JP3040901B2 - Control method by neural network and built-in control device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for applying a neural network to control of various control objects.

[0002]

2. Description of the Related Art Conventionally, in a neural network used for controlling various controlled objects, an output from the neural network is sent as it is to a controller for controlling the controlled object provided connected to the controlled object. Command value, the amount of operation transmitted to the control target, and the like.

That is, for example, a predetermined input signal is input to a neural network, and learning of the neural network is performed until an error between an output from the neural network and a desired output for the input signal becomes equal to or less than a predetermined value. The learned neural network is used for controlling the control items to be controlled.

In such learning, for example, a control amount is used as the input signal, and an output of the neural network corresponding to the input signal is used as an operation amount corresponding to the control amount.

[0005] Then, a certain control amount is input to the neural network in which learning has been completed, and the output from the neural network at that time is used as an operation amount, which is directly input to a controller provided in the control device to be controlled. Control items.

[0006]

However, the above prior art has, for example, the following problems.

[0007] Since the output from the neural network directly corresponds to a command value for the controller for controlling the control target and the manipulated variable sent to the control target, the output error of the neural network directly changes the command value. And the accuracy of the control amount,
The control system was degraded significantly or the error caused the control system to become unstable.

It is an object of the present invention to provide a control method incorporating a neural network for controlling an output of a neural network with high accuracy while reducing the output error of the neural network.

[0009]

[Means for Solving the Problems] In order to solve the above problems, the following means are conceivable.

[0010] At least one first input value, which is a deviation amount from a control target value of the control object, and a state quantity of the control object, which is constituted by having an input layer, an intermediate layer, and an output layer. A control device with a built-in neural network including a neural network that receives at least one or more second input values from an input layer and outputs an operation amount given to a control target from an output layer, the control device including an input layer, an intermediate layer, and an output layer The input layer of the new neural network includes a second input value and a predetermined value corresponding to the first input value. And outputs an output value from the output layer by an operation performed by the intermediate layer and the output layer based on the received value, and the subtraction unit outputs the output value of the two neural networks. Calculates the difference between the force value is a device that outputs as the operation amount.

The following control method is also conceivable.

That is, a neural network having an input layer, an intermediate layer, and an output layer is used, and at least one or more first input values, which are deviation amounts from a control target value of a control object, and control. In a control method using a neural network in which at least one or more second input values, which are state quantities of an object, are given to an input layer, and an operation amount given to a control object is output from an output layer, the input layer, the intermediate layer,
A new neural network configured with an output layer is prepared, and an input layer of the new neural network is predetermined in correspondence with the second input value and the first input value. A specific value is given, an output value is output from the output layer by a calculation performed by the intermediate layer and the output layer based on the received value, a difference between the output values of the two neural networks is calculated, and output as the manipulated variable. This is a control method using a neural network.

[0013]

As described above, according to the present invention, in addition to the first neural network for performing a normal neural network operation, a second neural network for newly calculating the output error of the first neural network is provided. It is solved by preparing.

The second neural network calculates an output error of the first neural network near a preset input value. A value obtained by subtracting the output of the second neural network from the output of the first neural network becomes a final output signal, which is a command value for a controller for controlling the controlled object and an operation amount transmitted to the controlled object. Can be handled.

Thus, if the second neural network calculates the error with respect to the input, which is the most important for stabilizing the control system, the neural network can calculate an accurate output value corresponding to the vicinity of the input value. Calculation can be performed with high accuracy, and the control performance of the control system is improved.

[0016]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows a configuration example of a neural network control device according to the present invention.

The neural network control device 100 is a device that calculates a desired command value to be given to the control target 104 and outputs the command value to the controller 103.

The neural network control device 100 comprises a neural network operation means 101, a target value storage unit 107, and a command value calculation means 102, and further comprises an input means 108.

The neural network operation means 101, the target value storage unit 107, and the command value calculation means 102 are, for example, CP
It is realized by electronic devices such as U, ROM (a program for performing a predetermined process is pre-installed), RAM, and various CMOSs. The input unit 108 can be realized by, for example, a keyboard, a mouse, and the like.

Although not particularly shown, a configuration having a display means for displaying various information such as the command value may be considered. In this case, the display means is realized by, for example, a CRT, an EL display, a liquid crystal display, or the like.

Further, the neural network operation means 101
Has two neural networks, and the command value calculation means 102 has a model 105.

The model 105 expresses the control target by a mathematical expression including an uncertain parameter, and it is possible to determine the operation amount with respect to the control target value using the mathematical expression.

The controller 103 outputs a command value to an actuator 106 provided in the control target 104 based on the command value given by the neural network control device 100.

The controller 103 includes, for example, a CPU, an R
It is realized by an electronic device such as an OM (a program for performing a predetermined process is stored in advance), a RAM, and various CMOSs.

The control target 104 is provided with a first sensor 109 for detecting a control amount and a second sensor 110 for detecting a state of the control target 104 other than the control amount.
The signals detected by these sensors are sent to the neural network control device 100.

