JPH09114503A - Controller - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform accurate control in consideration of a temporal factor present in a loop for feedback control. SOLUTION: According to the response time pΔt from when a manipulated variable is inputted to a controlled system 1 until when the influence of the manipulated variable appears in a controlled variable, and observation data which is stored in an observation data storage part and consists of a controlled variable measured in past one-unit time and the manipulated variable inputted to the controlled system 1, a learning part 4 regulates the parameter W of a regulation part 2 by learning and outputs it to the regulation part 2.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a controller for controlling the dissolved oxygen concentration in an aeration tank of a sewage treatment facility to a constant value.

[0002]

2. Description of the Related Art It is desirable that a control device for controlling a system of this type adapts to a gradual change in the characteristics of a controlled object and can cope with a sudden change in the state of the controlled object that affects the controlled variable. However, the dissolved oxygen concentration control in the aeration tank of the sewage treatment facility uses feedback control that determines the aeration air volume from the error between the observed dissolved oxygen concentration in the sewage and the target dissolved oxygen concentration. If the sewage inflow rate changes abruptly, feedback control will not be able to perform accurate control, and a monitor must manually adjust the aeration air volume. In addition, not only abrupt changes in dissolved oxygen concentration due to changes in sewage inflow, but also characteristics of the sewage treatment process change significantly depending on the season, so the current situation is that the feedback control gain constant must be adjusted.

Therefore, the adjustment unit is adjusted by learning using the observation data to adapt to the gradual characteristic change of the control target, and the control unit repeats the prediction of the control amount of the control target and the correction of the operation amount. By having a loop, it is possible to deal with a sudden change in the state of the controlled object that affects the controlled variable.

[0004]

However, even if such measures are taken, accurate control cannot be performed in an actual system because various time factors exist in the feedback control loop. . That is, for example, there is actually a time difference between the time when the operation amount is given to the controlled object and the time when the influence of the operation amount appears in the control amount. Further, there is a temporal difference between the time when the adjustment amount is determined by the adjustment unit and the time when the operation amount is applied to the controlled object from the operation unit based on the adjustment amount. Furthermore, there is a time delay before the measured value of the controlled variable measured by the controlled object is transmitted to the adjustment unit.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a control device capable of performing more accurate control in consideration of a time factor existing in a feedback control loop. Has an aim.

[0006]

In order to achieve the above object, the invention according to claim 1 is a control device for controlling the controlled variable of a controlled object by changing the manipulated variable input to the controlled object. An adjusting unit that calculates an adjustment signal based on the control amount and the target control amount measured by the control target, and an operation unit that outputs the operation amount to the control target based on the adjustment signal calculated by the adjusting unit, and the past. An observation data storage unit that stores observation data having a control amount and an operation amount that are observed in, and control from input of the operation amount from the operation unit to the control target until the influence of the operation amount appears in the control amount. A learning unit that learns the adjustment unit based on the response time of the target and the observation data stored in the observation data storage unit.

According to a second aspect of the present invention, in a control device for controlling a control amount of a control target by changing an operation amount input to the control target, the control amount and the target control amount measured by the control target are set. Based on the adjustment signal calculated by the adjustment unit, based on the adjustment signal calculated by the adjustment unit, the operation unit for outputting the operation amount to the control target, the observation data having the control amount and the operation amount observed in the past Based on the operation amount output from the operation unit and the control amount measured by the control target, the observation data storage unit that stores the Based on the response time calculation unit that calculates the response time of the controlled object until the influence appears in the control amount, the response time calculated by the response time calculation unit, and the observation data stored in the observation data storage unit. Te comprises a learning unit for learning the adjustment portion.

According to a third aspect of the present invention, in the control device according to the first or second aspect, the adjusting section calculates the virtual operation amount based on the control amount and the target control amount measured by the controlled object. A virtual operation amount calculation unit, a control amount prediction unit that predicts a control amount after the response time based on the control amount and the virtual operation amount measured in the control target, and a response time predicted by the control amount prediction unit An operation amount correction unit that corrects the virtual operation amount based on the subsequent predicted control amount and the target control amount,
The control amount predicting unit predicts the control amount and the operation amount correcting unit corrects the virtual operation amount at least once, and inputs the virtual operation amount obtained as an adjustment signal to the operation unit.

According to a fourth aspect of the present invention, in the control device according to the third aspect, the control amount predicting unit is composed of a neural circuit model, and the calculation of the manipulated variable correcting unit is performed by an error back propagation calculation of the neural circuit model. Realized, the learning unit creates learning data based on the response time and the observation data stored in the observation data storage unit, and learns the neural circuit model that constitutes the control amount prediction unit.

According to a fifth aspect of the present invention, in the control device according to the fourth aspect, the learning unit creates the learning data based on the observation data stored in the observation data storage unit at the same time. The observation data having the controlled variable and the manipulated variable of is used as an input signal, and the controlled variable after the response time is used as a teacher signal for learning.

According to a sixth aspect of the present invention, in a control device for controlling the control amount of the control target by changing the operation amount input to the control target, the control amount and the target control amount measured by the control target are set. Based on the adjustment signal calculated by the adjustment unit, based on the adjustment signal calculated by the adjustment unit, the operation unit for outputting the operation amount to the control target, the observation data having the control amount and the operation amount observed in the past An observation data storage unit that stores the response time of the control target from the time when the operation amount is input from the operation unit to the control target until the influence of the operation amount appears on the control amount, and the adjustment calculated by the adjustment unit. The learning unit learns the adjustment unit based on the operation delay time until the signal is realized as an operation amount through the operation unit and the observation data stored in the observation data storage unit.

