JPH07121206A - Method for control by neural network, and internal control unit - Google Patents

Abstract

PURPOSE:To reduce the output error of a 1st neural network corresponding to a specified value by regarding the value, obtained by canceling the output error with the output of a 2nd neural network, as the output of a neural network arithmetic means. CONSTITUTION:The neural network arithmetic means 101 has 1st and 2nd neural networks 301 and 302; and the neural network 301 calculates a correction quantity of, for example, a total heat absorption coefficient and the neural network 302 calculates the error quantity calculated by the network 301. To reduce the difference, the correction quantity of the total heat absorption coefficient is calculated and outputted to a model. The model varies the total heat absorption coefficient as an indeterminate parameter constituting a mathematical expression on the basis of the inputted correction quantity. Namely, the value obtained by subtracting the difference from the output of the network 302 from the output of the network 301 becomes the output of the neural network arithmetic means 101, and consequently the total heat absorption coefficient of the model is corrected.

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 various controlled objects.

[0002]

2. Description of the Related Art Conventionally, in a neural network used to control various controlled objects, the output from the neural network is sent as it is to a controller for controlling the controlled object, which is provided in connection with the controlled object. Command value, the amount of operation sent to the controlled object, etc.

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 with respect to the input signal becomes a predetermined value or less. Then, the neural network that has been subjected to such learning is used to control the control items to be controlled.

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

Then, a certain control amount is input to the learned neural network, and the output from the neural network at that time is used as an operation amount, which is directly input to the controller provided in the control device to be controlled. It controlled the control items.

[0006]

However, the above-mentioned conventional techniques have the following problems, for example.

Since the output from the neural network directly corresponds to the command value to the controller for controlling the controlled object and the operation amount sent to the controlled object, the output error of the neural network is the command value as it is. And the accuracy of control is
It was significantly reduced, and the control system was destabilized by the error.

SUMMARY OF THE INVENTION An object of the present invention is to provide a control means method having a built-in neural net for reducing the output error of the neural network and controlling the controlled object with high accuracy.

[0009]

[Means for Solving the Problems] In order to solve the above problems, the following means can be considered.

It is configured to have an input layer, an intermediate layer, and an output layer, and is a deviation amount from a control target value of a control target, which is at least one or more first input value and a state amount which the control target has. A neural network built-in control device having a neural network that receives at least one second input value from an input layer and outputs an operation amount given to a controlled object from an output layer, having an input layer, an intermediate layer, and an output layer. And a subtraction means, the input layer of the new neural network has a predetermined specific value corresponding to the second input value and the first input value. Is received, the output value is output from the output layer by the calculation 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 first input value and control which is a deviation amount from a control target value of a controlled object are used. In a control method by a neural network in which at least one second input value, which is a state quantity of an object, is given to an input layer and an operation amount given to a controlled object is outputted from an output layer, an input layer, an intermediate layer,
A new neural network having an output layer is prepared, and the input layer of the new neural network is set in advance corresponding to the second input value and the first input value. A specific value is given, and the output value is output from the output layer by the calculation performed by the intermediate layer and the output layer based on the received value, and the 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 which performs a normal neural network operation, a second neural network which newly calculates the output error of the first neural network is newly provided. It is solved by preparing.

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

As a result, 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 obtains an accurate output value corresponding to this input value neighborhood. It can be calculated 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 controlled object 104 and outputs the command value to the controller 103.

The neural network control device 100 comprises a neural network computing means 101, a target value storage section 107, a command value calculating means 102, and an input means 108.

The neural network calculation means 101, the target value storage section 107, and the command value calculation means 102 are, for example, CP.
It is realized by an electronic device such as U, ROM (a program that performs a predetermined process is built in in advance), RAM, and various CMOSs. The input unit 108 can be realized by, for example, a keyboard or a mouse.

Although not shown in the drawing, it is also possible to consider a structure provided with a display means for displaying various information such as the command value. In this case, the display means is realized by, for example, a CRT, EL display, liquid crystal display, or the like.

The neural network computing means 101
Is equipped with two neural networks, and the command value calculation means 102 is equipped with a model 105.

The model 105 expresses the controlled object by a mathematical expression including an uncertain parameter, and the mathematical expression can be used to determine the manipulated variable with respect to the control target value.

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

The controller 103 is, for example, a CPU or R
It is realized by an electronic device such as an OM (a program that performs a predetermined process is built-in in advance), a RAM, and various CMOSs.

