 2. [0091]

FIG. 14 is an explanatory view of the FF control according to the example 2. In the example 2, likewise the example 1, the neural network of the FF control-1 (first FF control neural net) receives the input of disturbance, and outputs the variation of the controlled object state variable (state variable in the example 2) . In the example 2, as the FB control is not executed to the controlled plant, a control gain of the predetermined value is given to the output from the FF control-1, which will be added to the control operation amount to be input to the controlled plant. The control gain is not necessarily limited to the specific value. For example, the control gain may be set to 1, indicating that no control gain is given. [0092]
In this example, likewise the example 1, the control rule execution device 10 as the FF control system samples the disturbance (for example, speed change), the state variable variation (shape change), and the operation amount variation (AS-U, intermediate roll) at the predetermined sampling period, and accumulates the acquired actual data. The control rule for the FF control may be generated by learning the accumulated actual data. In the example 2, when the shape of the plate to be produced through the rolling mill is deteriorated, the operator is expected to manually correct the shape deviation. [0093]
Although the FB control system is not provided in this example, the relation between the shape and the operation point operation amount is predicted to a certain degree, and the operation point operation amount is arithmetically obtained using the deviation of the state variable (shape) of the controlled object. Correction of the FF control output may improve the FF control accuracy. [0094]

Assuming that the operator's manual operation is inhibited while the rolling speed is changing so far as the shape fluctuation is operationally permissible, only the shape fluctuation is learned with no manual operation so that the simple FB control is configured by the FF control-1 as shown in FIG. 14. The FF control-1 is allowed to predict the shape fluctuation, based on which the control is executed.
[Example 3]
[0095]
The examples 1 and 2 have been described, exemplifying that the FF control is executed by learning the relation between the fluctuating disturbance and variation of control operation amount and/or between the fluctuating disturbance and variation of the state variable. However, the FF control may be executed by learning the relation between the non-fluctuating input amount unlike the disturbance and the variation of the control operation amount, and between the non-fluctuating input amount and the state variable of the controlled object. The non-fluctuating input amount represents an initial setting for fixed control operations other than those to be executed.
[0096]
FIG. 15 is an explanatory view of the initial setting for the rolling mill according to the example 3. The initial setting represents the operation amount for control operation other than the one for the shape to be set adapted to the target shape while feeding the plate at the plate leaping speed. The plate leaping speed denotes the initial rolling speed at the start of rolling by the rolling mill. The initial setting is performed to the control operation amount of the control operation point to be set preliminarily before starting the rolling. As for the control operation amount, the absolute value (the initially set value of the control operation point) is set rather than setting the deviation to be processed.

[0097]
Non-coincidence of the initially set value with the correct value indicates that the shape of the plate to be processed by the rolling mill at the plate leaping speed is in the deteriorated state. This may require the operator to intervene the process through manual operations. When the shape deviation is lessened to a certain degree through the manual operation, the operator attempts to increase the rolling speed of the rolling mill from the plate leaping speed.
[0098]
The timing when the shape deviation has been made small to a certain degree is set to the data acquisition timing. As the middle section and the lower section of FIG. 15 shows, the shape deviation and the control operation amount are acquired at the data acquisition timing, and accumulated as learning data. It is therefore possible to learn the control operation point operation amount to be initially set to the operation point, and the shape deviation as a result of the initial setting.
[0099]
In the timing when the operator starts accelerating the rolling speed, or the shape deviation becomes lower than the predetermined threshold value, the timing may be determined as the data acquisition timing at which the shape deviation has been made small to the certain degree.
[0100]
FIG. 16 is an explanatory view of the FF control according to the example 3, exemplifying the initial setting of the shape. The initial setting (operation data) denotes the plate thickness, tension and the like to be initially set to the rolling mill for producing the product with predetermined specification. The initial setting for the shape is