Specific examples of the control target, the first and second sensors, and the actuator will be described later with reference to specific examples.

The neural network learning device 111 is a LAN 112 using an optical fiber or the like as data transmission means.
Is connected to the neural network control device 100.

The neural network learning device 111 is realized by an electronic device such as a CPU, a ROM (a program for performing a predetermined process is pre-installed), a RAM, and various CMOSs. This is made into one chip and L
It may be realized by SI or the like.

Of course, the neural network learning device 11
1 may be configured to be provided in the neural network control device 100.

The neural network learning device 111 determines the computation content of the neural network computing means 101 provided in the neural network control device and converts the content to LA.
The data is sent to the neural network operation means 101 via N112.

First, the neural network control device 1
The flow of the process of 00 will be described.

As described above, the neural network control device 100 includes the command value calculation means 102 and the target value storage 10
6. It has a neural network operation means 101 for performing an operation by a neural network. The target value storage unit 106 stores a value that is a target control amount of the control target 104, and the value is stored in, for example, the input unit 10.
8 may be set and changed.

The command value calculation means 102 controls the control object 104.
Is calculated based on a model 105 that is modeled using a mathematical expression, so that the output of the control target 104 matches the control target value.

The neural network operation means 101 includes a first
The difference between the control amount acquired by the first sensor 109 and the control amount calculated using the model 105, which is the input value of Modify the parameters.

When another value (for example, a physical quantity such as temperature) representing the state of the control object affects the relationship between the difference and the correction amount of the parameter, as shown in FIG. Is input by the second sensor 110 as a second input value, and is input to the neural network operation means 101. The physical quantity captured by the second sensor 110 is input to the first and second neural networks exactly the same.

The command value calculation means 102 includes a model 105,
That is, every time the uncertain parameter of the formula is corrected,
The command value is recalculated based on the corrected model 105, and the calculation result is used as a new command value in the controller 103.
Output to

Next, a detailed configuration and processing of each section will be described.

In this embodiment, a heating furnace plant for hot rolling is employed as the control object 104.

FIG. 2 shows a configuration example of the heating furnace. Heating furnace 2
At 00, the slab 204, which is the material to be heated, is inserted from the inlet of the heating furnace and heated by the gas burner 201 while moving to the exit side of the heating furnace. In the correspondence with the configuration of FIG.
And the first sensor 109 corresponds to the thermometer 203 that measures the surface temperature at the outlet of the slab 204. Further, the second sensor 110 corresponds to a thermometer 205 for measuring the insertion temperature of the slab 204 and a thermometer 202 for measuring the temperature inside the heating furnace 200.

In the heating furnace plant, the control target is to bring the surface temperature at the outlet of the slab 204 to the target temperature by appropriately operating the gas burner 201.

Model 10 provided in command value calculation means 102
In FIG. 5, the amount of radiation heat transferred to the slab 204 is modeled. A typical model corresponding to the amount of radiation heat transfer is the Stefan-Boltzmann heat conduction equation. The following equation (1) shows a general Stefan-Boltzmann heat conduction equation.

[0043]

(Equation 1)

Here, Q is the amount of heat transferred from the furnace to the slab,
T is the furnace temperature, which is detected by the thermometer 202. θ is the temperature φcg of the slab 204 and the heat transfer coefficient called the overall heat absorption coefficient. The value of the overall heat absorption coefficient is
In general, it is very difficult to calculate exactly. In this embodiment, the purpose is to tune the overall heat absorption coefficient by the neural network operation means 101 so as to match the overall heat absorption coefficient of the heating furnace 200. It is.

The command value calculation means 102 calculates the model 10
5, a furnace temperature command value such that the temperature of the slab 204 at the heating furnace outlet satisfies the target temperature is calculated, and the calculation result is output to the controller 103. The controller 103 controls the gas burner 2 so that the furnace temperature matches the command value.
01 is controlled.

FIG. 3 shows a neural network operation means 101.
An example of the configuration will be described.

The neural network operation means 101 comprises:
, And a second neural network 302.

In the first neural network 301, the correction amount of the overall heat absorption coefficient is changed, and in the second neural network 302, the first neural network 3
The error amount of the correction amount calculated by 01 is calculated.

The first neural network 301 receives, as a first input value, a difference between a control amount actually taken in from the thermometer 203 and a control amount calculated using the model 105. Further, as the second input values, the temperature inside the furnace and the insertion temperature of the slab 202, which are respectively taken by the thermometers 202 and 205, and the value of the overall heat absorption coefficient currently set in the model 105 are input.

Then, in order to reduce the difference, the correction amount of the overall heat absorption coefficient is calculated and output to the model 105.
In the model 105, the overall heat absorption coefficient, which is an uncertain parameter constituting the equation, is changed based on the input correction amount.

