JPH0470906A - Control system and optimization deciding device - Google Patents

Abstract

PURPOSE:To improve the controllability by quantizing the perception and the experiences of an expert operator, deciding the manipulated variable of an actuator, and also obtaining an operating point that is changed by the control. CONSTITUTION:An optimization deciding device 5 is provided with a feature extracting means 8 which extracts the features out of the output of a sensor 3 and a will deciding mechanism 9 which decides a control policy, a control parameter, a control algorithm, etc., with the outputs of the mechanism 8 and the sensor 3 and based on the knowledge stored previously. Furthermore a thinking mechanism 10 is added to infer a will based on the conclusion obtained by the mechanism 9 because the mechanism 9 takes much time to complete an arithmetic operation together with a self-control mechanism 11 which decides a control executing command of a control mechanism 6 or changes a control command with an instruction of a man-machine interface 7. Then the manipulated variable of an actuator 2 to be set by the assurance of a characteristic pattern is decided by a fuzzy inference, an operating point is stored as an operating command to an actuator 2, and the pertinent command by learning. Thus the satisfactory control performance is ensured.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention provides a control system for controlling a controlled object operated by a plurality of actuators, in which the operation of the controlled object and the plurality of actuators is comprehensively judged, and individual The present invention relates to a control system and an optimality determination device that determine the optimal control amount of an actuator.

[Conventional technology]

Conventional control devices perform control using multiple force signals that represent the operating status of the controlled object, but with a method that controls by looking at individual outputs, control becomes local and the entire system is affected. The disadvantage is that it cannot take into account the optimality of

Therefore, recently there is a tendency to use a plurality of signals to find overall optimality. The operation of the conventional control device and method will be explained using a rolling mill system that is complex and cannot be controlled by a single control device.

A rolling mill is a system that obtains a steel material of a desired thickness by controlling the distance between opposing rolls and the tension applied to the rolled material. However, flat steel materials cannot be obtained due to thermal deformation due to heat loss generated during rolling and roll deformation due to mechanical deformation, etc. Therefore, shape control has been developed to obtain flat properties.

However, because the physical characteristics of rolling change significantly due to various factors, even if a control model is created near a specific operating point, the model often differs from the actual operation of the rolling mill. It ends up. Furthermore, the operating point of feedback control moves, and control margin is lost. For this reason, feedback control, which would produce good results if the model was accurate, could not fully demonstrate its ability, and there was a problem in that it could not surpass the skills of a skilled operator who operates based on intuition and experience.

[Problem to be solved by the invention]

The above-mentioned conventional technology does not take into account the point of making use of the know-how of a skilled operator, and also has problems in control performance, such as the operating point being changed by control and the control margin being lost.

An object of the present invention is to provide a control system and an optimality determination device that incorporate the know-how of a skilled operator, are highly expandable, and effectively utilize information on operating points moved by control.

[Means to solve the problem]

The above object is achieved by quantifying the intuition and experience of a skilled operator, determining the operation amount of the actuator, and determining the operating point changed by control.

[Effect]

Skilled operators extract characteristic patterns from control variables and perform vague operations (fuzzy control). Similarly,
The reliability of a characteristic pattern is obtained by calculating the sum of products of the control amount and passing the result through a nonlinear circuit, and the operation amount of the actuator is determined by fuzzy inference based on the reliability of each characteristic pattern. In addition, with an optimality determination mechanism that stores the operating point moved by the control as an operation command for the actuator and generates the operation command through learning, the control system operates like a skilled operator, resulting in good control performance. can be obtained.

Further, even if the know-how of a skilled operator is stored as control knowledge and the control is performed using the knowledge, the same control performance as described above can be obtained.

〔Example〕

An embodiment of the control system of the present invention is shown in FIG. 1 below.

The controlled object includes one or more actuators 2 and a sensor 3 that detects the operating states of the controlled object 1 and the actuators 2. The output of the sensor 3 is transmitted to a control and diagnostic mechanism 4.
and is controlled to generate a control execution command for each actuator.

The control and diagnosis mechanism 4 includes an optimality determination device 5 which receives the output of the sensor 3 and generates a command to the control mechanism 6, and a control device which generates a control execution command for the actuator 2 using the command and the output from the sensor 3. It consists of a mechanism 6 and a man-machine interface 7 that inputs information on the host computer or operator operations and outputs diagnostic results.

The optimality determination bag [5 includes a feature extraction mechanism 8 that extracts features from the output of the sensor 3, and a control policy based on pre-stored knowledge using the feature extraction mechanism 8 and the output of the sensor 3. A decision-making mechanism 9 that determines control parameters, control algorithms, etc., a thinking mechanism 10 that makes analogies based on the conclusion obtained in advance by the decision-making mechanism 9 because the calculations of the decision-making mechanism 9 take time, and It is comprised of a self-adjustment mechanism 11 that judges a control execution command from the control mechanism 6 or changes a control command according to an instruction from the man-machine interface 7.

On the other hand, the conventional system does not have the optimality determination device 5,
Generally, the information from the sensor 3 is transmitted to the control mechanism 6, resulting in a system that controls each actuator. For this reason, the optimization of the entire system could not be achieved, but by adding the optimality determining device 5 that determines the optimality of the entire system as in the present invention, it is possible to determine the optimality of the entire system. The determination device has the effect of being able to flexibly respond to changes in the actuator and changes in the controlled object.

An embodiment in which the present invention is applied to a rolling mill control system will be described below with reference to FIG.

The rolling mill system of controlled object 1 is configured to roll a rolled material 23 sandwiched between a pair of opposing work rolls 22 by using a rolling force acting between the work rolls 22 and a tension force acting on the rolled material 23.
The rolled material is made thinner by crushing and pulling force,
To obtain a desired thickness, an intermediate roll 24 is placed between the work roll 22 and a backup roll 25 is placed between the intermediate roll 24. The backup roll 25 is provided with a reduction control mechanism 26 that uses force such as hydraulic pressure.
The rolling force is applied to the intermediate roll 24 through the contact surface between the backup roll 25 and the intermediate roll 24, and the rolling force transmitted to the intermediate roll 24 is applied to the intermediate roll 24 and the work roll 22. , and is transmitted to the rolled material 23 through the contact surface between the work roll 22 and the rolled material 23, and the rolled material 23 undergoes plastic deformation due to the rolling force,
The desired thickness is achieved. Work roll 22, intermediate roll 24
, the backup roll 25 is also called a rolling roll.

By the way, since the roll widths of the rolling rolls 22, 24, and 25 are wider than the rolling material 23 in the liquid, and rolling force is applied thereto, the rolls are deformed. For example, in the work roll 22, the portion of the rolled material that is out of the liquid is bent by the rolling force.

As a result, the ends of the rolled material 23 are crushed and have a convex cross-sectional shape. In order to prevent this, the work roll bender 27 is moved in the direction in which the distance between the work rolls 22 is widened relative to the axis of the work rolls 22.
A work roll pending force Fw is applied to prevent the ends of the rolled material 23 from being crushed. Similarly, an intermediate roll pending force Fl is applied to the shaft of the intermediate roll 24 by an intermediate roll bender 28.

Further, the intermediate roll shift 29 moves the intermediate roll 24 in the direction of the liquid. This movement causes the rolls 22, 24
．． 25 makes the force applied to the rolled material 23 asymmetrical, thereby controlling the shape of the plate thickness of the rolled material.

On the other hand, the energy applied to the rolling mill for rolling is not only used for plastic deformation of the rolled material 23 but also becomes sound, vibration, and heat. This energy changed into heat is transferred to the rolled material 23.
and increases the temperature of the work roll 22. Due to this temperature rise, the work roll 2
2 expands and the roll diameter changes, but the roll diameter generally changes non-uniformly. Therefore, in order to uniformly control the roll diameter, a coolant control mechanism 3o is installed that applies cooling liquid to the work roll 22 through a plurality of nozzles (not shown) arranged in the direction of the liquid and the nozzles. .

A speed control mechanism 31 composed of an electric motor or the like for moving the rolled material 23 is connected to the shaft of the work roll 22 .

The control system for the rolling mill includes the rolling control mechanism 26;
A command generation mechanism 32 that generates operation commands for actuators such as the work roll bender 27, intermediate roll bender 28, intermediate roll shift 29, coolant control mechanism 3o, and speed control mechanism 31;
A pattern recognition mechanism 33 that determines which type of pattern the shape of the rolled material 23 belongs to among a plurality of pre-stored patterns and outputs the reliability of the pattern; Detects the plate pressure shape of
a shape detection mechanism 34 that applies pressure, the shape detection mechanism 34;
It is comprised of a storage mechanism 35 that stores the output of the command generation mechanism 32, and a learning mechanism 36 that uses the information in the storage mechanism 35 to change the parameters of the pattern recognition mechanism 33 through learning. Note that the controlled object 1 is a work roll 22,
It is composed of an intermediate roll 24 and a backup roll 25.

