JP2728492B2 - Control method and apparatus using neural network - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial applications]

The present invention relates to a control method and an apparatus using a neural network in which an operation amount of an actuator (operating means) for operating a controlled object such as a rolling mill is determined using a neural network and an inference mechanism.

[Prior art]

For example, a rolling system is a system that obtains a steel material having a desired thickness by controlling the interval between rolls facing each other and the tension applied to the rolled material. However, a flat steel material cannot be obtained due to roll deformation caused by heat deformation, mechanical deformation, or the like due to heat loss generated during rolling. Therefore, shape control has been developed to obtain flat characteristics. However, since the physical characteristics of rolling change significantly due to various factors, even if a control model is created and controlled in the vicinity of a specific operating point, the model often differs from the actual operation of the rolling mill. I will. For this reason, if the model is accurate, the feedback control that produces a good result cannot exert its ability sufficiently, and there is a problem that it cannot exceed the skillful operator who operates with intuition and experience. On the other hand, in recent years, in order to reduce the burden on the operator, the operation status of the actuator that operates the rolling mill and the shape of the rolled material are detected using various sensors, and these detection signals are sent to an inference mechanism such as a production system. There has been considered a system that determines the operation amount of an actuator by taking in and performing inference processing.

[Problems to be solved by the invention]

If there are many combinations of multiple pieces of detection information, the inference processing becomes enormous, and the range that can be processed by the current processing device due to interference between control rules for inference must be very narrow, The knowledge used for inference is also limited. Therefore, there is a problem that an appropriate operation amount of the actuator cannot be obtained according to the situation of the controlled object. As a result, the controlled object cannot be accurately controlled. SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a control method and an operation using a neural network that can improve the accuracy of determining an operation amount of an actuator.

[Means for Solving the Problems]

A feature of the present invention is that a combination of a plurality of detection signals relating to a controlled object is recognized as a detection pattern, and this detection pattern is compared with a preset setting pattern, and the detection pattern is similar to the setting pattern. The numerical value indicating the ratio is calculated using a neural network, and the numerical value indicating the similarity ratio and the inference using the control rule predetermined for the operating device are executed to obtain the operation amount of the operating device. is there.

[Action]

A plurality of detection signals of the controlled object, which are changed by the operation of the operation means, are combined and recognized as a pattern, and a control rule of an inference mechanism is created for the pattern. Therefore, the number of control rules can be reduced as compared with the case where a control rule is created for each detection signal. As the inference mechanism reduces the number of control rules, the interference between control rules is significantly reduced.
The accuracy of operation amount determination can be improved.