performed in an initial setting section (shape) including an initial setting-1 and an initial setting-2.
[0101]
The neural network of the initial setting-1 (first initial setting neural network) receives an input of the initial setting
(operation data) as indicated by the solid arrow of FIG. 16, and outputs the controlled object state variable. The first initial setting neural network predicts the controlled object state variable based on the initial setting (operation data). A predetermined control gain is added to the predicted controlled object state variable so as to be used for correcting the control operation amount. The teacher data for learning of the first initial setting neural network correspond to the controlled object state variable to be output from the neural network as indicated by the dashed arrow of FIG. 16. The relation between the controlled object state variable and the initial setting (operation data) is learned.
[0102]
The neural network of the initial setting-2 (second initial setting neural network) receives an input of the initial setting
(operation data) as indicated by the solid arrow of FIG. 16, and outputs the control operation amount of the initial setting. Accordingly, the teacher data used for learning of the second initial setting neural network correspond to the control operation amount to be output from the neural network as indicated by the dashed arrow of FIG. 16. The relation between the control operation amount and the initial setting
(operation data) is learned.
[0103]
In the initial setting timing, the second initial setting neural network outputs an initial setting value of the control operation amount, and the first initial setting neural network outputs the shape deviation

as the controlled object state variable. The predetermined control gain is added to the shape deviation for correcting the control operation amount so that the corrected amount is input to control the controlled plant. [0104]
In this example, the initial setting may be learned before starting operation of the rolling mill as described above. [0105]
Actually, the plant control apparatus of the example will be implemented as a calculation system including a plurality of program groups installed therein. [0106]
For example, the program group includes: a control rule execution program which generates a control output that allows processing to be executed by the control execution device in accordance with the determined relation between the actual data of the controlled plant and the control operation; a control output determination program which determines capability-incapability of supplying the control output from the control rule execution program, and notifies the control method learning device of errors in the actual data and the control operation; a control output restraining program which restrains supply of the control output to the controlled plant if it is determined by the control output determination program that the actual data of the controlled plant are deteriorated upon supply of the control output to the controlled plant; a control result determination program for executing the control result determination process when the control execution device supplies the control output to the controlled plant for allowing the control method learning device to execute the process so as to determine whether the actual data have been made better or worse than those before the control in the time delay timing until the control

effect is reflected in the actual data; a learning data generation program which acquires the teacher data using the control result acquired from the control result determination program, and the control output; and a control rule learning program for learning the learning data including the actual data and the teacher data. [0107]
Learning of the control method learning device provides the respective relations between the actual data and the control operation for multiple control targets in accordance with conditions of the controlled plant. The acquired relation between the actual data and the control operation may be used in the control rule execution program as the combination of the actual data of the controlled plant, and the control operation. [0108]
It is necessary to set initial values for the neural net upon application of the apparatus according to the example to the actual plant. Preferably, the combination of the actual data and the control operation is prepared through simulation using the control model of the controlled plant before controlling the controlled plant so as to reduce the time for leaning the relation between the actual data and the control operation in the controlled plant. [0109]
The examples as described above include configurations to be described below. It is to be understood that the configurations of the respective examples are not limited to those as follows. [0110]
The disclosed control apparatus controls a controlled object subjected to a control operation and a disturbance to the control operation. The control apparatus includes a feedforward control method learning device that configures a first feedforward control neural

network for predicting a variation of a control state variable of the controlled object from the disturbance to the controlled object by learning of learning data as a combination of the control state variable as a result of controlling the controlled object, and the disturbance to the controlled object, and a feedforward control execution device that predicts the variation of the control state variable of the controlled object from the disturbance to the controlled object based on the first feedforward control neural network, and directly or indirectly corrects a control operation amount as an operation amount of the control operation based on the variation of the control state variable to generate a control output to the controlled object. [0111]
In the above-described structure, the first feedforward control neural network is configured by learning the relation between the control state variable as the result of controlling the controlled object, and the disturbance to the controlled object so that the control operation amount is corrected. This makes it possible to execute highly accurate control to the controlled object. [0112]
The disclosed control apparatus further includes a feedback control device that corrects the control operation amount based on the control state variable of the controlled object to generate the control output. The feedforward control method learning device further configures a second feedforward control neural network for predicting a variation of the control operation amount to the controlled object from the disturbance to the controlled object by learning of the learning data as a combination of the control operation amount and the disturbance. The feedforward control execution device predicts the variation of the control operation amount to the controlled object from the disturbance to the controlled object based on the second feedforward