The actuator 2 includes a reduction control mechanism 26, a work roll bender 27, an intermediate roll bender 28, an intermediate roll shift 29, a coolant control mechanism 30, and a speed control mechanism 31.

The sensor 3 includes a shape detection mechanism 34 and, although not shown here, a speed detection mechanism, a rolling load detection mechanism, various position detection mechanisms, tension detection, plate thickness detection, and other detectors.

The control and diagnosis mechanism 4 includes a storage mechanism 35 and a command generation mechanism 32.
, a pattern recognition mechanism 33, and a learning mechanism 36.

FIG. 3 shows a detailed diagram of the pattern recognition mechanism 33. The pattern recognition mechanism 33 includes an input layer 51 including input cells 37, 38, an intermediate layer 39 connected to the input layer 51 and including cells 40, 41, and an intermediate layer 47 connected in series to the intermediate layer 39. .49 and an output layer 50. Cell 4o has weighting functions 43.46. Adder 44 and function unit 45
The cell 41 includes a weighting function 48, an adder 49,
and a function unit 50. Shape detection mechanism 34
and the output of the storage mechanism 35 is the pattern recognition mechanism 33.
is input into the input cell 37.38, and the input cell 37.3
8, the input signal is converted into a function value and output to the intermediate layer 39, and the input cell 37.3 input to the intermediate layer 39
The output of 8 is input to the cell 40.41 of the intermediate layer 39.

The signal inputted to the cell 40 as the output of the input cell 37 is multiplied by 0110 by the weighting function 43 and inputted to the adder 44, and the output of the input cell 38 is inputted to the adder 44 via the weighting function 46. . The adder 44 uses the weighting function 43
．． The outputs of 46 are added and output to a function unit 45. The function unit 45 performs a linear or nonlinear function operation on the input signal, and outputs the result to the intermediate layer 47 at the next stage.

Similarly, the outputs of the input cells 37 and 38 are input to the cell 41, and the output of the input layer 37 is multiplied by 0121 by the weighting function 48 and output to the next intermediate layer 47 via the adder 49 and the function unit 50. .

The intermediate layer 47 has the same structure as the intermediate layer 39, and the output of the intermediate layer 39 is used instead of the output of the input layer 37, 38.

Here, when the weight of the weighting functions 43, 46.48 is expressed as ω'IJ, ω is at the i-th cell of the i-th intermediate layer,
Indicates the weight to be applied to the fifth output of the k-1th intermediate layer (input cell only when k=1).

The signals input to the pattern recognition mechanism 33 as described above are transmitted to the input cells 37, 38, . Multi-stage intermediate layer 39.47.4
9, and an output layer 50 which is the same as the intermediate layer cell without the weighting function and adder. Note that the input layer 51 represents a collection of all input cells 37 and 38.

The feature of this pattern recognition mechanism 33 is that it only requires a simple product-sum operation, and there is no repeated operation such as feedback, and that each product-sum term in the intermediate layer can be processed in parallel when implemented in hardware. It is possible to perform high-speed calculations.

It is also possible to store command values for each actuator in accordance with each output pattern in advance in the output layer 50 of this pattern recognition mechanism, and to directly instruct the actuators with the command value closest to the output pattern. Although this method has good responsiveness, the control accuracy is slightly worse than the method described below.

Next, the processing result of the pattern recognition mechanism 33 is applied to the rolling mill system, which is the controlled object 1, via the command generation mechanism 32 shown in FIG. That is, the output of the pattern recognition mechanism 33 is determined by the operation amount determining means 5 provided in the command generation mechanism 32.
2 is input. The manipulated variable determining means 52 selects the most effective processing mechanism for processing the input signal from among the plurality of internally prepared processing mechanisms, executes the process, and outputs the manipulated variable. Using the results of the operation amount determining means, the command value calculation means generates specific command values for each actuator, such as a roll-down command for the roll-down control mechanism 26, an intermediate roll bender command for the intermediate roll bender 28, and the like. Note that the output of the shape detection mechanism 34 is directly input to the command generation mechanism 32 without going through the pattern recognition mechanism 33, and is processed by the optimal processing mechanism among the plurality of processing mechanisms prepared inside. It is also possible. However, in this case, when using various inference mechanisms, it is necessary to enrich the knowledge base in order to sufficiently reflect the operating methods of specialized operators.

FIG. 5 shows the configuration of the manipulated variable determining means 52. As shown in FIG. The operation amount determining means 52 receives signals from the shape detection mechanism 34 and the pattern recognition mechanism 33 and activates the control mechanism 141. The control mechanism 141, depending on the type of problem,
Knowledge base 56 is used to determine which inference to invoke. That is, the control mechanism 141 activates the production inference mechanism 142 when it is necessary to find a cause using a syllogism;
If there is an ambiguous factor, the fuzzy inference mechanism 143 is activated, and for a problem that has a certain degree of framework, the frame inference mechanism 144 is activated, and relationships such as causal relationships and equipment configurations are network-like. The semantic net inference mechanism 145 is activated for a problem, and the script inference mechanism 146 is activated for a problem in which the objects to be diagnosed operate in a temporal order. Furthermore, the control mechanism 141 activates the optimization arithmetic mechanism 111 to quickly find an optimal solution for empirical problems that cannot be solved by the various reasoning mechanisms described above, can memorize patterns, extract features, and provide answers. Feature extraction/answer mechanism 110 (Rum
(consisting of an Elhart-type neurocomputer). The processing result of the manipulated variable determining means 52 is outputted to the command value calculating means 53 via the control mechanism 141.

FIG. 6 shows the structure of the knowledge base 56, which is the knowledge necessary for inference. In the knowledge base 56, knowledge 106 input from the outside is based on the experience of a control expert, etc.
production rules 147 for performing inferences using a syllogistic approach; fuzzy rules 1489 for inferences based on ambiguous information; and knowledge frames 149 that can be described in a certain framework, such as the configuration of parts to be diagnosed. A semantic network 150 that organizes relationships between parts and common sense relationships in the form of a network, a script 151 that organizes and stores tasks when the diagnosis target performs certain tasks in order, and , above knowledge 147~
151 and is classified and stored as other knowledge 152 that cannot be described.

FIG. 7 shows an explanatory diagram of the operation of the manipulated variable determining means 52. The processing of the control mechanism 141 is performed by a pattern recognition mechanism 33, a shape detection mechanism 34. a processing step 200 for organizing information from storage 35 and converting it into data that can be used for further processing;
Repetitive processing step 201 that retrieves the data prepared in step 200 above and passes it to step 202 until it is exhausted.
, a judgment step 202 for determining the inference mechanism and process to be activated from the information collected in step 201;
and various inference mechanisms 142 to 146, feature extraction and response mechanism 110, optimization calculation mechanism 111, and general control mechanism 203 that executes algorithms of classical control such as PID control and modern control such as multivariable control; It consists of an end processing step 204 that executes reset of flags etc. necessary to end each step.

Here, the role of each processing mechanism will be described. The production inference mechanism 142 is suitable for control in which an expert operator assembles a logically established relationship using fragmentary production rules. The fuzzy inference mechanism 143 quantifies the operator's vague (qualitative) knowledge that cannot be quantified, such as the operator moving the actuator a little if the state of interest of the controlled object changes, so that it can be processed by a computer. It is suitable for determining the manipulated variable.

The frame inference mechanism 144 uses the knowledge of frames that describe relationships between control devices, and performs operations based on the relationships between these devices when returning to the original state when the state of the controlled object of interest changes. It is suitable for determining the amount of processing to be performed for each related device.

The semantic net inference mechanism 145 organizes and systematizes the frames, which are the fragmentary knowledge, to create a network, so it can determine the influence of the operation result of a specific actuator and create a compensation system. suitable for.

Since the script inference mechanism 146 makes inferences based on procedural knowledge when a specific state occurs, it is suitable for sequence control-like control where a failure must be dealt with in a fixed procedure.

Further, the feature extraction response mechanism 110 determines in advance the relationship between the input patterns of the pattern recognition mechanism 33, shape detection mechanism 34, and storage mechanism 35 and the outputs output by the inference mechanisms 142 to 146 when the input patterns are input. If the inference mechanisms 142 to 146 perform inference and determine the output, the learning feature is that the same result can be outputted at high speed. The optimization calculation mechanism 111 uses the steepest slope method, dynamic programming, linear Programming, ascending method, conjugate slope method or Hopfield method
field) type neurocomputer, etc., and provides an optimal response even to nonlinear control objects.