The second neural network 302
It has the same configuration as the first neural network 301, and when the correction amount is “0”, which is the difference between the control amount detected from the thermometer 203 and the control amount calculated using the model 105. The output error of the first neural network is calculated. Therefore, “0” is input to the input corresponding to the difference between the control amount detected from the thermometer 203 and the control amount calculated using the model 105, and the other input is the first neural network. A second input value that is the same value as that of the network 301 is input.

The difference between the output of the first neural network 301 and the output of the second neural network 302 is obtained by the neural network operation means 10.
The output is 1, and the overall heat absorption coefficient of the model 105 is corrected according to this value.

As described above, since the first and second neural networks have the same configuration, a specific value “corresponding to the“ control amount-control amount by model ”which is an input signal of the first neural network 301 is used. When "0" is input to the second neural network 302, the error is reduced by utilizing the fact that the output of the neural network 302 at this time becomes the error amount of the first neural network 301.

Next, FIG. 4 shows a flowchart of a process performed by the neural network operation means 101.

First, in s4-1, the difference between the control amount (hereinafter, referred to as “y”) detected by the thermometer 203 and the control amount (hereinafter, referred to as “yd”) calculated using the model 105. Further, a parameter group other than the control amount (corresponding to the second input value), such as the furnace temperature, the insertion temperature, the overall heat absorption coefficient, and the like are input to the first neural network 301, and the calculation result is corrected. Let it be 1.

Next, in s4-2, "0" is input to the input corresponding to "y-yd", and the second neural network 302 executes an operation. The calculation result at this time is defined as an error amount.

Next, in s4-3, the correction amount 2 is obtained by subtracting the error amount obtained in s4-2 from the correction amount 1 obtained in s4-1. Then, the correction amount 2 is output in s4-4, and the process ends.

Next, a detailed description will be given of the neural network 301 (in the present embodiment, the first neural network and the second neural network have the same configuration, so only the first neural network will be described).

FIG. 5 shows the first neural network 3
01 shows a configuration example.

This neural network has an input layer, an intermediate layer, and an output layer.

The first neural network 301 is
A neuron 501 is configured by being connected in layers by a synapse 502. In the example shown in FIG. 5, the input layer 503,
Although a three-layer network including the intermediate layer 504 and the output layer 505 is shown, it goes without saying that a network having four or more layers including a plurality of intermediate layers 504 may be constructed.

Assuming that the input input to the input layer 503 is Ii and the output of the intermediate layer 504 is Mj, the relationship between Ii and Mj is given by the following equation, for example.

[0063]

(Equation 2)

[0064]

(Equation 3)

[0065]

(Equation 4)

Wij (j is the number of the hidden neuron) is
The weight given to the synapse 502 connecting the i-th input neuron and the j-th hidden neuron, θj is a constant corresponding to the neuron j, and f (uj) is generally a differentiable monotone saturation function. But usually, as in equation (2),
A sigmoid function is used. Further, using Mj thus obtained, the output Ok of the k-th neuron in the output layer 505 becomes

[0067]

(Equation 5)

[0068]

(Equation 6)

[0069]

(Equation 7)

Is obtained. Vjk is the weight connecting the j-th intermediate layer neuron and the k-th output layer neuron, and θk is a constant corresponding to the output layer neuron.

As described above, the neural network computing means 103 performs the computation for determining the output Ok corresponding to the input Ii.

In the neural network operation means 101, the second neural network 302 has the same structure as the first neural network 301.
A configuration in which the neuron in the input layer 503 corresponding to the difference between the control amount detected from the control unit 03 and the control amount calculated using the model 105 and the synapse connected thereto may be omitted.

FIG. 6 shows a neural network learning device 111.
An example of the configuration will be described.

The neural network learning device 111 has a storage unit 6 in addition to the neural network 603 having the same structure as the first neural network 301.
01, learning means 602, input means 604, output means 60
5 is provided. These are buses 60
6 for transmitting and receiving necessary signals.

The storage means 601 is realized by, for example, a RAM, a disk drive device having a disk as a storage medium mounted thereon, and the learning means 602 is realized by, for example,
It is realized by an electronic device such as an OM (a program for performing a predetermined process is stored in advance), a RAM, and various CMOSs. The input unit 604 is, for example, a keyboard, a mouse, or the like. The output unit 605 is, for example, a keyboard, a mouse, or the like.
This is realized by a CRT, an EL display, a liquid crystal display, or the like.

The storage means 601 stores at least one input signal and at least one output signal (hereinafter referred to as an “input / output pair”) that the neural network 603 wants to learn.

The learning means 602 fetches one input / output pair from the storage means 601 one by one via the bus 606, and supplies an input signal among them to the neural network 603 via the bus 606. The neural network 603 performs an operation based on the given input signal, and sends out the operation result to the learning means 602.

The learning means 602 detects an error between the transmitted output and the output signal of the input / output pair, and, according to the detected error, weights Wij, Vij of the neural network 603.
Modify jk to reduce the error.

Various methods have been proposed for correcting Wj and Vj.