【Example】

 FIG. 1 shows an embodiment of the present invention. The controlled object 1 includes a plurality of actuators,
The actuator issues a control execution command for each actuator.
It is controlled by the actuator control device G6 to be generated.
The operation of the controlled object 1 and the plurality of actuators are respectively
Detecting device composed of various detectors arranged
Detected by G14. Multiple detector outputs of this detector G14
The signal is input to the pattern control device G13. Pattern system
Control device G13 detects multiple input detector output signals
The input part recognized as a pattern and this detection pattern are recorded.
The memory function to remember and the detection pattern are stored first
Compare with multiple setting patterns and check similarity with setting patterns
The processing unit to be output and each actu
Determines the amount of operation of the
Command generation to generate command signal to control device G6 and learning device G16
And a mechanism. The learning device G16 is a putter
The setting pattern is written to the pattern storage function of the
When remembering, the operator sets the set pattern in advance.
Operation when inputting and operating the pattern control device G13
The function to recognize the quantity and the
Setting pattern storage contents when the tutor is changed
Function to change the system in advance and the system status
It has a function to do. The conventional system does not have the pattern control device G13.
Is sent from the detection device G14 to the actuator control device G6.
Control system for each individual actuator.
I was Add pattern control device G13 as in the present invention
By doing so, a plurality of addresses are provided to the controlled object 1 according to the situation.
In addition to appropriately determining the amount of operation of the actuator,
If one of the tutors fails, the other actuator
You can bar and report the situation accurately. Actuate
Can respond flexibly to changes in data and controlled objects.
It has the effect of. Next, details will be given using an example in which the present invention is applied to rolling mill control.
Will be described. Hereinafter, an embodiment applied to a rolling mill control system of the present invention.
Will be described with reference to FIG. The rolling mill as the controlled object 1 is a set of opposing works
Rolled material 3 sandwiched between rolls 2
The pressure acting between the work rolls 2 between the rolled material 3 sandwiched between
By the rolling force and the tension acting on the rolled material 3, so-called crushing,
Rolled material 3 is thinned by pulling force to obtain desired thickness
And intermediate rolls 4, intermediate the work rolls 2.
Backup roll 5 is arranged with roll 4 interposed
You. For the backup roll 5, a force such as an oil pressure is used.
The rolling force is applied by the rolling reduction mechanism 6
Is through the contact surface between the backup roll 5 and the intermediate roll 4.
And transmitted to the intermediate roll 4 and transmitted to the intermediate roll 4.
Rolling force is applied to the intermediate roll 4, the work roll 2, and the work roll.
To the rolled material 3 through the contact surface between the roll 2 and the rolled material 3.
And the rolled material 3 undergoes plastic deformation due to the rolling force.
It will be the desired thickness. By the way, the roll width of the rolling rolls 2, 4, and 5 is
Because it is wider than the sheet width and the rolling force is applied,
Deforms. For example, in a work roll 2, a rolled material 3
The part outside the plate width is bent by the rolling force.
U. As a result, the end of the rolled material 3 is crushed and the cross section of the convex shape is obtained.
become. In order to prevent this,
Work roll bender 7
The work roll bending force Fw is applied to the end of the rolled material 3.
Prevents parts from being crushed. Similarly, the axis of the intermediate roll 5
Roll bending by middle roll vendor 8
Force F₁Add. Further, the intermediate roll shift 9 moves the intermediate roll 4 in the sheet width direction.
Move in the direction. By this movement, rolls 2, 4, 5 and rolling
By making the force applied to the material 3 asymmetric, the rolling material 3
Control the thickness of the plate. On the other hand, the energy added to the rolling mill 1 to perform the rolling
Energy is not only spent on the plastic deformation of the rolled material 3, but also on sound and vibration.
It becomes motion and heat. The energy converted into heat is the rolled material 3
And the temperature of the work roll 2 is reduced.
To raise. Due to this temperature rise, work roll 2
Expands and the roll diameter changes, but the roll diameter is generally
Deforms unevenly. Therefore, control the roll diameter uniformly.
For this reason, a plurality of nozzles (shown in the drawing)
Not supplied), and the cooling liquid is supplied to the work roll 2 through the nozzle.
Is installed. For moving the rolled material 3 on the axis of the work roll 2,
The speed control mechanism 11 composed of an electric motor etc. is connected.
You. The control for the rolling mill 1 is a draft control mechanism 6, a work roll
Vendor 7, Intermediate roll vendor 8, Intermediate roll shifter 9, Cooler
Control mechanism 10, speed control mechanism 11, etc.
Command generation mechanism 12 for generating an operation command for the
For the generating mechanism 12, the shape of the rolled material 3 is stored in advance.
Belongs to which type of setting pattern
To determine the similarity ratio with each setting pattern
Pattern recognition mechanism 13 that outputs similarity, the pattern recognition
A form that detects and outputs the thickness of the rolled material 3 to the mechanism 13
Shape detection mechanism 14, the shape detection mechanism 14 and the command generation mechanism 12
The storage mechanism 15 for storing the output and the information of the storage mechanism 15 are used.
Parameter of the pattern recognition mechanism 13 is changed by learning.
Learning mechanism 16 with the function of
Consists of Roll-down control mechanism 6, work roll bender 7, intermediate roll
Vendor 8, intermediate roll shifter 9, coolant control mechanism 10
And the first actuator control device with the speed control mechanism 11.
G6. And these mechanisms 6 to 11 are rolling mills
Perform different operations on. Command generation mechanism 1
2, pattern recognition mechanism 13 and storage mechanism 15
The control device G13 is configured. FIG. 3 shows a detailed view of the shape pattern recognition mechanism 13. form
Output of the shape detection mechanism 14 and the storage mechanism 15 (shape detection signal and shape
State recognition signals) are input cells 17, 18 of the pattern recognition mechanism 13.
And the input signal is converted to a function value in the input cell 17.
It is converted, output to the middle layer 19, and input to the middle layer 19.
The output of the force cell 17 is input to the cells 20, 21 of the intermediate layer 19.
The signal input to cell 20 at the output of input cell 17 is a weight function
23 in wl₁₁Multiplied and input to the adder 24.
The output of the module 18 is input to an adder 24 via a weight function 26.
The adder 24 adds the outputs of the weight functions 23 and 26, and
5 and a linear or non-linear function operation is performed by the function unit 25.
Then, it is output to the next intermediate layer 27. The cell 20 is heavy
It comprises only functions 23 and 26, an adder 24 and a function unit 25. Similarly, the output of input cells 17, 18 is input to cell 21,
The output of the input layer 17 is wl by the weight function 28_{twenty one}Doubled adder 24,