FIG. 8 shows an explanatory diagram of the operation of the proxy inference mechanism 142. The production inference mechanism 142 activated by the control mechanism 141 takes out the input process 54 stored in the memory at the time of activation from the control mechanism 141 and the information stored in the input process 54 one by one. If there is none, a termination determination step 55 is executed to terminate the processing of the production inference mechanism 142.

Using the pattern type and its certainty factor extracted in the termination judgment step 55, rules are extracted one by one from the knowledge base 56, and in step 57, the input pattern type and the antecedent part of the rule are compared. Step 58 executes the next process 59 if the comparison results match, and executes step 57 if they do not match. Step 59 replaces the input with the conclusion part of the rule when there is a match. At this time, the confidence level is handled using the mini-max theory, and is replaced with the minimum value or maximum value before replacement. In step 60, if the conclusion part of the replaced rule is an operation command, step 61
If the conclusion part does not match, step 57 is executed to perform further inference.

When the conclusion part is an operation command, the process 61 outputs the conclusion part and the confidence obtained in the processing step to the command value calculation means 53.

FIG. 9 shows the command value calculation means 53. Command value calculation means 5
3 is a memory 62 for storing commands and their reliability which are the inference results obtained by the manipulated variable determining means 52; It is determined whether or not all the commands in the memory have been processed, and if they have been processed, the command value calculation means 53 is terminated (step 63); if not processed, the actuators 27, 28, 29, 30 of the lowering control mechanism 26, etc. .31 each command is taken out,
Consists of a process 64 in which the center of gravity of the manipulated variable is determined based on the degree and confidence of the actuator operation determined through various inferences, and the center of gravity of the manipulated variables of the same actuator is collected to determine a new center of gravity and is used as a command for the corresponding actuator. be done.

By providing such a command value calculation means 53, various inferences 142 to 146. Feature extraction/answer mechanism 110, optimization calculation mechanism 111. - It has the feature that commands to the actuators individually obtained by the membrane control mechanism 203 can be handled in a unified manner.

FIG. 10 shows an input switching device 125 necessary for the learning.
The configuration is shown below. The input switching device 125 uses a switch mechanism 156 controlled by a learning mechanism to switch one of the output of the shape detection mechanism 34 and the output of the learning mechanism 36 to the input layer 5.
1. The state of the switch mechanism 156 in FIG. 10 indicates a state in which learning is performed.

FIG. 11 shows the configuration of the learning mechanism 36. The learning mechanism 36 includes an input pattern generation mechanism 65 and an output pattern generation mechanism 6.
7, an output matching mechanism 66, and a learning control mechanism 68. The output matching mechanism 66 outputs the output o1. of the distributor 139 for outputting the output of the output layer 50 to the command generation mechanism 32 and the matching mechanism 66. oI, on and the output OT of the output pattern generation mechanism 67, OTJ, oT, l
Adders 161, 162, 163 calculate the difference between the deviation e
□.

IBJ)an and output to the learning control mechanism 68. Note that the output 01+Oly On of the distributor 139 is the output of the input pattern generation mechanism 65 that is output from the pattern recognition mechanism 33 (
It is generated by being input to the input layer 51 of a Rumelhart (Ru+*elhart type neurocomputer). At this time, the input pattern generation mechanism 65 and the output pattern generation mechanism 67 are controlled by the learning control mechanism 68.

Figure 12 shows the weight function ω''1743 in the learning process.
The relationship between the learning control mechanism 68 and the learning control mechanism 68 is shown. In response to the deviation ek that is the output of the adder 161, the learning control mechanism 68 calculates the weight function ω1,
43 is changed in the direction that the deviation decreases.

FIG. 13 shows a processing outline 170 of the learning control mechanism 68. When the learning mechanism 36 is activated, the process 170 of the learning control mechanism 68 is activated. The processing 170 includes the input pattern generation mechanism 65. The output pattern generating mechanism 67 is started, and the input as a teacher signal and a pre-processing step 171 for generating the desired output are performed, and the following steps 173 and 174 are performed until the value of the deviation ek or the sum of the squares of the deviations falls within the allowable range. Step 172 of repeating ゜175. Step 173 of sequentially extracting the intermediate layer of interest from the intermediate layer close to the output layer 50 toward the input layer 51; Step 174 of sequentially extracting the cells of interest in the intermediate layer; and the cells extracted in the direction where the deviation ek becomes smaller. It consists of step 175 of changing the weight function ω, , 43 of , and step 176 of terminating the learning process.

By providing such a learning mechanism, a new phenomenon that has not been considered until now occurs, and once a countermeasure against it has been determined, it has the characteristic of being able to reflect that knowledge.

FIG. 14 shows the configuration of the storage mechanism 35 of FIG. 2. The storage mechanism 35 includes the command generation mechanism 32. A memory element 69 to which the output of the shape detection mechanism 34 is input; a memory element 70 to which the contents of the memory element 69 are transferred after a certain period of time; and a memory element 71 to which data is sequentially transferred to the memory elements and arrives after a certain period of time. each memory element 69,
The contents of 70.degree. 71 are input to the pattern recognition mechanism 33.degree. learning mechanism 36 via an arithmetic mechanism 510 for performing pattern differentiation and integration.

This storage mechanism 35 allows consideration of temporal changes in the shape detection mechanism 34 and the command generation mechanism 32. For example, operations such as differentiation and integration can be performed.

FIG. 15 shows a mechanism for recognizing a pattern using an input near the nozzle because the influence of the nozzle in coolant control affects only a certain length from the nozzle position. The output of the shape detection mechanism 34 is sent to the memory 7 of the pattern recognition mechanism 33.
2 and the memory 72 is input to the memory element 74 via the gate circuit 73, and the signal input to the memory element 74 is input to the memory element 77.78 via the gate circuit 75.76. , gate circuit 73°76
When the gate circuit 75 is turned off, the gate circuit 75 is turned on, and in synchronization with the clock, the information in the memory element 74 is transferred to the memory 77. After a certain period of time, the signal from the memory element 74 reaches the memory element 78, and the information in the memory element 77 is transferred to the memory element 77. The signal is memory element 7
4, memory elements 74, 77, 78 on the next clock
When the signals of
The information of the memory elements 74, 77, and 79 is input to the input layer 51.

By providing such a memory 72, the input layer 51 . Middle layer 39°47.49. This has the effect of greatly reducing the number of cells in the output layer 50.

FIG. 16 shows an input pattern generation mechanism 65 of the learning mechanism 36, an output pattern generation mechanism 67, and a controlled object simulator 80.
Here is an example using.

The shape pattern generated by the operator's operation or data in the output pattern generation mechanism 67 is input to a command generation mechanism 32A that is provided separately in the learning mechanism and has the same function as the command generation mechanism 32 in FIG.
At A, commands for various actuators are generated according to the pattern, and the commands are input to the controlled object simulator 80 provided in the input pattern generation mechanism 65, and the various actuators 26, 27, 28, 29, 30. 3
1 and the rolling mill of the controlled object 1, and when the response is poor, the command generation mechanism 32A and the parameter adjustment mechanism 81 for changing the parameters of the controlled object simulator 80 are used to simulate the output of the controlled object simulator 80. pattern recognition mechanism 3.
3 input.

The operation of the control method having the configuration described above will be described below using a specific example.

The initial value of the weighting function ω1゜48 of the intermediate layer 39, 47, 49 of the neurocomputer constituting the pattern recognition mechanism 33 is initially set to a random number or an appropriate value, such as a value that the weighting function can take (O ~ 1.0). Set it to half (0, 5) of ). At this time, for example, the concave rolling mill shape pattern generated by the input pattern generation mechanism 65 in FIG. 17 may be input.

The force signal line 9 is concave at the output of the output layer 5o.
The output of 0 does not become 1, and the probability that the output of the output line 91 of the output layer 5o is convex does not become zero.

Therefore, the learning mechanism 36 corresponding to the output line 90 of the output layer 50
The output line 92 of the output pattern generation mechanism 67 is 1, and the output line 9 of the 8-force pattern generation mechanism 67 corresponding to the output line 91 is
Output the output of 3 to zero.

Upon receiving these outputs, the output matching mechanism 66 receives the deviation between the ideal output (output of the output pattern generation mechanism 67) and the output of the pattern recognition mechanism 33.The learning control mechanism 68
The magnitude of the weight function ωiJ of the pattern recognition mechanism 33 is changed in a direction in which the deviation decreases in proportion to the magnitude of the deviation (see FIG. 12). A typical example of this algorithm is the steepest slope method.

The weight of the weight function is sequentially changed according to the process shown in FIG. 13, and when the sum of squares of ek shown in FIG. 12 falls within the allowable range, the operation of the learning mechanism 36 ends.

After learning is completed, when the same waveform as the output pattern of the input pattern generation mechanism 65 of FIG. 17 is input from the shape detection mechanism 34 of FIG. and outputs zero from the output line 91 of the output layer 50.