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

Specific examples of the controlled object, 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 a 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 previously built in), a RAM, various CMOSs and the like. This is one chip and L
It may be realized by SI or the like.

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

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

First, the neural network control device 1
The processing flow 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 unit 10.
6. It is configured to have a neural network computing means 101 for computing by a neural network. The target value storage unit 106 stores a value that is a target control amount of the controlled object 104, and the value is, for example, the input unit 10.
It is only necessary to set and change the setting via 8.

The command value calculation means 102 is a controlled object 104.
Is calculated based on the model 105 using a mathematical expression so that the command value to be output to the controller 103 is calculated so that the output of the controlled object 104 matches the control target value.

The neural network computing means 101 is the first
Using the difference between the control amount captured by the first sensor 109 and the control amount calculated using the model 105, which is the input value of, the uncertainty of the model 105 is determined so as to reduce this difference. Correct the correct parameters.

If another value (for example, a physical quantity such as temperature) representing the state of the controlled object has an influence on the relationship between the difference and the correction amount of the parameter, as shown in FIG. Is taken as the second input value by the second sensor 110 and input to the neural network computing means 101. The physical quantity taken in by the second sensor 110 is input to the first and second neural networks exactly the same.

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

Next, the detailed configuration and processing of each unit will be described.

In this embodiment, as the controlled object 104, a heating furnace plant in hot rolling is adopted.

FIG. 2 shows a structural example of the heating furnace. Heating furnace 2
At 00, the slab 204, which is the material to be heated, is inserted from the heating furnace inlet and heated by the gas burner 201 while moving to the heating furnace exit side. In correspondence with the configuration of FIG. 1, the actuator 106 is a gas burner 201.
The first sensor 109 corresponds to the thermometer 203 that measures the surface temperature at the outlet of the slab 204. The second sensor 110 corresponds to a thermometer 205 that measures the insertion temperature of the slab 204 and a thermometer 202 that measures 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 radiant heat transfer given to the slab 204 is modeled. A model corresponding to the amount of radiant heat transfer is represented by Stefan Boltzmann's 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 transfer from the furnace to the slab,
T is the temperature inside the furnace, which is detected by the thermometer 202. θ is the temperature φcg of the slab 204, and is the heat transfer coefficient called the overall heat absorption coefficient. The value of the overall heat absorption coefficient is
It is generally very difficult to calculate exactly, and in the present embodiment, it is the purpose of the neural network computing means 101 to tune this overall heat absorption coefficient to match the overall heat absorption coefficient of the heating furnace 200. Is.

The command value calculating means 102 is the model 10
Based on 5, the furnace temperature command value is calculated so that the temperature at the heating furnace outlet of the slab 204 satisfies the target temperature, and the calculation result is output to the controller 103. The controller 103 controls the gas burner 2 so that the furnace temperature and the command value match.
01 fuel flow rate is controlled.

FIG. 3 shows the neural network computing means 101.
A configuration example of is shown.

The neural network computing means 101 is the first
And a second neural network 302.

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

As a first input value, the first neural network 301 receives the difference between the control amount actually taken in from the thermometer 203 and the control amount calculated using the model 105. Further, as the second input values, the in-furnace temperature and the insertion temperature of the slab 202, which are respectively taken in 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 forming the mathematical formula, is changed based on the input correction amount.

The second neural network 302 is
When the correction amount, which is the difference between the control amount detected by the thermometer 203 and the control amount calculated using the model 105, has the same configuration as the first neural network 301 and is “0”. 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 inputs are the first neural network. The second input value which is the same value as 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 taken to obtain the neural network computing 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 structure, a specific value "corresponding to" control amount-control amount by model "which is an input signal of the first neural network 301 is specified. 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 the processing performed by the neural network computing means 101.

First, in s4-1, the difference between the control amount detected by the thermometer 203 (hereinafter referred to as "y") and the control amount calculated using the model 105 (hereinafter referred to as "yd"). In addition, 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, etc., is input to the first neural network 301, and the calculation result is corrected Set to 1.

Next, in s4-2, "0" is input to the input corresponding to "y-yd", and the second neural network 302 is caused to execute the operation. The calculation result at this time is the 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, the neural network 301 (in this embodiment, the first neural network and the second neural network have the same configuration, only the first neural network will be described) will be described in detail.

FIG. 5 shows the first neural network 3
A configuration example of 01 is shown.

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

The first neural network 301 is
Neurons 501 are configured by being connected in layers by synapses 502. In the example shown in FIG. 5, the input layer 503,
Although a network having a three-layer structure 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.