It is output to the next intermediate layer 27 via the function unit 25. The intermediate layer 27 has the same structure as the intermediate layer 19, and the input layer 1
The output of the middle layer 19 is used instead of the output of 7,18.
is there. Here, when the weights of the weight functions 23, 26, and 28 are represented by wkij,
wkij is the k-1th cell in the i-th cell of the k-th intermediate layer.
I-th in the middle layer of the eyes (however, when k = 1, the input cell)
Indicates the weight applied to the output. The signal input to the pattern recognition mechanism 13 as described above
Through the input cells 17, 18 and a plurality of intermediate layers 19, 27, 29,
Output in the form of removing the weight function and adder from the cells in the inter-layer
Output via the power layer 30. The input layer 31 is an input cell
Represents the sum of 17,18. The feature of this pattern recognition mechanism 13 is that it is a simple product-sum operation.
No repetitive operations such as corners and feedback, and
If each product-sum term in the hidden layer is implemented in hardware,
, And high-speed operation is possible. Following the output layer 30 of this pattern recognition mechanism,
Stores the command value for each actuator according to the turn
And directly access the command value closest to the output pattern.
It is also possible to instruct the tutor. In this method
Responsiveness is good, but control accuracy is slightly worse than the method described later.
It becomes. Next, the processing result of the pattern recognition mechanism 13 is shown in FIG.
It is applied to the rolling mill 1 through a logical processing mechanism. That is,
The output of the turn recognition mechanism 13 is provided to the command generation mechanism 12.
Is input to the manipulated variable determining means 32. Operation amount determination means 32
Now, among the multiple processing mechanisms prepared inside,
Select the most effective processing mechanism to process the
And output the operation amount. Using the result of the manipulated variable determination means 32
The command value calculation means 33 is a specific finger of each actuator.
Command, for example, a reduction command to the reduction control mechanism 6,
Generates an intermediate roll vendor command for the
You. FIG. 5 shows the configuration of the manipulated variable determining means 32.
You. The operation amount determination means 32 includes a shape detection mechanism 14, a pattern recognition
Upon receiving a signal from the mechanism 13, the control mechanism 141 is activated. System
The control mechanism 141 uses the knowledge base 36 according to the type of problem.
To determine the inference mechanism to be activated. As an example, this determination method is pre-programmed.
Have been. For example, the operator looks at a pattern-like waveform.
Assume that you are implementing ambiguous control. This operet
When configuring a system that incorporates data operations,
Upon receiving a signal from the turn recognition mechanism 13, the control mechanism 141
Activates the fuzzy inference mechanism 143 and activates the pattern recognition mechanism 13
Hand over the signal from Also, the production inference mechanism 142 changes the control policy.
Is necessary, for example, a failure or a large deviation even if control is performed.
Is activated by a failure signal or deviation signal, etc.
You. The script inference mechanism 146 is a sequence control type
Control from one sequence to another
Is shifted (by instructions from the upper control system, etc.)
Use this when predicting the state of the cans in advance and determining the next operation.
Used. In addition, the control mechanism 141 is in the form of each control mechanism 6, 7, 8, 9, 10, 11, etc.
Monitor the state and use the knowledge base for various inference mechanisms 142, 143,
Start 144,145,146. That is, the control mechanism 141
Production when legally required to reach a conclusion
Activate the inference mechanism 142, and if there is an
Activate the Zi inference mechanism 143 and ask
Activate the frame inference mechanism 144 for the
Network-related issues such as
The semantic net inference mechanism 145 is started for the
For problems where elephants are working in chronological order
Activate the script inference mechanism 146. Further, the control mechanism 141
Is an empirical problem that cannot be solved by the above various inference mechanisms.
Activate the optimization operation mechanism 111 to find the optimal solution,
Patterns can be memorized, features can be extracted and answers can be
Feature extraction answer mechanism 110 (Rumelha
rt type Niuro computer). Manipulated variable
The processing result of the deciding means 32 is transmitted to the command value meter via the control mechanism 141.
Output to the arithmetic unit. Figure 6 shows the structure of the knowledge base 36, which is the knowledge required for inference.
Is shown. Knowledge base 36 is used for experience of control experts
Inference of knowledge 106 input from outside based on syllogism
Production rules 147 for implementing, ambiguous information
Fuzzy rule 14 which is the knowledge to make inference based on
8, Knowledge that can be described in a certain framework such as the configuration of parts to be diagnosed
149, the relationship between parts and parts, and the common sense
Meaning network that is organized in a network form
150, when the diagnosis target proceeds in order
Script 151 that organizes and memorizes the work of
Classified as other knowledge 152 that cannot be described with knowledge 147 to 151
It is memorized. FIG. 7 is an explanatory diagram of the operation of the operation amount determining means 32. System
The processing of the control mechanism 141 includes the pattern recognition mechanism 13, the shape detection mechanism
14, Organize information from the storage mechanism 15 and use it for the following processing
Processing step 200, step 200 to convert to data that can be
Until the data prepared in step is exhausted, step 202
Step 201, the information collected in step 201
Report to determine the inference mechanism to be activated and the processing to be performed.
Disconnection step 202, and various inference mechanisms 142-146, feature extraction
Answer mechanism 110, optimization operation mechanism 111, and old
Executes modern control algorithms such as common control and multivariable control
The general control mechanism 203 performs the above steps.
End processing for executing resetting of flags necessary for
It comprises a management step 204. Here, the role of each inference processing mechanism will be described. Production
The inference mechanism 142 is a fragmentary
Using logical production rules
Suitable for assembly control. The fuzzy inference mechanism 142
Operate if the state of interest of the controlled object changes
Data can be quantified by moving the actuator a little.
Computers can handle ambiguous knowledge of operators who cannot
It is suitable to determine the manipulated variable by quantification as described above. The frame inference mechanism 144 describes the relationship between the control devices and the like.
Using the knowledge of the frame to be controlled,
When returning to the original state when the state of the
The amount of processing to perform operations based on the relationship between devices for each related device
Suitable to decide. The semantic network inference mechanism 145 is a piece of knowledge.
To organize and organize networks to create networks
Operation result of a specific actuator
The effect of
doing. The script inference mechanism 146 is used when a specific state occurs.