Next, if the waveform shown in FIG. 18, which is said to be convex, is input, and learning has not yet been completed, the output of the output line 91 representing the convex shape of the pattern recognition mechanism 33 is 1, and other The pattern in which the output 90 of is zero does not occur. Therefore, as described above, using a typical convex pattern as an input signal, the output of the output pattern generation mechanism 67 is generated by the output line 9.
The learning mechanism 36 changes the weight function ω''IJ so that the values corresponding to the outputs of 1 and 90 become 1 and 0, respectively, and when the learning is completed, the pattern recognition mechanism 33 When a convex waveform is input, the output line 91 of the output layer 50 in FIG. 17 becomes 1 and the output line 90 becomes 0.

As a result, the waveform shown in FIG. 19(a) is the pattern recognition mechanism 3
3, and its output is an output line 9 indicating that it is a convex waveform inputted in advance from the output layer 5° as described above.
1, the degree of resemblance to the waveform is output with a certainty of 40%, and at the same time, the degree of similarity to that waveform is output with a certainty of 50% from the output line 90 indicating that it is a concave waveform.

FIG. 20 shows the shape of a rolled material in consideration of temporal changes in the rolled material. The state directly below the rolling mill work roll 22 is t. And the value at that time is X. It is. The sampling period of the computer is T. Then, the height of the plate thickness at time t, which is To seconds ago, is xl, T, and the height of the plate thickness at time tn, Xn seconds ago, is X, .

That is, at time tl, the height xl is input to the storage mechanism 3S and stored in the memory element 69 of FIG. t, which is the next sampling point. When the height Xo of is input to the storage mechanism 35, data x1 of the memory element 69 is transferred to the memory element 70 at that timing, and the contents of the memory element 69 are rewritten to xo.

On the other hand, the calculation mechanism 510 performs various calculations using the contents of the memory elements 69 and 70. for example.

When a differential value is required, the calculation is (xox))/To, and when an integrator is required, the following calculation is performed: (X + x,) XT. In other words, the differentiator is used to find the rate of change of the shape. Therefore, the pattern recognition mechanism 33 can improve its responsiveness to changes.

On the other hand, an integrator can exhibit characteristics such as having a noise removing effect.

The pattern recognition mechanism 33 can have functions such as a differentiator, an integrator, and a proportional element that does not include a temporal element.

Furthermore, the data stored in the storage mechanism 35 can also be stored as needed.
It can be used for the input pattern generation mechanism 65 used during learning.

By the way, as shown in FIG. 21, the thickness of the rolled material in the roll axis direction of the rolling mill at time t0 is expressed as XOI
...X. -1, x, binding, 10 at the same position
The state of the plate thickness seconds (sampling period) ago is
l...X. −6,X, then at a certain point t
At time k, the memory element 79, which is the memory element 74 and 77 in FIG. 15, has a configuration similar to that of the memory mechanism 35 described above, so that time elements 79 and 79 before time Tn are stored.

FIG. 22 shows an example of production rules or fuzzy rules. (Corresponds to production rule 147 and fuzzy rule 148 in FIG. 6).

When the pattern recognition mechanism 33 obtains an output as a concave 50% confidence level, it is compared with the premise of the production rule and matches the concave rule 180.

As a result, a rule 181 that weakens the vendor (Small) is obtained. On the other hand, with a convex confidence of 40%, it matches the premise 182, and as a result, it strengthens the vendor (large degree)
is obtained.

As a result, as shown in FIG. 23, the command generation mechanism 32:
As a result of the comparison with the rule, the vendor's operation amount has a convex shape with a certainty of 50%, so it is represented by the area of the diagonal line in B. on the other hand,
Since the confidence of concave shape is 40% and the confidence of S is 40%,
This is the area of the shaded part S in FIG. 22. Next, in the command generating mechanism 32, 55%, which is the value of the center of gravity C which is a combination of the centers of gravity A and B in the shaded area, becomes the amount of operation by the bender.

Next, in cases where the influence of the actuator is local, unlike that of the bender or shifter, such as coolant control, the waveform of FIG. 24(a) is converted to the memory element 74.
7. 78. A part of the waveform stored in the memory element (see ■ in FIG. 24(a)) is processed by the if pattern recognition mechanism 33 and command generation mechanism 32, and by controlling one nozzle A of the coolant control device 30. The amount of coolant is controlled and the roll is flattened.

Now, xQ, x on both sides of x, - in FIG. 21, which corresponds to nozzle A. A conclusion 185 can be obtained from FIG. 22 that if X, -1 is large when compared with the value of -2, the center is large. On the other hand, as the relationship of X-,,x,-1,Xm-1X
If n-L is positive, Xo-1 tends to increase, so the differential coefficient becomes positive, agreeing with the premise 186, and as a result, the coolant is turned on. The extent of this is a large CB). As a result, x e-L and x n-L hardly change.

When control of nozzle A is completed, memory element 74 in FIG.
．． Shift the contents of 77, 78, and 79 by one.

As a result, the waveform input to the pattern recognition mechanism 33 is
The area indicated by ◎ in FIG. 24(a) is input, and processing 33
．． 32, one nozzle B of the coolant control mechanism 30 is controlled.

When processing is performed in this way, the pattern shown in FIG. 24(a) is started, and if the memory contents are further shifted, the pattern shown in FIG.
) is reproduced in the memory 72.

After a certain period of time has passed since the pattern shown in FIG. 24(a) was previously stored in the memory 72, the contents of the memory element 74 shown in FIG. 15 are transferred to the memory element 79, and the waveform of the shape detection mechanism 34 is stored in the memory element 74. let

Furthermore, it is obvious from FIG. 14 that if the arithmetic mechanism 510 shown in FIG. 14 is provided between the memory 72 and the input layer 51, it becomes possible to control the waveform change rate, etc.

Next, a method of learning a reference pattern for the pattern recognition mechanism 33 will be described.

Waveforms 82 and 83 in FIG. 19 are generated by the input pattern generation mechanism 65 in FIG. 11 and output to the input layer 51. This pattern is written into the memory of the input pattern generation mechanism 65, or a pattern stored in the storage mechanism 35 of FIG. 2 is used. The signal input to the input layer is sent to the intermediate layer 39. ...
, 47, and appears as an output from the output layer 50. At this time, the weight function ω of the intermediate layer! 1. is the initial value, and the input pattern generation mechanism 65 is
Corresponding to the output of 1 output terminal
, and the other terminals become zero) is input to the matching mechanism 66. When learning is not completed,
Output pattern of output layer 50 and output pattern generation mechanism 6
7 have different waveforms. As a result, the butting mechanism 6
6 outputs an output according to the degree of pattern difference. By finding the root mean square of this value and the deviation, the power spectrum of the deviation, etc. can be found. Various methods can be considered for changing the weighting function ωIiJ in order from the intermediate layer 49 near the output layer to the intermediate layer 39 near the input layer 51 according to the deviation. This is an optimization problem in which the deviation is minimized, and the steepest slope method is used, for example.As a specific method, the weighting function ω"
When IJ is slightly changed upward, looking at the direction in which the deviation value changes, the value of the weighting function ω'14 is changed in the direction in which it decreases.
, and the amount of movement is large when the change in the deviation value is small, and conversely, the amount of movement is made small when the change in the deviation value is large. Furthermore, when the change of the weighting function ω"iJ of the intermediate layer 39 closest to the input layer is completed, the deviation value of the matching mechanism 66 is checked again, and learning is completed when the value falls within the allowable error range. .

This control is performed by a learning control mechanism 68. However, it is not clear why the pattern recognition mechanism 33, which uses the learned results for pattern discrimination, is able to identify patterns or why the learning is successful. It is said that the number of patterns is large compared to the number of patterns, and there is a degree of freedom in the values, and even if the values are slightly out of order, or even if many patterns are memorized, it is possible to obtain good recognition results.

On the other hand, it is very difficult to decide what kind of patterns should be used for the input pattern generation mechanism 65 and the output pattern generation mechanism 67. Fortunately, a method of accurately deriving a model by operating the controlled object 1 near a certain operating point has been established in the field of control theory using a theory called system identification. However, it is difficult to model the object with strong nonlinearity in the entire operating region.

Therefore, we create a model in a specific operating region, implement control, use simulation to determine the relationship between the input and response of the control system that works well under that condition, and use this as data for learning. This procedure is performed by sequentially moving the operating point over the entire operating range of the control system and learning the optimal modeling and control commands at each time. That is, the controlled object simulator 80 in FIG.
The parameters of the simulator 80 are adjusted to operate the simulator 80 accurately at a specific operating point. Thereafter, the input pattern generating mechanism 65 . Parameter adjustment mechanism 81. Controlled object simulator 80. The command generation mechanism 32A is operated, and the outputs of the output pattern generation mechanism 67 and the controlled object simulator 80 are used as the output pattern and input pattern of the learning mechanism 36, respectively.