If 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.

[0063]

[Equation 2]

[0064]

[Equation 3]

[0065]

[Equation 4]

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

[0067]

[Equation 5]

[0068]

[Equation 6]

[0069]

[Equation 7]

It becomes Vjk is a weight connecting the jth intermediate layer neuron and the kth output layer neuron, and θk is a constant corresponding to the output layer neuron.

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

In the neural network computing means 101, the second neural network 302 has the same structure as the first neural network 301, but the thermometer 2
A configuration in which the neuron in the input layer 503 corresponding to the difference between the control amount detected from 03 and the control amount calculated using the model 105 and the synapse connected to this neuron may be omitted.

FIG. 6 shows a neural network learning device 111.
A configuration example of is shown.

In the neural network learning device 111, in addition to the neural network 603 having the same structure as the first neural network 301, the storage means 6
01, learning means 602, input means 604, output means 60
5 is provided. These are buses 60
They are coupled via 6 to exchange necessary signals.

The storage means 601 is realized by, for example, a RAM or a disk drive device equipped with a disk as a storage medium, and the learning means 602 is, for example, a CPU or R.
It is realized by an electronic device such as an OM (a program that performs a predetermined process is built-in in advance), a RAM, and various CMOSs. The input unit 604 is, for example, a keyboard, a mouse, etc., and the output unit 605 is, for example,
It is realized by CRT, EL display, liquid crystal display and the like.

The storage means 601 stores at least one input signal and output signal (hereinafter referred to as “input / output pair”) to be learned by the neural network 603.

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

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

Various methods have been proposed for correcting Wj and Vj.

As a typical method, there is an error back-propagation learning method using the steepest descent method. This learning method is performed by the output signal of the input / output pair and the neural network 60 corresponding thereto.
This is a method of modifying the weights Wj and Vj in the direction of decreasing this 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 amounts with the output of 3. .

The following equation shows this energy function.

[0082]

[Equation 8]

Tk is an output signal of the 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 by 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.

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. Item 533-Item 536.
October. 1986).

Now, the learning means 602 has the storage means 601.
The input / output pairs stored in are sequentially fetched and the above-mentioned operation is repeated. Finally, the neural network 603 stores the relationship between the input and the output of the storage unit 601.

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

By the above processing, the weights Wij and Vjk of the neural network 603 for which learning has been completed are transmitted to the neural network computing means 101, and the first neural network 301 and the second neural network 302 are transmitted. Calculation is performed based on the "weight".

Next, FIG. 7 shows an example of a method of setting the value of the input / output pair of the storage means 601.

First, in s7-1, the value of the overall heat absorption coefficient to be tuned (hereinafter, φc
g)) is set.

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

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

The controlled variable t1 is added using the new parameter φcg to which the parameter error Δφcg is added in s7-4.
To calculate.

At s7-5, the difference between the control amount t0 calculated at s7-2 and the control amount t1 calculated at s7-4 is calculated,
"t1-t0" calculated in s7-3 and the parameter error Δ
φcg is stored in the storage means as an input / output pair. s7
At −6, the parameter φcg is the upper limit φcg of the parameter.
It is judged whether or not it exceeds max.

If it exceeds, it ends. If not, s7
The process is moved to -3. By the above operation, the input / output pair is created and stored in the storage unit 601.

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

FIG. 8 shows the neural network computing means 101.
In, a configuration example of a switch 801 for cutting off the output of the second neural network 302 is shown.

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

The switch 801 is realized by an electronic device such as an analog switch (means that is turned on when a voltage higher than a predetermined value is input),
The opening and closing of the switch 801 may be configured to give a predetermined signal to the analog switch in response to the setting made by the user through the input means 108.

Further, a configuration in which the control amount detected by the thermometer 203 is detected to be close to the target value and is automatically turned on is also conceivable. Such automatic control is performed by, for example, a CPU,
ROM (predetermined program built in), R
It can be realized by means including an AM or 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 controlled variable (hereinafter "y") detected by the thermometer 203 and the controlled variable (hereinafter "yd") calculated using the model 105, and the control Parameter group other than quantity, furnace temperature, insertion temperature,
The overall heat absorption coefficient is input to the neural network 301, and the calculation result is saved as the correction amount 1.

At s9-2, the switch 801 is turned "on".
It is determined whether or not to do it.