As a representative method, there is an error back propagation learning method using the steepest descent method. This learning method is based on an output signal of an input / output pair and a neural network 60 corresponding to the output signal.
In this method, the weights Wj and Vj are corrected in the direction of decreasing the function by the steepest descent method, which is one of the optimal solution search methods, based on the energy function that gives the sum of the error amount with the output of No. 3. .

The energy function is shown in the following equation.

[0082]

(Equation 8)

Tk is an output signal of an input / output pair corresponding to the output Ok. An example of a procedure for correcting the weights Wj and Vj by the steepest descent method is shown in equations (9) and (10). Wj (n), Vj (n)
Represents the weight of the n-th correction.

[0084]

(Equation 9)

[0085]

(Equation 10)

By performing the processing of the above equation for all weights, the (n + 1) th correction is completed.

The details of this method are described in, for example, a magazine (David
E. Rumelhart. Geoffrey e. Hinton & ronald J. willia
ms. "Learning representations by back-propagating
error. "Nature. Vol.323. No.9. Paragraphs 533-536.
Oct. 1986).

Now, the learning means 602 includes the storage means 601
The input / output pairs stored in the memory are successively taken in, and the above operation is repeated. Finally, the relationship between the input and the output of the storage means 601 is stored in the neural network 603.

The input means 604 allows the user to
1 is a unit for inputting an input / output pair and changing a learning parameter of the learning unit 602. Output means 60
Reference numeral 5 denotes a unit for monitoring an input / output pair, displaying the structure of the neural network 301, and grasping the progress of learning (for example, displaying and outputting model parameters at that time).

With the above processing, the weights Wij and Vjk of the trained neural network 603 are transmitted to the neural network calculating means 101, and the first neural network 301 and the second neural network 302 transmit The calculation is performed based on the “weight”.

Next, an example of a method of setting the value of the input / output pair of the storage means 601 is shown in FIG.

First, in s7-1, the value of the overall heat absorption coefficient (hereinafter, φc
g).

Next, in s7-2, a control amount (extraction temperature) t0 is calculated from the model 105 in which φcg is set.

At s7-3, a parameter error Δφcg is added to the parameter φcg set at s7-1.

The control amount t1 is calculated using the new parameter φcg to which the parameter error Δφcg has been added in s7-4.
Is calculated.

In s7-5, the difference between the control amount t0 calculated in s7-2 and the control amount t1 calculated in s7-4 is calculated.
“t1-t0” calculated in s7-3 and parameter error Δ
φcg is stored in the storage means as an input / output pair. s7
At −6, the parameter φcg becomes the upper limit φcg of the parameter.
It is determined whether or not max has been exceeded.

If it exceeds, the process ends, otherwise, s7
The processing is shifted to -3. With the above operation, the input / output pair is created and stored in the storage unit 601.

In the above description, the neural network operation means 101 and the neural network learning device 111 are configured as separate devices, but may be configured as one device. That is, for example, a configuration in which the neural network learning device 111 is incorporated in the neural network operation means 101 may be employed.

FIG. 8 shows a neural network operation means 101.
5 shows a configuration example in which a switch 801 for cutting off the output of the second neural network 302 is provided.

In this embodiment, the switch 801 is provided on the output side of the second neural network 302. When the switch 801 is turned off, the output of the second neural network 302 becomes invalid and the first neural network 302 becomes invalid. The output of the network 301 is output as it is from the neural network operation means 101 as a correction amount.

The switch 801 is realized by an electronic device such as an analog switch (which is turned on when a predetermined voltage or more is input).
Opening and closing of the switch 801 may be configured such that, when the user arbitrarily sets the switch 801 through the input unit 108, a predetermined signal is supplied to the analog switch in response to the setting.

Further, a configuration is also conceivable in which the control amount detected by the thermometer 203 approaches the target value and is automatically turned on. Such automatic control includes, for example, a CPU,
ROM (predetermined program is pre-installed), R
It can be realized by means configured to have AM and the like.

Next, FIG. 9 shows a flowchart of processing for automatically opening and closing the switch 801.

First, in s9-1, the difference between the control amount (hereinafter, “y”) detected by the thermometer 203 and the control amount (hereinafter, “yd”) calculated using the model 105, and the control Parameters other than the quantity, furnace temperature, insertion temperature,
The overall heat absorption coefficient is input to the neural network 301, and the calculation result is stored as the correction amount 1.

In s9-2, the switch 801 is turned on.
It is determined whether or not to perform.

The following equation can be considered as an example of a calculation equation for the determination.

[0107]

[Equation 11]

Here, Δyth is a threshold value for determining whether the switch is on or off, and the switch should be turned on when “y−yd” satisfies the above equation. Good.

According to the above-mentioned determination formula, when the accuracy of the error calculated by the second neural network 302 is not good, when the input value is far from near zero (when the input value is near “0”), the first The output of the neural network 301 is directly a control amount, and the second neural network 3
02 is a near-zero input value from which an error can be calculated with high accuracy, and the output of the second neural network 302 is subtracted.