In a control system with such a configuration, the pattern recognition mechanism abstracts the target waveform, and the control mechanism can perform control including ambiguity.

Although the embodiment has been described using a rolling mill system as a specific example of the present invention, the control object 1. It is obvious that the various actuators 26, 27, 28, 29, 30, 31 are not limited to the rolling mill system, and can be applied to general control objects, actuators, and controllers. For example, it can be used to control a system such as a railway traffic control system that recognizes train schedule patterns and returns delayed trains to normal schedules according to various schedule change rules. That is, the train operation is expressed in a diagram, and the pattern recognition mechanism 33 extracts the characteristics of the delay. Next, based on the feature amounts, the inference mechanism creates a timetable using various rules such as, for example, passing trains at stations. In response to the results of the inference mechanism, the command calculation means 53 generates operating commands for individual trains. The train, which is the actuator, operates according to the instructions.

FIG. 25 describes shape control of a Sensima mill, which is the next application example of the present invention. The Sensima Mill, which is the controlled object 1, has a work roll 1001. 1st intermediate 0-#1002. 2nd intermediate a-#1003. and AS-U roll 1004. AS-Do-JL/ is split o-sil 10
05. Axis 006゜Saddles 1007
It consists of The shape of the rolled material 23 in the roll axis direction is detected by a shape detection mechanism 34 and input to a neurocomputer that constitutes the feature extraction mechanism 8. The output of the feature extraction mechanism is inputted as a certainty factor to the command generation mechanism 32 that constitutes the control mechanism 6. In this case, the command generation mechanism 3
2 performs fuzzy control calculations. The result of the fuzzy control calculation is output to the actuator as a command. The actuator operates the 5addles 1007, resulting in the deformation of the shaft 1006 and controlling the shape of the rolled material 23.

FIG. 26A shows an axial cross-sectional view of the Sensima mill. The rolled material 23 and the work roll 1o01 are in contact with each other, the work roll 1001 is in contact with the first intermediate roll 1002, and the first intermediate roll 1002 is in contact with the work roll 1001.
A split roll 1005 of the AS-U roll 1004 is in contact with the intermediate roll 1002, and the shaft 1006 of the split roll is in contact with the intermediate roll 1002.
is attached via a bearing, 5addle
It is transformed by operating s 1007.

FIG. 26B shows that the deformation of the shaft 1006 due to the operation of the 5addles 1007 causes the split roll 1005. Second intermediate roll 1003. First intermediate roll 1002° Work roll 1
The rolled material 23 is deformed via 001. The operating waveforms are drawn to emphasize the results.

FIG. 26C shows an operation waveform of the rolled material 23 when the first intermediate roll whose end portion is inclined is operated.

That is, when the first intermediate roll is moved in the axial direction, no load is applied to the rolled material at a location where the roll axis is tilted. As a result, the ends of the rolled material become thicker.

FIGS. 27A to 27D show the configuration and operation example of the neurocomputer once again. FIG. 27A shows the structure of a neuron, where the input xi is input to the adder 44 via the weight 43ω1, and the output y of the adder 44 is input to the Logitech function unit 45.
The signal is input to the Logitech function unit 45 and output. The relationship between the input y and the output 2 of the Logitech function is expressed by the following formula.

z=1/ (1+ [exp (-y)] ) No. 27B
The figure shows an overview of a neurocomputer. input signal 10
10 is input to the input cell 37 of the input layer 51, and the input layer 5
The output of 1 passes through the intermediate layer 39 and is output from the output layer 50 as a feature quantity.

The output of the rough pattern generation mechanism 67 and the output of the output layer 50 are input to an adder 161, and the deviation that is the output of the adder 161 is input to a learning control mechanism 68, so that learning of the neurocomputer is performed.

Next, the learning process is shown in FIGS. 27C and 27D. That is, when the input terminal is concave 101°-a during learning, the output line 1011 is 1 and the output line 1012 is O.
The output line 1o13 from the teacher signal 67 is 1. Output line 1014
Output is O. Before learning ends, output layer 5
Since the output lines 1oll and 1012 of 0 are not 1.0, a deviation ej from the teacher signal appears in the output of the adder 161, and as described in FIG. 11, the learning control mechanism 68
The weighting function ωl in FIG. 3 is changed so that the deviation ea (J=1...tn) becomes zero.

As a result, when the recess 1010-a is input, the output line 10
The values of 11.1012 become 1 and 0, respectively.

Next, input the convex 1010-b and learn so that the output lines 1013 and 1014 of the teacher signal 67 become O and l, respectively.

FIG. 28 shows the operation of the neural network after learning all the waveforms to be stored.

The output waveform 1010-c of the shape detection mechanism 34 has a convex 10
The component of 10-a is 30%, and the component of concave 1010-b is 60%.
%. The output waveform 1010-c is input to the pattern recognition mechanism 33 which is a neurocomputer, and from the pattern recognition mechanism 33, the pattern convex 101
0-a and pattern concavity 1010-b are each 0
．． It is output as 3°0.6. On the other hand, the operator has the know-how to operate the central AS-U roll if the operation waveform is concave. If the r″ antecedent part” corresponding to the know-how described in the fuzzy inference rule 148 is concave, then the -LI of the “conclusion part” AS-U roll is
The rule is ``operate the Jth item.'' The output of the feature extraction mechanism corresponds to the confidence of the antecedent part of the fuzzy rule. Fuzzy control combines the conclusions of individual inference rules and determines commands for each actuator. That is, fuzzy control corresponds to the function of the command generation mechanism shown in FIG.

Next, FIG. 29 shows an example in which control is performed by determining the actual actuator operation waveform as a fuzzy rule in addition to the operator's know-how. The operator operates the individual actuators 2 using the man-machine interface 7 in FIG. 1. For example, in the case of FIG.
dles 1007 to operate the AS-U role. As a result, the plate thickness shape changes and the operating waveform 1015
is obtained. This operating waveform is stored in the storage mechanism 35 shown in FIG. A controllable waveform 101 that allows a target shape to be obtained by operating an actuator from the operation waveforms stored in the learning mechanism 36 and the storage mechanism 35.
6 is found, and the neural network of the pattern recognition mechanism 33 is made to learn. .

That is, the controllable waveform 1016 in this case is the target shape minus the operation waveform 1015.

During actual operation, the operation waveform 1017 is input to the pattern recognition mechanism 33, and the pattern recognition mechanism 33
The component of the controllable waveform 1016 included in 017 is
It is input to the synthesis process 1018 of the conclusion part of the fuzzy inference shown in the figure. The output of the synthesis process of the fuzzy conclusion part is input to the actuator 2 to obtain the desired rolled material.

With this kind of configuration, control can be carried out by incorporating warm operator know-how, and control can be performed by actually operating the actuator and its operation waveform, which is easier than control using conventional models. , good control becomes possible.

On the other hand, even in a control system where the object to be controlled is modeled relatively accurately, such as plate thickness control in a rolling mill, the control may fail due to aging, non-linearity of the object to be controlled, etc. FIG. 30 shows a control system in a case where the control system has changed over time.

The states of the controlled object 1 and the actuator 2 are determined by the control mechanism 6 via the sensor 3. It is input to filter 200. Filter 200 is a type of feature extraction mechanism 8 shown in FIG. The role of the filter 200 is to extract features and reduce noise. The difference from FIG. 1 is that commands are also input to the self-adjusting mechanism 11 via a filter 200. The filter 200, the state a that is the output from the control mechanism, and the control state are input to the self-adjustment mechanism 11.

The self-adjustment mechanism 11 receives a start signal from an operator panel, which is a man-machine interface 7, and a host system 20.
The operating conditions etc. from 1 are input.

FIG. 31 explains the control mechanism 6 in detail.

The status from the controlled object 1 and the actuator 2 is transmitted via the sensor 3 to the filter 200. It is input to the comparator 202 of the control mechanism 6. The comparator 202 generates a state deviation, which is the difference between the target state of a part of the target, which is the output of the self-adjusting mechanism 11, and the state sent from the sensor 3.
03. The controller b203 is designed using a control law such as an optimal regulator, for example.

The control command that is the output of the controller b203 is sent to the adder 2
04. It is output to the self-adjusting mechanism 11.

The adder 204 adds the target value from the self** mechanism 11 and the control command to generate a command. The command is input to the filter 20o and the controller a205. Controller a205 and controller b203 can be determined by classical or advanced control theory.

Here, the comparator 202 handles the control system as a deviation value system, and the adder 204 handles the # control system as a regulator problem. Of course, there are configurations in which the control system is described as an absolute value system, and systems that are described as servo problems, but essentially, the absolute value system is a deviation value system.