In order to infer the procedural knowledge of
Sequence control that must respond in order
Suitable for control. The feature extraction response mechanism 110 includes a pattern recognition mechanism 13,
Input patterns of the shape detection mechanism 14 and the storage mechanism 15 and the input
Inference mechanism 142-146 issued when pattern was input
If the output relationship is learned in advance, the inference mechanisms 142 to 146
Is faster and the same as inferring to determine output
The feature is that the result of can be output. Optimization operation mechanism 111
Means that the controlled object 1 usually has a strong nonlinearity,
If the operating point changes due to the cause, it is necessary to reset the operation.
In that case, the steepest gradient method, dynamic programming
, Linear programming, hill climbing, conjugate gradient method,
Is used for algorithms such as Hopfield-type neurocomputers
Optimized response for nonlinear controlled objects
Do. FIG. 8 is a diagram for explaining the operation of the production mechanism 142.
You. Production inference machine started by control mechanism 141
The structure 142 is an input from the control mechanism 141 to be stored in the memory at the time of startup.
Take out each piece of information stored in force processing 34 and input processing 34
If there is no pattern information in the memory,
End judgment machine that ends the processing of the duplication inference mechanism 142
Execute structure 35. Pattern extracted by end judgment mechanism 35
Using knowledge type and its confidence, knowledge base 36 or rule 1
At step 37, the type of the input pattern and the
Compare the assumptions of the rules. Using the comparison results,
Step 38 performs the next processing 39 if they match, and steps if they do not match.
Step 37 is executed. Step 39 is to enter
To the conclusion of the rule. Taking confidence at this time
The treatment is the mini-max theory, where the minimum or maximum
Replace with large value. Step 40 is the conclusion of the replaced rule.
If the logical part is an operation command, step 41 is followed and the conclusion part is not.
In the case of a match, perform step 37 to perform further inference.
Let go. When the conclusion part is the operation command, the process 41 calculates the command value.
To the means 33, the conclusion part and the certainty determined in the above processing step
Is output. FIG. 9 shows the command value calculation means 33. Command value calculation means 33
Are commands and inference results obtained by the manipulated variable determination means 32,
The memory 42 that stores the certainty factor and all the memory commands are processed.
Judge whether or not it has been processed, and if it has been processed, the command value
Step 43 for terminating calculation means 33, if not processed
If the actuator 7,8,9,10,11 such as the reduction control mechanism 6
Actuators obtained by various inferences
Calculate the center of gravity of the operation amount based on the degree of operation and confidence, and
New center of gravity by combining the centers of operation of actuators
From processing 44 which determines the corresponding actuator command
Be composed. By providing such a command value calculation means 33, various inferences can be made.
Mechanisms 142 to 146, feature extraction circuit mechanism 110, optimization operation mechanism 11
1, actuators individually required by general control mechanism 203
There is a feature that can handle the command to the unified. Fig. 10 shows the configuration of the input switching device 125 required for learning.
Is shown. Input switching device 125 is controlled by the learning mechanism
Output of the shape detection mechanism 14 using the switch mechanism 156
And one of the outputs of the learning mechanism 16 is output to the input layer 31.
is there. The state of the switch mechanism 156 in FIG.
Indicates the state to be performed. FIG. 11 shows the configuration of the learning mechanism 16. The learning mechanism 16
Force pattern generator 45, output pattern generator 47, output protrusion
It comprises a matching mechanism 46 and a learning control mechanism 48. Out
The force matching mechanism 46 outputs the output of the output layer 30 to the command generation mechanism 12.
Output o of distributor 139 for output to butt mechanism 46₁, o_i, o
_nAnd the output O of the output pattern generating mechanism 47_T1, O_Ti, O_TnDifference with
Is calculated by adders 161, 162, 163₁, e_i, e_nAsking
Therefore, it outputs to the learning control mechanism 48. The output o of the distributor 139
₁, o_i, o_nIs the output pattern recognition of the input pattern generation mechanism 47
Input layer 1 of mechanism 13 (Rumelhart type neurocomputer)
Occurs when 9 is input. At this time,
Turn generation mechanism 45 and output pattern generation mechanism 47 are learning control
Controlled by mechanism 48. Figure 12 shows the weight function w in the learning process._ij23 and learning control
The relationship of the mechanism 48 is shown. Deviation e which is the output of adder 161_kTo
Then, the learning control mechanism 48 constitutes the pattern recognition mechanism 13.
The value of the load function wij23 of the cell 20
Change the direction. FIG. 13 shows a processing outline 170 of the learning control mechanism 48 in FIG.
When the learning mechanism 16 is activated, the processing 170 of the learning control mechanism 48
Is started. The processing 170 is performed by the input pattern generation mechanism 45 and the output.
Activate the force pattern generation mechanism 47 and input the teacher signal
And preprocessing 171 to generate the desired output, deviation e_kValue, or
The following steps must be taken until the sum of the squares of
Step 172 to repeat steps 173, 174, 175, close to the output layer 30
Extract intermediate layers of interest from the intermediate layer toward the input layer 31 sequentially
Step 173, the cells of interest in the intermediate layer sequentially
Step 174, and the deviation e_kDirection in which
To change the weight function wij23 of the extracted cell
75 and step 176 to end the learning process
Consists of By providing such a learning mechanism,
New phenomena that did not occur, and
Once determined, there is a characteristic that can reflect that knowledge. FIG. 14 shows the configuration of the storage mechanism 15 of FIG. Memory
The structure 15 receives the outputs of the command generation mechanism 12 and the shape detection mechanism 14
Memory element 49, the contents of memory element 49
Memory elements 50 that are transferred afterwards, and
Memory element to which data is transferred and arrives after a specified time
51, and the contents of each memory element 49, 50, 51 are
Through an arithmetic mechanism 501 for differentiating and integrating
It is input to the turn recognition mechanism 13 and the learning mechanism 16. This storage mechanism 15 allows the shape detection mechanism 14 and command generation
The temporal change of the mechanism 12 can be considered, and for example, differentiation, integration, etc.
The movement will give. Figure 15 shows the effect of the nozzle on coolant control.
Since only a certain length is affected from the position of the
Here is a mechanism for recognizing patterns using nearby inputs. shape
The output of the detection mechanism 14 is input to the memory 52 of the pattern detection mechanism 13.
The signal input to the memory 52 is input through the gate circuit 53.
Input to the memory element 54 and input to the memory element 54
Input to memory elements 57 and 58 via gate circuits 55 and 56
When the gate circuits 53 and 56 are turned off, the gate circuit 55