Next, if it is determined in s9-2 that the power should be turned off, the process proceeds to s9-5, and the correction amount 1 is output to the model 105 as the correction amount 2. If it is determined that the power should be turned on, the process proceeds to s9-3, where "0" is input to the input corresponding to "y-yd", and the second neural network 302 performs an operation.

In s9-4, the correction amount 2 is obtained by subtracting the error amount that is the calculation result obtained in s9-3 from the correction amount 1 stored in s9-1. And s9-5
To output the correction amount 2, and the process ends.

Next, an embodiment having a limiter for limiting the output of the neural network operation means 101 so as not to exceed the output range of the output signal given by the input / output pair will be described.

FIG. 10 shows a configuration example of the neural network operation means 101 at this time.

[0114] The limiter 1001 receives the results of calculations by the first neural network and the second neural network. The processing of the limiter 1001 is, for example, when the above-mentioned operation result exceeds the maximum value or the minimum value of the output signal of the input / output pair, the output of the neural network operation means 101 is set to the maximum value or the minimum value of the output signal. If not, the calculation result is output directly from the neural network calculation means 101.

The following equation shows an example of an equation for determining the output of the limiter 1001 at this time.

The limiter 1001 may be realized by hardware using electronic devices such as an operation amplifier and a resistor, or may be a ROM, a CPU, an R, or the like.
A configuration having an AM may be used, and the processing may be executed by a program stored in the ROM.

[0117]

(Equation 12)

[0118]

(Equation 13)

[0119]

[Equation 14]

Here, O is the operation result by the first neural network and the second neural network, and Ymax and Ymin are the maximum value and the minimum value of the output signal of the input / output pair, respectively. According to the above equation, the output range of the neural network operation means 101 can be limited, so that the stability of the control system is guaranteed.

Next, FIG. 11 shows a simulation result of tuning of model parameters by the neural network control device 100 of the present embodiment.

In this simulation, the value of the overall heat absorption coefficient φcg of the actual furnace was set to “0.5” and the model 10
The overall heat absorption coefficient φcg set to 5 is “0.65”
Set to.

This figure shows that the neural network operating means 101 according to the present invention replaces φcg of the model 105 with φ of the actual furnace.
The state of tuning is shown in cg.

As a comparison object, a neural network control device in which the neural network operation means 101 is composed of only the first neural network 301 is employed as conventional means.

Δθout is an extraction temperature error between the measured temperature of the slab 204 and the target temperature, and becomes 0 when the target temperature is reached.

According to the result of tuning by the conventional means, φcg of the model 105 converges to a value deviating from φcg of the actual furnace, and the extraction temperature error remains, whereas the present invention provides For example, as the number of times of tuning increases, φcg of the actual furnace is reproduced, and finally the extraction temperature error becomes almost “0”, and tuning is successful.

In the above embodiment, the neural network operation means has one output, but the output value is plural (for example, when a control object requiring a plurality of correction amounts or operation amounts is considered). You may. This is true for all embodiments.

FIG. 1 shows a configuration having two neural networks. However, in addition to the first neural network, the second neural network is not provided. (An output value corresponding to a specific value output by the second neural network) may be obtained in advance, and this value may be subtracted from the output value of the neural network.

Next, FIG. 12 shows an embodiment in which the neural network control device according to the present invention is used as a feedback controller.

[0130] The neural network control device 1200 includes the neural network operation means 101 and the target value storage section 106.

[0131] The neural network operation means 101
, And a second neural network 302.

The target value storage unit 106 stores a target value of the control amount, and is realized by, for example, a RAM.

The neural network control device 1200 uses the deviation obtained by subtracting the target value from the control amount taken from the first sensor 1203 as a first input value other than the control amount taken from the second sensor 1204. Control target 1
The state quantity 201 is given to the first neural network 301 as a second input value. Further, given “0” to the second input value and the first input value,
It is provided to the second neural network 302.

Then, the neural network operation means determines the difference between the outputs of the first and second neural networks as an operation amount for making the control amount equal to the target value, and
1 is output to the actuator 1202 included in 1.

Now, the neural network control device 1200
The operation of the neural network control device 1200 will be described in detail by using an application example in which is applied to chemical injection control of a water purification plant.

FIG. 13 shows a diagram corresponding to the control target 1201.
The block diagram of the water purification plant is shown.

The actuator 1202 shown in FIG. 12 is a solidified material injector 1301 for inserting the solidified material into the first layer 1304.
Corresponding to The first sensor 1203 corresponds to the sensor 1303 for detecting the solidified material concentration and the impurity concentration, and the second sensor 1204 corresponds to the sensor 1302 for detecting the temperature of the water purification plant 1300.

[0138] The plant into which the source water was injected is the first layer 1
At 304, the solidified material is injected by the solidified material injector 1301 to solidify impurities contained in the source water.

The precipitate is filtered through the second layer 1305 and the third layer 1306, and purified as tap water.

In the third layer 1306, the concentration of impurities and the concentration of solidified material of water extracted by the sensor 1302 are detected.