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

[0107]

[Equation 11]

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

According to the above judgment formula, the accuracy of the error calculated by the second neural network 302 is not good, and the first value is detected when the input value is far from near zero (when the input value is near "0"). The output of the neural network 301 of FIG.
02 is a near-zero input value capable of calculating an error with high accuracy, and the output of the second neural network 302 is subtracted.

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

At s9-4, the correction amount 2 is obtained by subtracting the error amount which is the calculation result obtained at s9-3 from the correction amount 1 stored at s9-1. And s9-5
Then, the correction amount 2 is output, and the process ends.

Next, an embodiment provided with a limiter for limiting the output of the neural network computing 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 computing means 101 at this time.

The limiter 1001 receives the calculation results of the first neural network and the second neural network. The processing of the limiter 1001, for example, sets the output of the neural network calculation means 101 to the maximum value or the minimum value of the output signal when the above-mentioned calculation result exceeds the maximum value or the minimum value of the output signal of the input / output pair. If not, the calculation result is directly output 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 is realized by hardware, using an electronic device such as an operational amplifier and a resistor, a ROM, a CPU, and an R.
The configuration may be such that the AM is provided and the processing may be executed by the program stored in the ROM.

[0117]

[Equation 12]

[0118]

[Equation 13]

[0119]

[Equation 14]

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

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

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

In this figure, the neural network computing means 101 according to the present invention calculates φcg of the model 105 as φ of the actual furnace.
The tuning is shown in cg.

As a comparison target, a neural network control device in which the neural network calculation means 101 is composed of only the first neural network 301 is adopted as a 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 tuning result of the conventional means, φcg of the model 105 converges to a value deviated from φcg of the actual furnace and the extraction temperature error remains. 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 the tuning is successful.

In the above embodiment, the neural network computing means has one output, but the output value is plural (for example, in the case of considering a control object requiring a plurality of correction amounts or operation amounts). May be. This is true in all the examples.

Although FIG. 1 shows a configuration having two neural networks, the first neural network is not provided with a second neural network in addition to the first neural network. May be obtained (the output value corresponding to the specific value output by the second neural network), 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.

The neural network control device 1200 is constructed to have a neural network computing means 101 and a target value storage unit 106.

The neural network computing means 101 has a first
The neural network 301 and the second neural network 302 are included.

Further, the target value storage unit 106 stores the target value of the control amount, which is realized by, for example, a RAM or the like.

The neural network control device 1200 uses the deviation obtained by subtracting the target value from the control amount fetched from the first sensor 1203 as the first input value, except for the control amount fetched from the second sensor 1204. Control target 1
The state quantity 201 is given to the first neural network 301 as the second input value. In addition, “0” given to the second input value and the first input value is
It is given to the second neural network 302.

Then, the neural network computing means uses the difference between the outputs of the first and second neural networks as the manipulated variable for matching the controlled variable with the target value, and the controlled object 120.
1 is output to the actuator 1202.

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

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

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

The plant injected with the source water is the first layer 1
At 304, the coagulant is injected by the coagulant injector 1301 to coagulate 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 impurity concentration of the water extracted by the sensor 1302 and the coagulant concentration are detected.

Further, the temperature of the plant is detected by the sensor 1302.
Take in by. In the water purification plant, the impurity concentration and the coagulant concentration are control amounts, and the neural network control device 12
00 is until the impurity concentration and the coagulant concentration become “0”,
The operation amount is sent to the coagulant injector 1301.

FIG. 14 shows the neural network control device 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 sensor 1 which is the first input value is input to the first neural network 301.
Deviations between the impurity concentration and coagulant concentration captured by 303 and the target value (“0” in this embodiment) are input.
Further, the temperature of the plant captured by the sensor 1302 is input as a second input value as a signal that affects the relationship between the deviation and the appropriate manipulated variable, and the calculation result is saved as the manipulated variable 1. .

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

The calculation result is saved as an error amount.

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

The input / output pair of the storage means 601 used for learning of the first neural network 301 is prepared in the actual plant based on the input signal and corresponding to it, for example, based on the operation amount of the skilled driver. Good. Further, when a model for the water purification plant 1300 is created, it is also possible to create an input / output pair by simulation, experiment, etc. using this 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 structural example of this embodiment.

The neural network control device 1500 comprises a neural network 301 and an offset storage means 15.
01, the target value storage unit 107.

The offset storage means 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 takes in a value obtained by subtracting the target value from the control amount taken in by the sensor 1504, and calculates an operation amount suitable for reducing this value.

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

The offset storage means 1501 stores in advance the output value when the deviation "0" is input to the neural network 301.