Turns on and synchronizes with the clock to
To the memory 57, and after a certain time, the memory element 54
The signal reaches memory element 58 and the signal of memory element 57
Memory element 54, and in the next clock, memory elements 54, 57, 58
When the signal makes one cycle, gates 53 and 56 are turned on, and gate 55 is turned on.
Off and the contents of memory element 54 are stored in memory element 59.
The information of the memory elements 54, 57, 59 is input to the input layer 31.
You. By providing such a memory 52, pattern recognition can be performed.
Of the input layer 31, the intermediate layers 19, 27, 29, and the output layer 30 of the
This has the effect of greatly reducing the number. FIG. 16 shows the input pattern generating mechanism 45 and the output of the learning mechanism 16.
Using the controlled object simulator 60 for the pattern generation mechanism 47
Here is an example. The operator operates or operates the output pattern generation mechanism 47.
Is the shape pattern generated by the data.
It has the same function as the command generation mechanism 12 and is provided separately in the learning mechanism
Is input to the command generation mechanism 12
Generates various actuator commands according to the pattern,
The command is a control target provided in the input pattern generation mechanism 45.
Various activities that are input to the simulator 60 and that are controlled
The operation including the heaters 6, 7, 8, 9, 10, 11 and the rolling mill 1 is simulated.
If the response is poor, the command generation mechanism 12
Parameters for changing simulator 60 parameters
The output of the simulator 60 to be controlled is
Adjust it to the desired shape, and turn on the pattern recognition mechanism 13.
Force. The operation of the control method of the configuration described above will be described using a specific example.
It is described below. Neurocomputer constituting the pattern recognition mechanism 13
The initial value of the weight function wij28 for the middle layers 19, 27, 29 of
First, random number or appropriate value, for example, the value that the weight function can take
Set to half (0.5) of (assuming 0 to 1.0). At this time
In addition, for example, the input pattern generation mechanism 45 of FIG.
Even if a concave rolled material shape pattern is input, the output layer 30
The output of the output signal line 70 that is concave in force is 1
The output line 71 of the output layer 30 is convex.
The rate does not go to zero. Then, the output of the learning mechanism 16 corresponding to the output line 70 of the output layer 30 is output.
The output line 72 of the force pattern generating mechanism 47 is connected to 1 and the output line 71.
The output of the output line 73 of the corresponding output pattern generating mechanism 47 is set to zero.
Output to In response to these outputs, the output matching mechanism 46
Is the ideal output (output pattern generation mechanism 47) and putter
The learning control mechanism 48 receives the deviation of the output of the
The deviation reduces the magnitude of the load function wij of the turn recognition mechanism 13.
It is changed in the direction to decrease in proportion to the magnitude of the deviation.
A steepest gradient method is a typical example of this algorithm. Change the weight of the load function sequentially according to the processing in Fig. 13.
And the deviation e in FIG._kWhen the sum of the squares of
The operation of the learning mechanism 16 ends. After learning is completed, the output of the input pattern generation mechanism 45 in FIG.
The same waveform (shape detection pattern) as the pattern (shape setting pattern)
Is input from the shape detection mechanism 14 in FIG.
The pattern recognition mechanism 13 outputs 1 from the output line 70 of the output layer 30.
And outputs zero from the output line 71 of the output layer 30. Next, the waveform shown in Fig. 18, which is said to be convex, is input.
If the learning is not completed, pattern recognition
The output of the output line 71 expressing the convex shape of the mechanism 13 is 1, and the other
Does not result in a pattern in which the output 70 becomes zero. So the aforementioned
As shown in the figure, a typical convex pattern
The output of the force pattern generator 47 corresponds to the output of the output lines 71 and 70.
The corresponding values are set to 1,0 respectively. The learning mechanism 16
By changing the weight function wij, when the learning is completed,
The convex waveform of FIG. 18 is input to the turn recognition mechanism 13.
17, the output line 71 of the output layer 30 is 1 and the output line 70 is zero.
become. As a result, the waveform of FIG.
The output is output from the output layer 30 as described above.
The output line 71 indicates that the input waveform is a convex waveform.
The degree (proportion) of similarity to the waveform is 40% confidence
Output and at the same time indicate a concave waveform
It is output from the output line 70 as a certainty factor of 50%. like this
In addition, a pattern recognition mechanism 13 composed of a neural network
From the shape detection pattern is similar to the shape setting pattern
A numerical value indicating the percentage is output. Fig. 20 shows the shape of the rolled material considering the temporal change of the rolled material.
Is shown. The state immediately below the work roll 2 of the rolling mill is t₀And that
Is x₀It is. Set the sampling period of the computer to T₀When
Then T₀Seconds before t₁The height of the thickness at the time is x₁, T₀x_Two
Seconds before t_TwoThe height of the thickness at the time is x_Two, ... That is, t_TwoAt the height x_TwoIs input to the storage unit 15,
It is stored in the memory element 49 of FIG. Next sampling
Time t₁Height x₁Is input to the storage mechanism 15,
At that timing, the data x of the memory element 49_TwoIs the memory element
50 and the contents of memory element 49 are x₁To
Rewritten. On the other hand, the operation mechanism 510 is
Perform various calculations using the contents of 9,50. For example, if the derivative is
When needed (x_Two−x₁) / T₀, When an integrator is needed (x
₁+ X_Two) × T₀The following operation may be performed. That is, differentiation
The container can determine the rate of change of the shape.
The region recognition mechanism 13 can improve responsiveness to a change. On the other hand, integrators have a function of removing noise, etc.
The feature of can be given. These differentiators, integrators and
Functions such as proportional elements without time and time elements
Can be provided to the recognition mechanism 13. In addition, storage
Use the data stored in step 15 for learning as needed
It can be used for the input pattern generation mechanism 45. By the way, as shown in FIG.₀Rolling at the time
X the thickness of the rolled material in the roll axis direction of the mill⁰ ₀, X⁰ ₁… X⁰ _n-1, X⁰ _n
And T at the same position₀(Sampling cycle) Previous plate
Thick state, X⁰ ₀, X⁰ ₁… X⁰ _n-1, X⁰ _nThen, at some point t₀
Then, the memory elements 54, 57, and 58 in FIG.⁰ _n, X⁰ _n-1…, X⁰
₀Is stored. The memory element 59 includes the memory mechanism 1 described above.
Since the same operation as 5 is performed, T₀Time t before time t₁of
X is the data¹ ₀, X¹ ₁… X¹ _nIs stored in memory element 59 and elsewhere
ing. Figure 22 shows the production rules or fuzzy rules.
An example is shown (production rule 47, fuzzy
Corresponding to the rule 48). Output as concave 50% confidence by pattern recognition mechanism 13
(Numerical value indicating the similarity ratio)
And matches the concave rule 80. The result
As a result, to weaken the vendor, Small) Rule 81 is obtained
You. On the other hand, with a convex confidence of 40%,