The temperature of the plant is measured by a sensor 1302.
Capture by In the water purification plant, the concentration of impurities and the concentration of solidified material are controlled amounts, and the neural network controller 12
00 is the time until the impurity concentration and the solidification material concentration become “0”.
The operation amount is sent to the coagulating material injector 1301.

FIG. 14 shows the neural network controller 13.
00 shows the flow of processing.

In this embodiment, a case will be described in which the second neural network 302 calculates an error when the control amount matches the target value.

First, in s14-1, the first neural network 301 sends the first input value, the sensor 1
At 303, a deviation between the acquired impurity concentration and solidified material concentration and a target value (“0” in this embodiment) is input.
Further, the temperature of the plant taken by the sensor 1302 is input as a second input value as a signal affecting the relationship between the deviation and an appropriate operation amount corresponding to the deviation, and the calculation result is stored as the operation amount 1. .

Next, in s14-2, "0" is input to the second neural network 302 as an input corresponding to the deviation between the impurity concentration and the solidified material concentration input in s14-1 and the target value. Perform the same operation.

The operation result is stored as an error amount.

In s14-3, the value obtained by subtracting the error amount calculated in s14-2 from the value of the operation amount 1 stored in s14-1 is set as the operation amount 2. Then, in s14-4, the value of the manipulated variable 2 is output to the plant (specifically, to 1301 corresponding to the actuator).

An input / output pair of the storage means 601 used for learning of the first neural network 301 is created in an actual plant based on input signals and correspondingly, for example, operation amounts of a skilled driver. Good. When a model for the water purification plant 1300 is created, an input / output pair can be created by simulation, experiment, or the like using the model.

Next, an embodiment in which the output of the neural network 101 is determined only by the control amount will be described.

FIG. 15 shows a configuration example of this embodiment.

The neural network control device 1500 includes the neural network 301, the offset storage unit 15
01, a target value storage unit 107 is provided.

The offset storage unit 1501 and the target value storage unit 107 are realized by, for example, a RAM.

First, the overall processing will be described.

The neural network 301 fetches a value obtained by subtracting the target value from the control amount fetched from the sensor 1504, and calculates an operation amount appropriate for reducing this value.

The offset value stored in the offset storage means 1501 is subtracted from the calculated operation amount, and the value is output to the actuator 1503 of the controlled object 1502.

In the offset storage means 1501, an output value when "0" is input as a deviation to the neural network 301 is stored in advance.

FIG. 16 shows a configuration method of the offset storage means 1501.

First, in s16-1, "0" which is the value when each control amount matches the target value is input to the neural network 301.

Next, in s16-2, the output of the neural network 301 at this time is stored in the offset storage means 1501.

With the above processing, the offset storage means 1
The data of the offset value stored in 501 is created and stored.

Next, the neural network controller 15
In this embodiment, 00 is applied to the inverted pendulum control.

FIG. 17 shows a configuration example of the control target 1502. The inverted pendulum 1700 is mounted on a carriage 1701, and the angle θ and the angular velocity dθ / dt are captured by an angle sensor 1703 corresponding to the sensor 1404. For example, a gyro capable of detecting an angular velocity may be used as the angle sensor.

The angle θ and the angular velocity dθ / dt are controlled to be “0” by the force F generated by the driving unit 1702 corresponding to the actuator 1503.

In the case of an inverted pendulum, the control amount is an angle θ and an angular velocity dθ / dt, the target value of the control amount is “0”, and the operation amount is a force F.

Further, for example, when the driving section 1702 is provided with a motor and driving wheels driven by the motor, the force F, which is the operation amount, is given by the output torque of the motor and the like. To give, for example,
What is necessary is just to inject a predetermined drive current into a motor.

Next, FIG. 18 shows a flow of processing by the neural network control device 1500.

First, in s18-1, the target value, the angle θ and the angular velocity dθ / d captured by the angle sensor 1302 are obtained.
The respective deviation from t is calculated using the neural network 3
01 and calculate the obtained operation result.

Next, in s18-2, the value stored in the offset storage means 1501 is subtracted from the calculation result, and the subtracted value is set as the operation amount F.

Next, in s18-3, the operation amount F is output to the drive section 1702 of the control target 1502.

That is, for example, a current is injected into the motor so as to give the operation amount F.

In this embodiment, the method of realizing the neural network control device 1500 by software has been described. However, the neural network 301 and the offset storage means 1501 are provided separately,
It may be realized by hardware using a logic element such as a MOS.

Further, similarly to the above-described embodiment, a switch 80 for selecting whether or not to subtract the offset value stored in the offset storage means 1501 depending on the situation.
1 may be provided.

Next, an embodiment in the case where there are a plurality of outputs of the neural network 301 will be described.

FIG. 19 shows a PID (P is proportional, I
Represents an integral, and D represents a derivative, and refers to a control method combining these three operations.) An embodiment applied to automatic tuning of a gain parameter is shown.