Since the servo problem can be described as a regulator problem, such an assumption does not lose generality.

FIG. 32 is a detailed configuration diagram of the self-adjusting mechanism 11. The output control command from the control mechanism 6 is input to the evaluation mechanism 207 via the filter 206.

The evaluation mechanism 207 monitors whether the value of the control command exceeds the permissible range, and generates an activation signal to activate the learning mechanism 36 when the value of the control command exceeds the permissible range.

The control command, which is the output of the controller b, is used to correct deviations from the target value of the controlled object 1 due to disturbances, model errors of the control system, nonlinearity, aging, etc., by feedback. That is, the control command serves to amplify the error and return the controlled object 1 to the target state, and ideally the control command becomes zero.

Here, the error due to random disturbance is the error caused by the filter 200.
The state a is removed by passing through the state, and the state a becomes something that has a certain directionality, such as a model error or aging change, that is, something like an offset. When the value of the state a increases, it is necessary to change the operating point of the control system. This is because the output of the controller normally saturates. For example, if the offset increases and approaches the saturation value, even if a disturbance occurs, the controller output will become saturated and cannot respond to the disturbance.

The role of the evaluation mechanism 207 is, for example, to monitor the root mean square of the control command output from the filter 206, and when it exceeds a certain value, issue an activation signal to the learning mechanism 36 to change the operating point.

The learning mechanism 36 is activated by an activation signal from the operator console 7 or from the evaluation mechanism 207. Note that the operator monitors the operation of the controlled object 1 and manually operates the actuator based on experience and intuition to bring the controlled object into an ideal state. As shown in FIG. 30, the operator's operation is input to the control mechanism 6 as a manual operation command.

This manual operation command forcibly changes the target value shown in FIG. 31.

The activated learning mechanism 36 changes the parameters of the neural network 208 in response to the state a from the filter 2°O and the operating conditions from the host system 201.

The input to the neural network 208 during learning is the operating conditions etc. from the upper system 201, and the output is state a. Note that state a is created from a state and a command, as shown in FIG.

At the end of learning, the output of the neural network 208 is a target value that supports the operating point of the controlled object;
This is the target state which is the ideal state of the controlled object 1 when the target value is input. Here, the ideal state is a state of the controlled object and the actuator when the controlled object 1 is not affected by disturbance.

Note that the neural network 208 and learning mechanism 36 are collectively referred to as a neurocomputer 209.

With this configuration, the operating point of the control mechanism becomes the optimal point, and the controller a can respond to disturbances.Even if a large disturbance occurs, there is no need to compensate for the offset, so the controllable range can be expanded. .

FIG. 33 shows an outline of the operation when this embodiment is applied to rolling mill plate thickness control.

First, the normal operation represented by the solid line will be explained, and next, the operation during learning will be explained, which is represented by the broken line. Since rolling mills have strong nonlinearity, it is difficult to apply direct linear control theory to them. Therefore, the host system 201 inputs operating conditions such as base material plate thickness, product plate thickness, steel type, steel material, etc. to the neural network 208 as setup data. Assuming that the neural network 208 has been trained in advance, it adds a target value that determines the operating point of the rolling mill and a target state that is the operating state of the rolling mill at the operating point of the rolling mill in accordance with the input. Output to 204°202.

The adder 202 subtracts the operating state of the rolling mill from the target state to generate a control deviation, and the control deviation is input to the DDC control system 210, which corresponds to the controller b203 in FIG. 31.

The DDC control system 210 is designed using a regulator design algorithm, etc., feeds back control deviations, and determines command deviations.

The adder 204 adds the target value and the command deviation and outputs the result to the actuator 2 of the rolling mill. Here, in Fig. 31, the command is input to the actuator via controller a, but in Fig. 33, the gain of controller a is 1.
If so, then FIG. 33 and FIG. 31 are equivalent.

At this time, if the coefficients of the neural network 208 are optimal, the DDC control system 210 will operate in response to the disturbance.

By the way, when the output of the neural network 208 is not optimal, the error appears as a control deviation, and the value of the control command deviation increases. At this time, the command deviation due to disturbance becomes zero in terms of average value or average power. Therefore,
The evaluation mechanism 207 monitors the command deviation, and for example, when the value of the squared error of the command deviation exceeds the reference value, the evaluation mechanism 207
7 activates the learning mechanism 36. Storage mechanism 3 shown in FIG.
5 shows the adder 204 at the time when the learning mechanism 36 is activated.
The commands which are the output of the machine and the operating state of the rolling mill which is the controlled object 1 are stored, and the neurocomputer 209 is made to learn by using these commands and the operating state as teacher signals and the set-up data as input.

The difficulty in using neurocomputers for control is that there is no definitive way to provide a teacher signal.

In other words, the relationship between input and output is expressed by a nonlinear function,
There are an infinite number of correspondences. Moreover, in conventional neurocomputers, the human user provided the teacher signals. However, in complex systems with many control variables, the instructor could not fully grasp the behavior of the object to be controlled, so it was not possible to create the optimal teacher signal.

This method focuses on the fact that feedback control effectively cancels out errors caused by disturbances and model errors, and moves the operating point to the optimal position. This is a method of making a neurocomputer learn the operating state of a controlled object at an operating point. In other words, feedback control moves the operating point to an optimal point within its controllable range. Therefore, by using the results as a teacher signal, the optimal operating point is automatically determined.

On the other hand, the operator monitors the operating status of the controlled object and takes manual intervention if a problem occurs. Eventually, the behavior of the controlled object will improve and move in a positive direction. By using the operator's operation results as a teacher signal, the operator's know-how and intuition can be reflected in the neurocomputer.

Therefore, the operator manually intervenes to create the operation console 2 that constitutes the man-machine interface 7 shown in FIG.
Operate 11. The output of operation console 211 is applied to adder 204 and the command is changed. When these operations become stable, the operating point of the controlled object is at or near the optimum point.

The stable point can be determined by the operator's judgment that stability has been reached, by the elapse of a certain period of time from the operator's last operation, or by monitoring the steady-state deviation of the command deviation, and the learning mechanism 36 is activated. The operation of the system after the learning mechanism is activated is the same as when the learning mechanism 36 is activated by the evaluation mechanism 207.

FIG. 34 shows a case in which the present invention is applied to an elevator control device. An elevator palanquin 220 is balanced by a rope 223 via a counterweight 221 and a sheave 222.

An electric motor 223 is connected to the sheave 222.
3 rotates, the elevator palanquin 220 moves up and down. A pulse generator 224 is connected to the electric motor 223, and the output of the pulse generator 224 is sent to the sensor 3.
The signal is applied to the position detection mechanism 225 and speed detection mechanism 226 that constitute the . Position detection mechanism 225, speed detection mechanism 226
The output of
8 is input. The speed control mechanism 227 uses the position and speed information of the elevator palanquin 220 and the information of the elevator sequence control mechanism 228 to output a control command to the power conversion mechanism 229 according to a feedback control law.

The power conversion mechanism 229 is a mechanism that interprets the command from the speed control mechanism 227 and controls the electric power supplied to the electric motor 223. The power conversion mechanism 229 may be an inverter, a converter, a phase control, a cycloconverter, etc. depending on the type of the target electric motor.

On the other hand, the elevator passengers operate the call buttons 230, 231 arranged in the hall or the destination button in the palanquin, and follow the instructions in the elevator sequence control mechanism 2.
28 determines the destination light position and deceleration start position of the elevator palanquin.

By the way, elevators have strong nonlinearity, for example, they vibrate when passing through a specific floor. When vibration occurs, the control command for the speed control mechanism 227 increases because the control deviation increases. As a result, the self-adjusting mechanism 11 having the configuration shown in FIG. 32 determines that the root mean square value of the control commands of the speed control mechanism 227 exceeds a certain reference value, and the self-adjusting mechanism 11 with the configuration shown in FIG. A storage mechanism 35 that stores information on the position detection mechanism 225, speed detection mechanism 226, etc.
The learning mechanism 36 operates using the information as a teacher signal.

When the learning mechanism 36 completes its learning and the condition for the elevator to generate vibrations is reached, the self-adjusting mechanism 11 corrects the speed command to the speed control mechanism 227.

The above operation is schematically shown in FIG. The elevator sequence control mechanism 228 inputs a speed command for running the elevator to the speed control mechanism 227 via the self-adjustment mechanism 11 when a passenger operates the elevator carton 220 or the hall buttons 230, 231. When vibration occurs when the elevator passes a specific floor, the speed control mechanism generates a control command according to the deviation, and the power conversion mechanism 22
Control 9. Power conversion mechanism 229 controls the rotation speed of electric motor 223. As a result, the pulse generator 22
4. A speed curve is obtained through the speed and position detection mechanism 225, 226, and this speed curve is obtained through the speed control mechanism 227.
is working, so the speed command is more ideal than the speed command output from the elevator sequence mechanism 228.