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

First, in s16-1, the value "0" 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.

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

Next, the neural network control unit 15
An example in which 00 is applied to the inverted control of the inverted pendulum will be described.

FIG. 17 shows a configuration example of the controlled object 1502. The inverted pendulum 1700 is mounted on the carriage 1701, and the angle θ and the angular velocity dθ / dt are captured by the angle sensor 1703 corresponding to the sensor 1404. As the angle sensor, for example, a gyro that can detect the angular velocity may be used.

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

In the case of the inverted pendulum, the controlled variables are the angle θ and the angular velocity dθ / dt, the target value of the controlled variable is “0”, and the manipulated variable is the force F.

Further, for example, when the drive section 1702 is constituted by a motor and drive 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,
It suffices to inject a predetermined drive current into the motor.

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

First, in s18-1, the target value, the angle θ captured by the angle sensor 1302, and the angular velocity dθ / d.
The respective deviations from t are calculated by the neural network 3
Input to 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 taken as the manipulated variable F.

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

That is, for example, a current is injected into the motor so as to give the manipulated variable 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 individually provided, and these various C
It may be realized by hardware by a logic element such as MOS.

Further, similar to the above-described embodiment, the switch 80 which makes it possible to select 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 example in which the neural network 301 has a plurality of outputs will be described.

FIG. 19 shows the PID (P is proportional, I
Represents an integral and D represents a derivative, and refers to a control system in which these three operations are combined.) An example applied to the automatic tuning of a gain parameter is shown.

The neural network control device 1900 uses 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, input means 108 realized by a keyboard, a mouse, etc. is also provided.

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

First, the overall processing will be described.

In the present embodiment, as an example, a case of performing step input as an input waveform will be described.

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

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

The control amount of the controlled object 1902 is calculated by the analysis unit 19
It is input to 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 method of storing the target value. For example, sampling the waveform at Nyquist intervals or examining the distribution of the frequency spectrum, Various methods such as a method of sampling the magnitude of the spectrum at a predetermined frequency interval are conceivable, and any method may be used as long as the waveform can be extracted so as to be reproduced.

The difference between the output waveform thus extracted and the ideal waveform is input to the neural network 301. The neural network 301 outputs the gain correction amounts ΔkP, ΔkI, ΔkD, but the gains kP, k stored in the offset storage means 1501.
The offset value corresponding to each of I and kD is obtained by subtracting the correction amounts ΔkP, ΔkI, and ΔkD from P
Processing for adding 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 is modified,
From the PID controller 1901 to the control target 1902
The control amount of the controlled object 1902 is input to the analysis unit 1904 again.

The gain parameters are tuned by repeating the above processing.

FIG. 20 shows a flowchart of 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, in s20-2, the values stored in the offset storage means 1501, ΔkP0, ΔkI0, Δ.
Δk stored in kD0 as a correction amount
ΔkP0, ΔkI extracted from P, ΔkI, ΔkD
The values obtained by subtracting 0 and ΔkD0 from each other are used as correction amounts ΔkP and
Let ΔkI and ΔkD.

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

[0191]

According to the present invention, a 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 calculation means. The output error in the vicinity of the specified 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 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 which is a control target.

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

FIG. 4 is a flow chart showing a processing algorithm of the neural network computing 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 input / output pair creation algorithm.

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

FIG. 9 is a flow chart showing an algorithm of 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 simulation results.

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 flow chart showing an algorithm of a neural network computing means.

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

FIG. 16 is a flow chart showing an algorithm of offset storage means.

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

FIG. 18 is a flow chart showing an algorithm of the neural network computing 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 flow chart showing an algorithm of the neural network computing means.

[Explanation of symbols]

Reference numeral 100 ... Neural network control device, 101 ... Neural network calculation 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

Claims (12)