The neural network control device 1900 calculates P
ID controller 1901, input waveform generator 1903,
It comprises an analysis unit 1904, a neural network 301, an offset storage unit 1501, and a target value storage unit 107. Further, an input unit 108 realized by a keyboard, a mouse, or the like is also provided.

Each component is, for example, a CPU, a ROM,
(A program for performing a predetermined process is stored in advance),
It can be realized by an electronic device such as a RAM or various CMOSs.

First, the overall processing will be described.

In this embodiment, as an example, a case where step input is performed as an input waveform will be described.

The PID controller 1901 is controlled so that the output waveform of the controlled object 1902 in response to the step input generated from the input waveform generator 1903 matches the ideal response.
It is necessary to tune the gains kP, kI, and kD.

The target value storage unit 1 is input via the input unit 108.
At 07, the response waveform of the ideal step response is stored in advance.

The control amount of the control object 1902 is
04 and an output waveform is extracted.

The method of extracting the output waveform is related to the method of generating the input waveform and the manner of storing the target value. For example, sampling the waveform at the Nyquist interval, examining the distribution of the frequency spectrum, and Various methods, such as a method of sampling the magnitude of the spectrum at predetermined frequency intervals, can be considered, and any method may be used as long as the waveform is extracted so as to be reproduced.

The difference between the output waveform extracted in this way and the ideal waveform is input to the neural network 301. The neural network 301 outputs the gain correction amounts ΔkP, ΔkI, and ΔkD, and the gains kP, k stored in the offset storage unit 1501 are output.
The value obtained by subtracting the offset value corresponding to each of I, kD from the correction amounts ΔkP, ΔkI, ΔkD is P
Processing to be added to the gain of the ID controller 1901 is performed.

The offset value stored in the offset storage means 1501 is the output value of the neural network 301 when "0" is input as the deviation between the output waveform and the ideal waveform.

The parameter representing the gain has been corrected.
From the PID controller 1901, the control target 1902
Then, the operation amount is transmitted, and the control amount of the control target 1902 is input to the analysis unit 1904 again.

The above processing is repeated to tune the gain parameter.

FIG. 20 shows a flowchart of the processing by the offset storage means of this embodiment.

First, in s20-1, the calculation result by the neural network 301 is stored in ΔkP, ΔkI, and ΔkD as the correction amount of the gain parameter.

Next, at s20-2, the values ΔkP0, ΔkI0, ΔkP0 stored in the offset storage means 1501 are set.
kD0 is taken out and stored as a correction amount, Δk
ΔkP0, ΔkI extracted from P, ΔkI, ΔkD
0 and ΔkD0 are subtracted from the correction amount ΔkP,
Let ΔkI and ΔkD.

Then, in s20-3, the correction amounts ΔkP, ΔkP
If kI and ΔkD are output to the control target 1902, a series of processing is completed.

[0191]

According to the present invention, the value obtained by canceling the output error of the first neural network with respect to the designated value by the output of the second neural network is used as the output of the neural network operation means. An output error near the designated input value can be reduced. Therefore, the neural network control device can output the command value, the operation amount, and the like with high accuracy.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an example of a neural network control device according to the present invention.

FIG. 2 is a configuration diagram of a heating furnace to be controlled.

FIG. 3 is a configuration diagram of a neural network operation unit.

FIG. 4 is a flowchart showing a processing algorithm of a neural network operation means.

FIG. 5 is a configuration diagram of an example of a neural network.

FIG. 6 is a configuration diagram of a neural network learning device.

FIG. 7 is a flowchart showing an algorithm for creating an input / output pair.

FIG. 8 is a configuration diagram of an example of a neural network control device provided with a switch.

FIG. 9 is a flowchart showing an algorithm of a switch operation.

FIG. 10 is a configuration diagram of a neural network operation unit provided with a limiter.

FIG. 11 is an explanatory diagram of a simulation result.

FIG. 12 is a configuration diagram of an embodiment in which the present invention is applied to a controller.

FIG. 13 is a configuration diagram of a water purification plant.

FIG. 14 is a flowchart showing the algorithm of the neural network operation means.

FIG. 15 is a configuration diagram of a neural network control device incorporating an offset storage means.

FIG. 16 is a flowchart showing an algorithm of the offset storage means.

FIG. 17 is a configuration diagram of an inverted pendulum.

FIG. 18 is a flowchart showing the algorithm of the neural network operation means.

FIG. 19 is a configuration diagram of an embodiment in which the present invention is applied to a PID controller.

FIG. 20 is a flowchart showing the algorithm of the neural network operation means.