In order to generate this speed command, the speed control mechanism 227
Since the value of the control command is large, the self-adjusting mechanism 11 operates and performs learning. At this time, the speed curve obtained from the speed detection mechanism 225 and conditions under which vibrations occur, such as ascending or descending a specific floor, are also input into the teacher signal at this time. The self-adjustment mechanism 11 includes a speed control mechanism 227
The speed control device 227'4 outputs the speed curve when the conditions for generating the vibration are input so that the control command for the output is minimized.

FIG. 36 shows an example in which the present invention is applied to train control. The electric train 250 runs on electricity supplied from an overhead line 251 and a track 252'. The operation management system 253 detects the position of the train 250 via the train position detection mechanism 254, and generates commands to the signal 255 and the speed command generation mechanism 256 in accordance with the position of the train. Speed command generation mechanism 2
56 generates a speed command to the train control mechanism 256 in accordance with instructions from the traffic management system 253. The train control mechanism 256 controls the speed of the train 250 using the speed command and the train speed from the train position detection mechanism 254. Here, the train position detection mechanism 254, the speed command generation mechanism 25
6 and some of the functions of the traffic management system 253 may be loaded onto the train 250.

FIG. 37 shows an outline of the train control mechanism 256.

The outputs of the train position detection mechanism 254 and speed command generation mechanism 256 are input to the self-adjustment mechanism 11. Self-adjustment mechanism 11
Furthermore, the outputs of the control mechanism 227, power conversion mechanism 229, and pulse generator 224 are inputted as positions or speeds. In response to the speed command from the self-adjusting mechanism 11 and the position and speed from the pulse generator 224, the control mechanism 227 controls the power conversion mechanism 229 and the self-adjusting mechanism 11.
Outputs control commands to. The power conversion mechanism 229 is the transformer 2
It receives power from the secondary winding 57 and also supplies power to the electric motor 222 in response to the control command. A pantograph 258 and a wheel 2 are connected to the primary winding of the transformer 257.
59 is connected, and power supplied from the overhead line 251 and the line 252 is supplied to the power conversion device 229. The system configurations shown in FIGS. 36 and 37 are the same as the elevator system configurations shown in FIGS. 34 and 35. I will not discuss the operation of the train in detail here, but the operation is the same as that of an elevator.

Furthermore, although this embodiment describes the control of an AC electric train, the essence of the control system for a DC electric car is different only in the configuration of the power converter and transformer, so the idea of the present invention is as follows.
Can be extended to DC trains.

Furthermore, although the description has been made regarding power-distributed electric locomotives, the configuration of the control device for power-centralized electric locomotives is the same as that of electric trains, so it is obvious that the present invention can be applied to electric locomotives. It is.

FIG. 38 shows an embodiment in which the present invention is applied to control of an automobile. Power generated by an automobile engine 260 is transmitted to wheels 262 via a transmission 261. engine 26
0 is controlled by the control mechanism 227 via an actuator such as a carburetor, and the states of the engine 260 and transmission 261 are input to the control mechanism 227 and the self-adjustment mechanism 11. The control mechanism 227 includes an engine control section and a transmission control section, and includes an accelerator mechanism 26 operated by the driver of the automobile.
3. Controlled by instructions from the operation of the steering mechanism 264.

Recently, some cars are equipped with automatic driving devices that drive the car at a constant speed. This is done by controlling the control mechanism 227 from the outside.
Instruct the speed at which the vehicle should travel.

The control mechanism 227 uses feedback from the sensor to run at a constant speed.

The self-adjusting mechanism 11 receives the output from the sensor 3 and the information from the control mechanism 227, and when the control deviation becomes large, the self-adjusting mechanism operates and changes the command to the actuator 2. Possible causes of increase in control deviation include aging, lubricating oil replacement, climate change, etc., and these factors form a nonlinear control system that is intricately intertwined. here,
The operation of the control system of these automobiles will be explained in correspondence with the operation of the aforementioned rolling mill.

It is assumed that the vehicle is driven at a constant speed according to instructions from the driver. The driver looks at the speedometer or the like and depresses the accelerator 263 if the speed decreases. As a result, the operation console 2 of the rolling mill operator shown in FIG.
The operation is performed after issuing a start signal to the self-adjusting mechanism 11 in the same way as the instruction No. 11. There is nothing special about cars here.

Next, FIG. 39 shows a modification of FIG. 32.

Although the nonlinearity is strong, if the model to be controlled is relatively accurate, it is possible to use that model to correct the error. Operating conditions from the host system 201 and filter 20
State a from 0 is input to the setup control system 270. The set-up control system 270 sets a target state and a target, and outputs the operating condition and the deviation between the state a and the target state to the learning mechanism 36 as a state deviation. The difference from FIG. 32 is that the neurocomputer 209 operates not with absolute values but with a deviation value system that is a deviation from a certain reference (target value or target state determined by the setup control system 270).

FIG. 40 shows an example of the operation of the plate thickness control system of a rolling mill, and shows a modification of FIG. 33. The setup model 271 receives operating conditions from the host system 201 and outputs the target value and target state of the controlled object. The state of controlled object 1 is input to adder 202 via filter 200 . The adder 202 adds or subtracts the above state, the target state from the set-up model, and the target state deviation from the neural network to obtain the state deviation. The state deviation is input to the learning mechanism 36 of the neurocomputer 209 as a teacher signal via the DDC control system and the storage mechanism 35. The DDC control system 210 operates so that the control deviation becomes zero as a regulator problem, and is inputted to the adder 272 as a control command. Adder 272
Then, the control command and the manual command from the operation console are added and inputted to the adder 204 and evaluation mechanism 207 as a new control command. The adder 204 receives the control command, the target value from the setup model 271, and
The target value deviations from the neural network 208 are added and a command to the actuator 2 is generated. Here, the setup model 271, adders 202, 27
Setup control system 270 with 2,204 functions
It is called.

In a rolling mill equipped with a plate thickness control system having such a configuration, when the rolling mill operates, a temperature rise occurs, the frictional resistance of the lubricating oil decreases, and the plate thickness changes.

As a result, since the state deviates from the target state, the state deviation of the output of the adder 202 increases. Since the state deviation has become large, the DDC control system 210 generates a control command to reduce the deviation. Since the control command rapidly reduces the deviation and usually amplifies the deviation, the evaluation mechanism 207 operates and starts learning. The adder 204 adds the control command and the target value from the setup model 271 to generate a command to the actuator 2.

Next, the learning process of the neurocomputer 209 will be explained. The input of the teacher signal of the learning mechanism 36 is the operating condition from the upper system 201, and the output is the state deviation and control command. When learning is completed, the output of the neural network 208 is added to an adder 202 as a target state deviation and to an adder 204 as a target command value deviation. As a result, before learning,
The output of the adder 202 was equal to the deviation of the target state after learning, but after learning, the target state deviation is subtracted and becomes zero. - As for the force command, the output from the DDC control system is zero, but since the target value deviation is output from the neural network, the same command as the control state before learning can be obtained from the adder 204.

〔Effect of the invention〕

According to the present invention, qualitative knowledge is quantified, the actuator is operated based on this quantified information, the operating point moved by control is determined, and control is performed based on the operating point. , qualitative knowledge that was previously difficult to introduce into automatic control systems has been introduced into automatic control, and it has become possible to control a variety of control objects with sufficient control margin.