As a result, an operation amount (large in degree) that strengthens the vendor can be obtained. As a result, as shown in FIG.
As a result of matching with control rules, the operation amount of the vendor is convinced that it is convex
Since the degree is 50%, it is represented by the area of the shaded portion of B. Meanwhile, concave
Is 40% and S is 40% certainty.
It is the area of the shaded portion of S. Next, the command generation mechanism 12 is shaded
65%, which is the value of the center of gravity C obtained by combining the centers of gravity A and B of
Operation amount. Next, the influence of the actuator, such as coolant control,
However, unlike vendors and shifters,
As shown, the waveforms of FIG.
To memorize. A part of the waveform stored in the memory element (24th
(A part of FIG. A) is a pattern recognition mechanism 13, a command generation mechanism 1
2 and one nozzle A of the coolant control device 10
Controlling the amount of coolant by controlling
It is flattened. Now, x in FIG. 21 corresponding to nozzle A_0n-1Both sides of
x_0n, x_0n-2X when compared to the value of_0n-1Is larger,
As shown in FIG. 22, the conclusion 85 of the center size is obtained. one
Who, x_0n-1, x_1n-1X as the relation_0n-1, x_1n-1Is positive
If x_n-1Is increasing, so the derivative is positive.
It matches the part 86, and as a result, the coolant is turned on. So much
The degree is large (B). As a result, x_0n-1, x_1n-1Is almost strange
It does not change. When the control of the nozzle A is completed, the memory elements 54, 57,
The contents of 58 and 59 are shifted one by one. As a result,
The waveform input to the turn recognition mechanism 13 is shown in FIG.
The area indicated by b is input, and the processes 13 and 12 are performed.
One nozzle B of the runt control mechanism 10 is controlled. When the processing is performed in this manner, the pattern A shown in FIG.
To pattern B, and furthermore,
24, the waveform of FIG. 24 (a) is reproduced in the memory 52.
You. Last time, the pattern of FIG.
The contents of the memory element 54 in FIG.
Moved to the memory element 59 and the wave of the shape detection mechanism 14
Remember the shape. Further, between the memory 52 and the input layer 31, the functions shown in FIG.
If the arithmetic mechanism 510 is provided, it is possible to control the rate of change of the waveform, etc.
This is obvious from FIG. Next, the pattern recognition mechanism 13 learns a reference pattern.
Describe the learning method. The waveforms 62 and 63 in FIG. 18 are converted to the input pattern generation mechanism in FIG.
Generated at 45 and output to input layer 31. This pattern is
Writing to the memory of the force pattern generation mechanism 45, or
The pattern stored in the storage mechanism 15 is used. Input layer
Are input to the output layer 30 via the intermediate layers 19,.
Appear as output. At this time, the weight function ω of the hidden layer_kij
Is an initial value, and the input from the output pattern generating mechanism 47 is
A pattern recognizer corresponding to the output of the pattern generator 45
The pattern that you want to output from structure 13 (for example, input pattern
Is a standard pattern, and the output end of the output layer 30 is
When one child is assigned to the standard pattern, the assigned
The output terminal becomes 1 and the other terminals become zero.
Turn) is input to the butting mechanism 46. Learning is not complete
The output pattern of the output layer 30 and the output pattern
The waveform of the raw mechanism 47 is different. As a result, the butting mechanism
The output of 46 outputs according to the degree of the pattern difference. This
Calculate the mean of the deviation and the square mean of the deviation
Tram etc. are found. Intermediate near the output layer according to the above deviation
From the layer 27 to the intermediate layer 19 close to the input layer 31, the weight function w
Change kij. There are various ways to change the weight function wkij
However, if the above deviation is minimized,
In such an optimization problem, for example, the steepest gradient method is used. Ingredient
As a concrete method, the weight function wkij of interest
In the direction of the deviation value.
Moving the weight function value wkij in the decreasing direction
In addition, the movement amount is large when the change in the deviation value is small,
On the other hand, when the change in the deviation value is large, the moving amount is reduced.
Also, change the weight function wkij of the intermediate layer 19 closest to the input layer.
Is completed, the deviation value of the butting mechanism 46 is checked again.
Learning ends when the value falls within the allowable error range.
You. This control is performed by the learning control mechanism 48. Note that this
Pattern recognizer that uses learned results for pattern discrimination
Structure 13 explains why patterns can be identified, and why learning is successful
Although the behavior of whether to go is not clear,
Number is larger than the number of inputs and outputs, and the value
There is a lot of freedom, and even if the value is slightly out of order,
It can be said that a good recognition result can be obtained even if it is stored.
Have been done. On the other hand, the input pattern generation mechanism 45 and the output pattern generation
What pattern should be used for the raw mechanism 47
Or, there is a very difficult aspect. Fortunately, the operation of the controlled object 1
Operating model near a certain operating point leads to accurate model
In the field of control theory, the theory of system identification is
Established. However, non-linear targets are strong in all operating regions
Is difficult to model. Therefore, a model is created in a specific operation area and control is performed.
The relationship between the input and response of the control system
It is obtained by simulation and used as learning data.
This procedure is repeated for all operating areas of the control system.
Move and seek the optimal modeling and control commands at each moment
Let them learn. That is, the control target simulator 60 shown in FIG.
Adjust parameters to accurately simulate at specific operating points
Data 60 is operated. After this, the control target is a typical putter
Input pattern generation mechanism 47, parameter
Data adjustment mechanism 61, controlled object simulator 60, command generation mechanism 12
And outputs the outputs of these processes 47 and 60 to the learning mechanism 16 respectively.
Output pattern and input pattern. Such a control method is applicable to the pattern recognition mechanism.
Abstraction of waveform and control including ambiguity by control mechanism
it can. In addition, using a rolling mill system as a specific example of the present invention
The embodiment has been described.
Eta 6, 7, 8, 9, 10, 11 limited to rolling mill system
It is not necessary to perform
It is self-evident that it can be applied to controllers and controllers. For example,
Like a train operation management system, a train schedule pattern
Recognized and delayed columns according to various diamond rearrangement rules
It can be used to control systems that return cars to normal
You. In other words, the train operation is represented by a diagram,
The delayed features are extracted by the recognition mechanism 13. Next, its features
The inference mechanism based on the quantity, for example, overtakes trains at stations.
Create a diagram using various rules such as The inference machine
The command value calculation mechanism 33 receives the results of the
Generate a rotation command. The train, which is an actuator,
Operate according to instructions.