control neural network to correct the control operation amount based on the variation of the control operation amount, and corrects a controlled object state variable of the controlled object as an input to the feedback control device based on the variation of the control state variable of the controlled object predicted based on the first feedforward control neural network. [0113]
In the structure for executing the feedback control, the feedforward control is additionally executed to the controlled object using the second feedforward control neural network. Furthermore, the feedforward control is additionally executed to the feedback control using the first feedforward control neural network. This makes it possible to execute the highly accurate control to the controlled object. [0114]
In the disclosure, the feedback control device learns predicts the variation of the control operation amount based on the control state variable of the control object, and corrects the control operation amount based on the variation using a feedback control neural network configured by learning of the learning data as a combination of the control state variable of the controlled object and the control operation amount to the controlled object. [0115]
In the disclosure, the first feedforward control neural network and the second feedforward control neural network are integrated to form a neural network. The structure manages the first feedforward control neural network and the second feedforward control neural network collectively as the single neural network, and further manages the learning data as a set of learning data. [0116]

In the disclosure, the feedforward control method learning device configures the first feedforward control neural network for predicting the variation of the control state variable of the controlled object from the disturbance input to the controlled object, and a disturbance variation by learning of the learning data as a combination of the control state variable of the controlled object, the disturbance to the controlled object, and the disturbance variation, and the second feedforward control neural network for predicting the variation of the control operation amount to be input to the controlled object from the disturbance input to the controlled object, and the disturbance variation by learning of the learning data as a combination of the control operation amount to the controlled object, the disturbance to the controlled object, and the disturbance variation. The feedforward control execution device predicts the variation of the control state variable of the controlled object from the disturbance to the controlled object, and the disturbance variation based on the first feedforward control neural network, and predicts the variation of the control operation amount to the controlled object from the disturbance to the controlled object, and the disturbance variation based on the second feedforward control neural network.
[0117]
In the structure, as the feedforward control neural network receives an input of the disturbance and the disturbance variation, the output reflects the influence of the disturbance with high accuracy. This makes it possible to reduce the influence of the disturbance highly accurately.
[0118]
In the disclosure, the controlled object corresponds to a rolling mill. The control state variable corresponds to a plate shape. The

disturbance includes a rolling speed. The disturbance variation
includes a variation of the rolling speed.
[0119]
In controlling the rolling mill, time zones for accelerating and decelerating the rolling speed are provided in addition to the time zone for rolling at the constant speed. In the time zone for acceleration or deceleration, the shape fluctuation is likely to occur. The structure serves to process the rolling speed variation, that is, acceleration as the input to the neural network. It is therefore possible to improve the feedforward control accuracy in the time zone for acceleration or deceleration. [0120]
In the disclosure, the feedforward control method learning device configures a plurality of the first feedforward control neural networks and the second feedforward control neural networks under different operation conditions by learning actual data for each of the operation conditions individually. The feedforward control execution device selects one of the first and the second feedforward control neural networks based on the operation conditions. [0121]
The disclosed control apparatus controls a controlled object subjected to a control operation and a disturbance to the control operation. The control apparatus includes an initial setting method learning device that configures a first initial setting neural network for predicting a control state variable of the controlled object from an initial setting to the controlled object by learning of learning data as a combination of the control state variable as a result of controlling the controlled object and the initial setting to the controlled object, and configures a second initial setting neural network for predicting a control operation amount to the controlled object from the initial

setting to the controlled object by learning of learning data as a combination of the control operation amount as an operation amount of the control operation, and the initial setting, and an initial setting execution device that predicts the control operation amount to the controlled object from the initial setting to the controlled object based on the second initial setting neural network, and corrects the control operation amount predicted by the second initial setting neural network based on the control state variable of the controlled object that has been predicted based on the first initial setting neural network to output the corrected data to the controlled object. [0122]
The structure allows the second initial setting neural network to add the initial setting to the controlled object, and the first initial setting neural network to correct the initial setting. It is possible to perform the highly accurate initial setting to the controlled object. [0123]
The present invention relates to a control method and a control apparatus for a rolling mill as one of components constituting the rolling facility, and causes no specific problem upon practical application.