[Brief explanation of drawings]

Fig. 1 is a block diagram of an 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 pattern recognition mechanism diagram, Fig. 4 is a command generation mechanism diagram, and Fig. 4 is a diagram of a command generation mechanism. Figure 5 is a configuration diagram of the operation amount determining means, Figure 6 is a knowledge base configuration diagram, and Figure 7 is a diagram of the knowledge base configuration.
Figure 8 is an explanatory diagram of the operation of the manipulated variable determining means, Figure 8 is an explanatory diagram of the operation of the production mechanism, Figure 9 is a configuration diagram of the command value calculation means, Figure 1o is the configuration of the input switching device, and Figure 11 is the learning The structure of the mechanism, Figure 12 is a diagram of the relationship between the learning control mechanism and node weight functions, Figure 13 is a diagram of the basic processing of the learning control mechanism,
FIG. 14 is a configuration diagram of the storage mechanism, FIG. 15 is a diagram of the pattern recognition mechanism, FIG. 16 is a configuration diagram when the learning mechanism is equipped with a simulator, and FIG. 17 is an explanatory diagram of the operation of the pattern recognition mechanism.
Fig. 18 is an example of an input pattern, Fig. 19 is an explanatory diagram of the output of the pattern recognition mechanism, Figs. 20 and 21 are explanatory diagrams of temporal changes in rolled material, and Fig. 22 is an example of production rules and fuzzy rules. Fig. 23 is an explanatory diagram of the method of converting the similarity into a manipulated variable, Fig. 24 is an explanatory diagram of the input waveform processing situation, Fig. 25 is an example of application to the ZR mill, and Fig. 2
Figures 6A to 26C are structural diagrams of the ZR mill, and Figures 27A to 26C are structural diagrams of the ZR mill.
FIG. 27D shows a specific example of neurocomputer learning, and FIG. 28 shows a specific example of neural feature extraction and fuzzy control. Fig. 29 is an example of the operation waveform when the present invention is applied to a ZR mill, Fig. 30 is an overall explanatory diagram of the self-adjustment function, Fig. 31 is a specific configuration diagram of the control mechanism, and Fig. 32 is the self-adjustment mechanism. Fig. 33 is an operational diagram of the self-adjusting mechanism, Fig. 34 is an example of application of the present invention to an elevator, Fig. 35 is an operation waveform diagram of an elevator, and Fig. 36 is an application of the present invention to vehicle control. example,
Fig. 37 is a block diagram of a train control mechanism, Fig. 38 is an example of application of the present invention to automobile control, Fig. 39 is an example of self-organization when a model is used in combination, and Fig. 40 is a self-organization in combination with a model. This is a specific example. 1... Controlled object, 2... Actuator, 3...
Detection device (sensor), 5...Optimality determination device, 6...
- Control mechanism, 23... Rolled material, 26-31... Drive mechanism, 32... Command generation mechanism, 33... Pattern recognition mechanism, 35... Memory mechanism, 36... Learning mechanism, 3
9... First stage intermediate layer, 47. 49... Another intermediate layer, 50... Output layer, 51... Input layer, 65... Input pattern generation mechanism, 66... Comparison Mechanism, 67...
Output pattern generation mechanism, 68...Learning control mechanism, 209
...neurocomputer.

Claims (24)

[Claims] 1. A control system comprising a detection device including a control object operated by an actuator, a control mechanism that controls the actuator, and a detector that detects the operation of the control object and the actuator. A control system comprising: an optimality determination mechanism that comprehensively determines operations of a controlled object and an actuator, and transmits a control command to the actuator control mechanism based on the determination result. 2. In claim 1, the control mechanism detects a deviation between a command of the optimality determination mechanism that generates a command based on information from a host computer or an operation by an operator of the control device, and a controlled variable of the controlled object. A control system comprising a control device that determines a feedback control amount to decrease the amount and generates a control command to an actuator. 3. 3. The control system according to claim 2, wherein the optimality determination mechanism includes a mechanism that can change the command according to a deviation of the control mechanism. 4. 4. The control system according to claim 3, wherein the optimality determination mechanism includes a mechanism for recognizing that the deviation of the control mechanism exceeds a predetermined determination criterion. 5. In claim 4, the optimality determination mechanism has a mechanism that generates a command based on information from a host computer or an operation by an operator of a control device, and when the determination criterion is exceeded,
A control system characterized by having a mechanism for changing commands. 6. 5. The control system according to claim 4, wherein the mechanism for generating the command for the optimality determining mechanism changes the command so as to reduce a control deviation. 7. 7. The control system according to claim 6, wherein a neurocomputer is used as a mechanism for generating commands for the optimality determining device. 8. In claim 1, the optimality determination mechanism includes a decision-making mechanism that selects an optimal control strategy from a plurality of control strategies stored in advance; A control system comprising: a control mechanism that generates a processing command using a control strategy; and a command value calculation means that converts the manipulated variable determined by the manipulated variable determining means into a control command for each actuator. 9. In claim 8, the plurality of control strategies stored in the manipulated variable determining means include an inference control strategy constructed from a knowledge base and an inference section, a control strategy using a pattern recognition method, a feedback control, etc. A control system comprising either a control strategy using a compensator or a control strategy using a state model. 10. In claim 1, the optimality determination mechanism comprises:
means for storing a plurality of combination patterns of a plurality of detection signals of the detection device and operation amounts of an actuator corresponding to the plurality of combination patterns; A control system characterized in that it is configured to compare stored output patterns and output operation amounts of a plurality of actuators for the stored pattern with the greatest degree of similarity. 11. In claim 1, the optimality determination mechanism comprises:
a pattern recognition mechanism that determines the degree of similarity between the patterns of the output signals from a combination of a plurality of detection signals from the detection device; A control device comprising a command generation mechanism that converts into command signals for a plurality of actuators. 12. In claim 11, the pattern recognition mechanism includes an input layer that takes in a combination pattern of a plurality of output signals of the detection device, and a function that multiplies and adds the output signals of the input layer with weights, and uses the result as a designated function. The output signal of each node of the first intermediate layer is weighted and added as an input signal, and the result is mapped by a specified function. A control system comprising a plurality of stages of another intermediate layer composed of a plurality of nodes, and a final stage of the another intermediate layer is composed of an output layer that outputs pattern similarity as a determination result. . 13. In claim 12, the command generating mechanism uses a knowledge base and an inference mechanism to determine the amount of operation for the actuator from the similarity of the plurality of patterns output from the pattern generation mechanism, and A control system characterized in that it is configured to convert into a command signal for a control mechanism of an actuator. 14. Claim 1: A teaching device comprising: a mechanism for determining whether the determination result of the optimality determination device is good; a means for notifying the determination result to the outside; and a means for changing the contents of the optimality determination device. A control system characterized by: 15. In claim 1, a storage mechanism is provided for storing the output signal of the detection device in time series, the output of the detector and the output of the storage mechanism are input to the optimality determination device, and the optimality determination device is configured to perform the optimality determination. The control system is characterized by the ability of the device to make comprehensive judgments that take into account temporal changes. 16. In the control system, the control system includes a plurality of actuator control mechanisms that execute control according to detection signals of a plurality of detectors and output signals of the detectors, and a controlled object controlled by the operation of the plurality of actuators. Pattern recognition that recognizes a combination of multiple detection signals as a pattern, stores the signal combination pattern in advance, compares the pre-stored pattern with a new input pattern, and outputs the comparison result as pattern similarity. A control system comprising: a mechanism; and a command generation mechanism that converts the degree of similarity of the patterns into an operation amount for each actuator by fuzzy inference, and converts the operation amount into a command value for each actuator. 17. 17. The control system according to claim 16, wherein a neurocomputer is used for the pattern recognition mechanism. 18. Claim 17, further comprising a learning mechanism for applying an input pattern to the pattern recognition mechanism and changing the weight of each node of the pattern recognition mechanism in advance so that the pattern recognition mechanism has an ideal output. A control system featuring: 19. 19. The learning mechanism comprises an input pattern generation mechanism inputting to an input layer of a pattern recognition mechanism;
an output pattern generation mechanism that generates an ideal output of the pattern recognition mechanism; an output pattern of the pattern recognition mechanism;
a comparison mechanism for determining a deviation of the output pattern generation mechanism; a learning control that generates a command to change the weight of a node in the pattern generation mechanism based on the comparison result; and an operation command for the input pattern generation mechanism and the output pattern generation mechanism. A control system characterized by comprising a mechanism. 20. Claim 16, further comprising a storage mechanism for storing the detection signals of the plurality of detectors in time series, and inputting the output of the storage mechanism to the pattern recognition mechanism in parallel with the detection signals of the detectors. A control system characterized by: 21. In a control system that includes a controlled object that operates with a plurality of actuators and a detector that detects the operation of the plurality of actuators and the operation of the controlled object, the plurality of detection signals are used to recognize the operation of the entire control system, and the system A control system comprising: an optimality determination device that determines the control amount of each of the actuators so that the overall operation is optimal. 22. In a control system for a rolling mill that includes a plurality of drive mechanisms and rolls a material, detecting the states of the plurality of drive mechanisms and the state of the rolled material, and determining the quality of the operation of the entire system based on the detection signal, A control system configured to output a command signal to each of the drive mechanisms by classifying the amount of operation of each of the drive mechanisms. 23. It has a plurality of input terminals into which the operating states of a plurality of actuators can be input and a plurality of output terminals which output control commands for the plurality of actuators, and based on the input signals, determines whether the operation of the entire system is good or bad, and determines the determination result. Optimality determination device with a function to determine the output signal to each output terminal based on 24. Claim 23, further comprising a function of patterning a combination of operation signals of the entire system and storing the pattern, determining a degree of similarity between a plurality of stored patterns and a pattern of a new input signal, and determining the degree of similarity between the plurality of stored patterns and a pattern of a new input signal. The optimality determination device is composed of a recommendation function that obtains each force signal from the above.