【The invention's effect】

According to the present invention, a combination of a plurality of signals relating to a controlled object is recognized as a detection pattern, and a numerical value indicating a rate at which the detection pattern is similar to the set pattern is input to the inference mechanism to obtain an operation amount. Therefore, interference between the control rules in the inference mechanism is significantly reduced, and the accuracy of determining the operation amount of the operation means can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention, FIG. 2 is an embodiment in which the present invention is applied to a rolling mill control system, FIG. 3 is a diagram of a pattern recognition mechanism, FIG. 5 is a block diagram of the operation amount determining means, FIG. 6 is a knowledge base block diagram, FIG.
FIG. 8 is an explanatory diagram of the operation of the manipulated variable determining means, FIG. 8 is an explanatory diagram of the operation of the production mechanism, FIG. 9 is a configuration diagram of the command value calculation means, FIG. 10 is a configuration of the input switching device, and FIG. FIG. 12 is a diagram showing the relationship between the learning control mechanism and the weight function of the node, FIG. 13 is a basic processing diagram of the learning control mechanism, FIG. 14 is a configuration diagram of the storage mechanism, and FIG. 15 is a diagram of the pattern recognition mechanism. , 16th
The figure shows the configuration when the learning mechanism is equipped with a simulator.
FIG. 18 is an explanatory diagram of the operation of the pattern recognition mechanism, FIG. 18 is an example of an input pattern, FIG.
FIG. 20, FIG. 21 is an explanatory diagram of a temporal change of a rolled material, FIG. 22 is a diagram showing an example of a production rule and a fuzzy rule, FIG. 23 is an explanatory diagram of a method of converting a similarity into an operation amount,
FIG. 24 is an explanatory diagram of the processing status of the input waveform. 1 ... controlled object, 13 ... pattern recognition mechanism, 12 ... command generation mechanism, 16 ... learning mechanism, 19, 27, 29 ... middle layer, 30
... output layer, 31 ... input layer.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Tetsu Hattori 5-2-1 Omikacho, Hitachi City, Ibaraki Prefecture Inside the Omika Plant, Hitachi, Ltd. (72) Inventor Masaaki Nakajima 5-2-2 Omikacho, Hitachi City, Ibaraki Prefecture No. 1 Inside the Omika Plant of Hitachi, Ltd. (56) References JP-A-61-180304 (JP, A) JP-A-63-273014 (JP, A) Shigemi Osada, two others, "Principles of neurocomputers and Application to Robot Control "FUJITSU, Ohmsha, June 10, 1988, Vol. 39, No. 3, p. 175-184 Norio Inaba, "Using Neural Networks for Pattern Recognition, Signal Processing, and Knowledge Processing," Nikkei Electronics, August 10, 1987, No. 427, p. 115-124 Kunihiko Fukushima, "Pattern Recognition and Robot Vision", Optical Technology Contact, 1986, Vol. 24, No. 5, p. 387-394

Claims (20)