[Explanation of symbols]

100: Neural network control device, 101: Neural network operation means, 102: Command value calculation means, 105 ...
Model, 111 ... neural network learning device, 301 ...
1st neural network, 302 ... 2nd neural network, 801 ... switch, 1001 ... limiter, 1501 ... offset storage means

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP 5-181506 (JP, A) JP 5-197401 (JP, A) JP 5-150803 (JP, A) JP 4-A 15704 (JP, A) JP-A-5-333902 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G05B 11/00-13/04 G06F 15/18

Claims (12)

(57) [Claims] At least one or more first input values and control values, each having an input layer, an intermediate layer, and an output layer, and being a deviation amount between a control target value of a control target and a detected control amount. At least one or more second states that are the state quantities of the subject
A neural network-equipped control device including a neural network that receives an input value of (i) from an input layer and outputs an operation amount given to a control target from an output layer, comprising a new input layer, an intermediate layer, and an output layer. An input layer of the new neural network receives the second input value and a predetermined value corresponding to the first input value, and receives the received value. Output from the output layer by an operation performed by the intermediate layer and the output layer based on the subtraction means, wherein the subtraction means calculates the difference between the output values of the two neural networks and outputs the result as the manipulated variable. A control device with a built-in neural network. 2. A method using a neural network having an input layer, an intermediate layer, and an output layer, wherein at least one of a deviation amount between a control target value of a control target and a detected control amount is used.
One or more, which is the first input value and the state quantity with the controlled object, at least given one or more second input values to the input layer, the control by the neural network to output an operation amount to be given to the controlled object from the output layer A new neural network comprising an input layer, a hidden layer, and an output layer is provided, wherein the second input value and the first neural network are provided in an input layer of the new neural network. Given a predetermined specific value corresponding to the input value, based on the received value, the middle layer,
A control method using a neural network, comprising: outputting an output value from an output layer by an operation performed by an output layer; calculating a difference between output values of the two neural networks; and outputting the difference as an operation amount. Wherein the input layer, an intermediate layer, is configured to have an output layer, the input layer accepts one or more input values even without low,
In a neural network built-in control device having a neural network that outputs an operation amount given to a control target to an output layer, the neural network is configured such that a predetermined specific value is used as an input value.
Store output value of neural network as offset value
Offset storage means, the output value of the neural network and the offset
Calculate the difference with the value and output this as the manipulated variable
A neural network built-in control device , comprising: an offset calculating means . 4. Using a neural network having an input layer, a hidden layer, and an output layer, providing at least one or more input values to the input layer, and outputting an operation amount given to the control target to the output layer. In a control method using a neural network, an offset storage unit that stores an output value of the neural network as an offset value when a predetermined specific value is set as an input value is provided, and the output value of the neural network and the offset are provided. Control using a neural network, which calculates a difference from the value and outputs the calculated value as the manipulated variable.
How . 5. The neural network according to claim 1, wherein said new neural network is provided with selecting means for selecting whether or not to connect an output value of said neural network to said subtracting means. Control device with built-in network. 6. The control device with built-in neural network according to claim 3 , further comprising selection means for selecting whether to activate said offset calculation means. 7. The built-in neural network according to claim 5, further comprising: a starting unit that starts the selecting unit so as to select a predetermined selection item at a predetermined timing. Control device. 8. The apparatus according to claim 1, further comprising limiter means, wherein said limiter means limits one of an upper limit value and a lower limit value of an output value from said subtraction means. A control device with a built-in neural network. 9. The apparatus according to claim 3 , further comprising limiter means, wherein the limiter means sets one of an upper limit value and a lower limit value of an output value from said offset calculating means. A control device with a built-in neural network, which is controlled. 10. The method according to claim 5 , wherein the input value is a deviation amount between a control target value of a control target and a detected control amount, and the predetermined specific value is zero. Control device with built-in neural network. 11. When a control target is represented by a model having an input layer, a middle layer, and an output layer, and a control target is represented by a mathematical expression including parameters that cannot be accurately determined, a control target value and a detection target value are detected.
The input layer accepts at least one or more first input values that are deviation amounts from the controlled amount and at least one or more second input values that are state quantities of the controlled object. A new neural network comprising an input layer, an intermediate layer, and an output layer, comprising a neural network that outputs a correction amount of the parameter of the model to an output layer by an operation performed by the intermediate layer and the output layer based on the neural network. And a subtraction means. The input layer of the new neural network receives the second input value and a predetermined value corresponding to the first input value, and sets the received value to the received value. A predetermined output value is output from the output layer by an operation performed by the intermediate layer and the output layer based on the output value. The subtraction means calculates a difference between the output values of the two neural networks, Neural network internal control device and outputs the operation result as a correction amount of parameters. 12. A method using a neural network having an input layer, a middle layer, and an output layer, wherein at least a deviation between a control target value of a control target and a detected control amount is determined.
A state quantity with the one or more first input value and the controlled object, using a neural network to output at least given one or more second input values to the input layer, an output layer with the operation amount applied to the control object in the control method, first, to the input layer of the neural network, the second input value, and gives one specific value determined in advance in correspondence to the first input value, based on the values accepted intermediate The output value is output from the output layer by the operation performed by the layer and the output layer, and the output value is stored. Next, the difference between the output value of the neural network and the stored output value is calculated. And outputting the manipulated variable as the manipulated variable.