WE CLAIM

1.A control apparatus for controlling a controlled object subjected
to a control operation and a disturbance to the control operation, the
control apparatus comprising:
a feedforward control method learning device that configures a first feedforward control neural network for predicting a variation of a control state variable of the controlled object from the disturbance to the controlled object by learning of learning data as a combination of the control state variable as a result of controlling the controlled object, and the disturbance to the controlled object; and
a feedforward control execution device that predicts the variation of the control state variable of the controlled object from the disturbance to the controlled object based on the first feedforward control neural network, and directly or indirectly corrects a control operation amount as an operation amount of the control operation based on the variation of the control state variable to generate a control output to the controlled object.
2. The control apparatus according to claim 1, further comprising a
feedback control device that corrects the control operation amount based
on the control state variable of the controlled object to generate the
control output, wherein:
the feedforward control method learning device further configures a second feedforward control neural network for predicting a variation of the control operation amount to the controlled object from the disturbance to the controlled object by learning of the learning data as a combination of the control operation amount and the disturbance; and
the feedforward control execution device predicts the variation of the control operation amount to the controlled object from the disturbance to the controlled object based on the second feedforward

control neural network to correct the control operation amount based on the variation of the control operation amount, and corrects a controlled object state variable of the controlled object as an input to the feedback control device based on the variation of the control state variable of the controlled object predicted based on the first feedforward control neural network.
3. The control apparatus according to claim 2, wherein the feedback control device predicts the variation of the control operation amount based on the control state variable of the control object, and corrects the control operation amount based on the variation using a feedback control neural network configured by learning of the learning data as a combination of the control state variable of the controlled object and the control operation amount to the controlled object.
4. The control apparatus according to claim 2, wherein the first feedforward control neural network and the second feedforward control neural network are integrated to form a neural network.
5. The control apparatus according to claim 2, wherein:
the feedforward control method learning device configures:
the first feedforward control neural network for predicting the variation of the control state variable of the controlled object from the disturbance input to the controlled object, and a disturbance variation by learning of the learning data as a combination of the control state variable of the controlled object, the disturbance to the controlled object, and the disturbance variation; and
the second feedforward control neural network for predicting the variation of the control operation amount to be input to the controlled object from the disturbance input to the controlled object, and the

disturbance variation by learning of the learning data as a combination of the control operation amount to the controlled object, the disturbance to the controlled object, and the disturbance variation; the feedforward control execution device:
predicts the variation of the control state variable of the controlled object from the disturbance to the controlled object, and the disturbance variation based on the first feedforward control neural network: and
predicts the variation of the control operation amount to the controlled object from the disturbance to the controlled object, and the disturbance variation based on the second feedforward control neural network.
6. The control apparatus according to claim 5, wherein:
the controlled object corresponds to a rolling mill;
the control state variable corresponds to a plate shape;
the disturbance includes a rolling speed; and
the disturbance variation includes a variation of the rolling speed.
7. The control apparatus according to claim 2, wherein:
the feedforward control method learning device configures a plurality of the first feedforward control neural networks and the second feedforward control neural networks under different operation conditions by learning actual data for each of the operation conditions individually; and
the feed forward control execution device selects one of the first and the second feedforward control neural networks based on the operation conditions.

8. A control apparatus for controlling a controlled object subjected
to a control operation and a disturbance to the control operation, the
control apparatus comprising:
an initial setting method learning device that configures a first initial setting neural network for predicting a control state variable of the controlled object from an initial setting to the controlled object by learning of learning data as a combination of the control state variable as a result of controlling the controlled object and the initial setting to the controlled object, and configures a second initial setting neural network for predicting a control operation amount to the controlled object from the initial setting to the controlled object by learning of learning data as a combination of the control operation amount as an operation amount of the control operation, and the initial setting; and
an initial setting execution device that predicts the control operation amount to the controlled object from the initial setting to the controlled object based on the second initial setting neural network, and corrects the control operation amount predicted by the second initial setting neural network based on the control state variable of the controlled object that has been predicted based on the first initial setting neural network to output the corrected data to the controlled object.
9. A control method that controls a controlled object subjected to a
control operation and a disturbance to the control operation, the
control method comprising:
configuring a first feedforward control neural network for predicting a variation of a control state variable of the controlled object from the disturbance input to the controlled object by learning of learning data as a combination of the control state variable as a

result of controlling the controlled object and the disturbance to the controlled object;
predicting the variation of the control state variable of the controlled object from the disturbance to the controlled object based on the first feedforward control neural network; and
correcting a control operation amount as an operation amount of the control operation directly or indirectly based on the variation of the control state variable to generate a control output to the controlled object.