[Claims] 1. At least one first input value which is a deviation amount from a control target value of a controlled object and a state quantity which the controlled object has, which is configured by having an input layer, an intermediate layer and an output layer. In a control device with a built-in neural network that has a neural network that receives at least one or more second input values from the input layer and outputs the amount of operation given to the controlled object from the output layer, the input layer, the intermediate layer, and the output layer are A new neural network having the above configuration and a subtracting means are provided, and the 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 received, and 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. It calculates the difference between the output values, the neural network internal controller and outputs as the operation amount. 2. A neural network having an input layer, an intermediate layer, and an output layer is used, and at least one or more first input value and control which are deviation amounts from a control target value of a controlled object are used. At least 1, which is the state quantity that the subject has
In the control method by the neural network in which the second input value is given to the input layer and the operation amount given to the controlled object is outputted from the output layer, a new neural network having an input layer, an intermediate layer and an output layer is provided. A net is provided, and a predetermined specific value corresponding to the second input value and the first input value is given to the input layer of the new neural network, and an intermediate value is obtained based on the received value. layer,
An output value is output from the output layer by an operation performed by the output layer, a difference between the output values of the two neural networks is calculated, and the difference is output as the operation amount. 3. An operation configured to have an input layer, an intermediate layer, and an output layer, the input layer receiving at least one input value, which is a deviation amount from a control target value of a control target, and giving it to the control target. In a neural network built-in control device having a neural network for outputting a quantity to an output layer, an offset storing means and an offset calculating means for storing at least one set of a given specific value and an offset value corresponding thereto in advance are provided. The offset calculation means comprises a difference between the output value of the neural network and the offset value corresponding to the specific value when the value given to the neural network matches the specific value stored in the offset storage means. A control device with a built-in neural network, which calculates and outputs the calculated amount as the operation amount. 4. A neural network having an input layer, an intermediate layer, and an output layer is used, and at least one input value, which is a deviation amount from a control target value of a controlled object, is given to the input layer. In a control method by a neural network that outputs an operation amount given to a controlled object to an output layer, an offset storage means is prepared in advance for storing at least one set of a given specific value and an offset value corresponding thereto, When the value given to the neural network matches the specific value stored in the offset storage means, the difference between the output value of the neural network and the offset value corresponding to the specific value is calculated, and the difference is calculated. A control method by a neural network characterized by outputting as an operation amount. 5. The neural network according to claim 1, wherein the new neural network is provided with a selecting means for selecting whether to connect the output value of the neural network to the subtracting means. Network built-in control device. 6. The neural network built-in control device according to claim 2, further comprising selection means capable of selecting whether or not to activate the offset calculation means. 7. The neural network built-in control according to any one of claims 5 and 6, further comprising: a starting unit that starts the selecting unit so as to select a predetermined selection item at a predetermined timing. apparatus. 8. The limiter means according to claim 1, further comprising limiter means for limiting either one of an upper limit value and a lower limit value of an output value from the subtraction means. A controller with a built-in neural network, which is characterized by: 9. The method according to claim 2, further comprising a limiter means, wherein the limiter means sets one of an upper limit value and a lower limit value of an output value from the offset computing means. A control device with a built-in neural network characterized by being limited. 10. A mathematical expression representing a model when a neural network including an input layer, an intermediate layer, and an output layer and a controlled object are expressed by a mathematical model including parameters that are not exactly known. According to the control amount of the controlled object, a command value calculation means for giving an operation amount to the controlled object is provided so as to match the controlled target value, and the input layer is a deviation amount from the controlled target value of the controlled object. The first input value and the second input value which is the state quantity of the controlled object are accepted, and the intermediate layer and the output layer deviate from the control target quantity of the controlled object and the controlled variable by the model based on the accepted values. A controller with a built-in neural network, wherein a modified value of the parameter is output from the output layer so as to reduce 11. A deviation amount from a control target value when a control target is expressed by a model including an input layer, an intermediate layer, and an output layer and expressed by a mathematical expression including a parameter that is not exactly known. There is at least one or more first input values and at least one or more second quantities that are state quantities possessed by the controlled object.
The input layer accepts an input value of, and a neural network that outputs the correction amount of the parameter of the model to the output layer by an operation performed by the intermediate layer and the output layer based on the received value is provided. A new neural network having an output layer and subtraction means, wherein an input layer of the new neural network corresponds to the second input value and the first input value. A predetermined specific value is received, and 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 received value, and the subtraction unit outputs the output values of the two neural networks. Is calculated, and the calculation result is output as the correction amount of the parameter. 12. A neural network comprising an input layer, an intermediate layer and an output layer is used, and at least one or more first input value and control which are deviation amounts from a control target value of a controlled object are used. In a control method by a neural network in which at least one second input value, which is a state quantity of an object, is given to an input layer and an operation amount given to a controlled object is outputted from an output layer, first, an input layer of the neural network To the second input value, and a predetermined specific value corresponding to the first input value, the intermediate layer based on the received value,
By the operation performed by the output layer, the output value is output from the output layer, the output value is stored, and then the difference between the output value of the neural network and the stored output value is calculated, A control method using a neural network, wherein the control value is output as the operation amount.