(57) [Claims] In a control system having an operation means for operating a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detected pattern is compared with a preset setting pattern. And
A numerical value indicating the ratio of the detection pattern similar to the setting pattern is obtained by using a neural network, and the numerical value indicating the similarity ratio and the inference using the control rule predetermined for the operating means are performed. A control method using a neural network, wherein an operation amount of the operation means is obtained. 2. A control system having a plurality of operation means for operating a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detection pattern and a preset setting pattern are recognized. Comparing, using a neural network to determine a numerical value indicating the ratio of the detection pattern is similar to the set pattern,
A neural network which performs inference using a numerical value indicating the similarity ratio and a control rule predetermined for each of the plurality of operation means to obtain the operation amounts of the plurality of operation means. Control method using. 3. A control system having an operation means for operating a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detection pattern and a plurality of preset setting patterns are recognized. Are compared with each other, a numerical value indicating a ratio of the detection pattern being similar for each of the setting patterns is obtained by using a neural network, and a numerical value indicating the similarity ratio of each of the setting patterns and predetermined in the operation means. A control method using a neural network, wherein an inference using a control rule is executed to obtain an operation amount of the operation means. 4. A control system having a plurality of operation means for operating an object to be controlled, wherein a combination of a plurality of detection signals relating to the object to be controlled is recognized as a detection pattern, and the detection pattern and a plurality of predetermined presets are recognized. Each of the setting patterns is compared with each other, a numerical value indicating a rate at which the detection pattern is similar for each of the setting patterns is obtained using a neural network, a numerical value indicating the similarity rate for each of the setting patterns, and the plurality of operations are performed. A control method using a neural network, wherein an inference using a control rule predetermined for each means is executed to obtain an operation amount of each of the plurality of operation means. 5. A control system having a plurality of operation means for executing different operations on a controlled object, wherein a combination of a plurality of detection signals relating to the controlled object is recognized as a detection pattern, and the detected pattern is set in advance. A plurality of setting patterns are compared with each other, a numerical value indicating a rate at which the detection pattern is similar for each of the setting patterns is obtained using a neural network, and a numerical value indicating the similarity rate for each of the setting patterns And performing an inference using a control rule predetermined for each of the plurality of operation means to obtain an operation amount of a different operation performed by each of the plurality of operation means. Control method. 6. A control system having operating means for operating a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detected pattern is compared with a preset setting pattern. And
A numerical value indicating a ratio of the detection pattern similar to the setting pattern is obtained by using a neural network, and an operation amount of the operation unit is obtained by fuzzy inference based on the numerical value indicating the similarity ratio. A control method using a neural network. 7. A control system having a plurality of operation means for operating an object to be controlled, a combination of a plurality of detection signals related to the object to be controlled is recognized as a detection pattern, and the detection pattern and a preset setting pattern are recognized. Are compared with each other, a numerical value indicating a rate at which the detection pattern is similar to the set pattern is obtained using a neural network, and the operation amounts of the plurality of operating means are determined by fuzzy inference based on the numerical value indicating the similarity rate. A control method using a neural network, wherein 8. A control system having an operation means for operating a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detection pattern and a plurality of preset setting patterns are recognized. Are compared with each other, a numerical value indicating the ratio of the detected pattern being similar for each of the setting patterns is obtained using a neural network, and the operation is performed by fuzzy inference based on the numerical value indicating the similarity ratio for each of the setting patterns. A control method using a neural network, wherein an operation amount of the means is obtained. 9. A control system having a plurality of operation means for operating an object to be controlled, a combination of a plurality of detection signals relating to the object to be controlled is recognized as a detection pattern, and the detection pattern and a plurality of predetermined presets are recognized. Each of the setting patterns is compared with each other, a numerical value indicating a rate at which the detection pattern is similar for each of the setting patterns is obtained using a neural network, and fuzzy inference is performed based on the numerical value indicating the similarity rate for each of the setting patterns. A control method using a neural network, wherein an operation amount of each of the plurality of operation means is obtained. 10. A control system having a plurality of operation means for executing different operations on a controlled object, a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detection pattern is set in advance. A plurality of setting patterns are compared with each other, a numerical value indicating a rate at which the detection pattern is similar for each of the setting patterns is obtained using a neural network, and a numerical value indicating the similarity rate for each of the setting patterns A control method using a neural network, wherein an operation amount of a different operation performed by each of the plurality of operation means is obtained by fuzzy inference based on the following. 11. An operation means for operating an object to be controlled, and a combination of a plurality of detection signals relating to the object to be controlled is recognized as a detection pattern, and the detection pattern is compared with a preset setting pattern. Is a feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio similar to the setting pattern, and the operation is performed by inference using a numerical value indicating the similarity ratio and a control rule predetermined in the operating means. Means for obtaining an operation amount of the means and outputting an operation command to the operation means. 12. Recognizing a combination of a plurality of operation means for operating a controlled object and a plurality of detection signals relating to the controlled object as a detection pattern, comparing the detection pattern with a preset setting pattern, A feature amount extraction unit having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to the set pattern; a numerical value indicating the similarity ratio and a control rule predetermined for each of the plurality of operation units; Means for obtaining an operation amount for each of the plurality of operation means by inference used and outputting an operation command for each of the plurality of operation means. 13. An operation means for operating a controlled object, and a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detected pattern is compared with a plurality of preset setting patterns. A feature amount extracting unit having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to each of the setting patterns; a numerical value indicating a similarity ratio at each of the setting patterns; and the operating unit. A means for obtaining an operation amount of the operation means by inference using a control rule and outputting an operation command to the operation means. 14. Recognizing a combination of a plurality of operation means for operating a controlled object and a plurality of detection signals related to the controlled object as a detection pattern, and recognizing the detected pattern and a plurality of preset setting patterns. Compare each,
A feature amount extracting unit having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar for each of the setting patterns; a numerical value indicating a similarity ratio for each of the setting patterns; Means for obtaining an operation amount for each of the plurality of operation means by inference using a predetermined control rule, and outputting an operation command for each of the plurality of operation means. apparatus. 15. A plurality of operation means for executing different operations on a controlled object, and a combination of a plurality of detection signals related to the controlled object are recognized as a detection pattern. And a feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to each of the setting patterns, and a numerical value indicating the similarity ratio at each of the setting patterns Means for obtaining an operation amount for each of the plurality of operation means by inference using a control rule predetermined for each of the plurality of operation means, and outputting an operation command for each of the plurality of operation means. A control device using a neural network. 16. An operation means for operating an object to be controlled and a combination of a plurality of detection signals relating to the object to be controlled are recognized as a detection pattern, and the detection pattern is compared with a preset setting pattern. A feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio similar to the setting pattern; and an operating amount of the operating means being obtained by fuzzy inference based on the numerical value indicating the similarity ratio and operated by the operating means. Means for outputting a command, the control device using a neural network. 17. Recognizing a combination of a plurality of operating means for operating a controlled object and a plurality of detection signals related to the controlled object as a detection pattern, comparing the detection pattern with a preset setting pattern, A feature amount extraction unit having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to the setting pattern, and obtaining an operation amount for each of the plurality of operation units by fuzzy inference based on the numerical value indicating the similarity ratio And a means for outputting an operation command for each of the plurality of operation means. 18. An operation means for operating an object to be controlled, and a combination of a plurality of detection signals relating to the object to be controlled is recognized as a detection pattern, and the detection pattern is compared with a plurality of preset patterns, respectively. A feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to each of the setting patterns; and a fuzzy inference based on a numerical value indicating a similarity ratio at each of the setting patterns. Means for obtaining an operation amount of the means and outputting an operation command to the operation means. 19. Recognizing a combination of a plurality of operation means for operating a controlled object and a plurality of detection signals related to the controlled object as a detection pattern, and recognizing the detection pattern and a plurality of preset setting patterns. Compare each,
A feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar for each of the setting patterns; and the plurality of operations by fuzzy inference based on a numerical value indicating a similarity ratio for each of the setting patterns. A means for obtaining an operation amount for each means and outputting an operation command for each of the plurality of operation means. 20. A plurality of operation means for performing different operations on a controlled object, and a combination of a plurality of detection signals related to the controlled object is recognized as a detection pattern, and the detection pattern and a plurality of preset detection patterns are recognized. And a feature amount extracting means having a neural network for obtaining a numerical value indicating a ratio at which the detection pattern is similar to each of the setting patterns, and a numerical value indicating the similarity ratio at each of the setting patterns Means for obtaining an operation amount for each of the plurality of operation means by fuzzy inference based on the above and outputting an operation command for each of the plurality of operation means.