According to a seventh aspect of the present invention, in a control device for controlling the controlled variable of the controlled object by changing the manipulated variable input to the controlled object, the controlled variable and the target controlled variable measured by the controlled object are used. Based on the adjustment signal calculated by the adjustment unit, based on the adjustment signal calculated by the adjustment unit, the operation unit for outputting the operation amount to the control target, the observation data having the control amount and the operation amount observed in the past Based on the observation data storage unit that stores the adjustment signal calculated by the adjustment unit and the operation amount output by the operation unit, the adjustment signal calculated by the adjustment unit is realized as the operation amount through the operation unit. The operation delay time measuring unit that measures the operation delay time until the operation, the operation sending time measured by the operation delay time measuring unit, and the operation amount after the operation amount is input from the operation unit to the control target. According influence based on the stored observation data in the observation data memory unit and the response time of the control object to appear in the controlled variable comprises a learning unit for learning the adjustment portion.

According to an eighth aspect of the present invention, in the control device according to the sixth or seventh aspect, the adjusting section calculates the virtual operation amount based on the control amount measured by the control target and the target control amount. A virtual operation amount calculation unit, a control amount prediction unit that predicts a control amount after the response time based on the control amount and the virtual operation amount measured in the control target, and a response time predicted by the control amount prediction unit An operation amount correction unit that corrects the virtual operation amount based on the subsequent predicted control amount and the target control amount,
The control amount predicting unit predicts the control amount and the operation amount correcting unit corrects the virtual operation amount at least once, and inputs the virtual operation amount obtained as an adjustment signal to the operation unit.

According to a ninth aspect of the present invention, in the control device according to the eighth aspect, the control amount predicting unit is composed of a neural circuit model, and the calculation of the manipulated variable correcting unit is performed by an error back propagation calculation of the neural circuit model. Realized, the learning unit creates learning data based on the response time, the operation delay time, and the observation data stored in the observation data storage unit, and forms a neural circuit model that constitutes the control amount prediction unit. Learn.

According to a tenth aspect of the present invention, in the control device according to the ninth aspect, when the learning unit creates learning data based on the observation data stored in the observation data storage unit, the time of day As the input signal, the observation data having the control amount of and the operation amount after the operation delay time from the time,
Learning is performed by using the control amount after the response time as a teacher signal.
According to an eleventh aspect of the present invention, in a control device that controls a controlled variable of a controlled object by changing an operation amount input to the controlled object, adjustment is made based on the controlled variable and the target controlled variable measured by the controlled object. An adjustment unit that calculates a signal, an operation unit that outputs an operation amount to the controlled object based on the adjustment signal calculated by the adjustment unit, and an observation data that has a control amount and an operation amount that have been observed in the past are stored. The observation data storage unit, the response time of the control target from the input of the operation amount from the operation unit to the control target until the influence of the operation amount appears in the control amount and the control amount measured by the control target are The learning unit learns the adjustment unit based on the transmission time delay until the data is input to the adjustment unit via the transmission unit and the observation data stored in the observation data storage unit.
According to a twelfth aspect of the present invention, in a control device for controlling the controlled variable of the controlled object by changing the manipulated variable input to the controlled object, adjustment is made based on the controlled variable and the target controlled variable measured by the controlled object. An adjustment unit that calculates a signal, an operation unit that outputs an operation amount to the controlled object based on the adjustment signal calculated by the adjustment unit, and an observation data that has a control amount and an operation amount that have been observed in the past are stored. An observation data storage unit, a transmission dead time measuring unit that measures a transmission dead time by the transmission unit based on a control amount measured by the control target and a control amount input to the adjustment unit through the transmission unit, and The transmission dead time measured by the transmission dead time measuring unit and the response time of the control object from the time when the operation amount is input from the operation unit to the control target until the influence of the operation amount appears in the control amount, and Based on the stored observation data in the measurement data storage unit comprises a learning unit for learning the adjustment portion.

According to a thirteenth aspect of the present invention, in the control device according to the eleventh aspect or the twelfth aspect, the adjusting unit is based on the control amount and the target control amount measured by the control target,
A virtual operation amount calculation unit that calculates a virtual operation amount, a control amount prediction unit that predicts the control amount after the response time based on the control amount and the virtual operation amount measured in the control target, and the control amount prediction unit Based on the predicted control amount and the target control amount after the response time predicted by the operation amount correction unit that corrects the virtual operation amount, the control amount prediction by the control amount prediction unit, and the virtual operation amount by the operation amount correction unit. Of the virtual operation amount obtained by performing the correction of 1. at least once, and inputting the virtual operation amount to the operation unit as an adjustment signal.

According to a fourteenth aspect of the present invention, in the control device according to the thirteenth aspect, the control amount predicting unit is composed of a neural circuit model, and the calculation of the manipulated variable correcting unit is performed by an error back propagation calculation of the neural circuit model. Realized, the learning unit
Learning data is created based on the response time, the transmission dead time, and the observation data stored in the observation data storage unit, and the neural circuit model that constitutes the control amount prediction unit is learned.

According to a fifteenth aspect of the present invention, in the control device according to the fourteenth aspect, when the learning section creates the learning data based on the observation data stored in the observation data storage section, the transmission is performed. Learning is performed by using the observation data having the control amount before the dead time and the operation amount at the time as the input signal and the control amount after the response time as the teacher signal.

According to a sixteenth aspect of the present invention, in a control device for controlling the controlled variable of the controlled object by changing the manipulated variable input to the controlled object, the controlled variable and the target controlled variable measured by the controlled object are controlled. Based on the adjustment signal calculated by the adjustment unit, based on the adjustment signal calculated by the adjustment unit, the operation unit for outputting the operation amount to the control target, the observation data having the control amount and the operation amount observed in the past And an observation data storage unit that stores the response time of the control target from the input of the operation amount from the operation unit to the control target until the influence of the operation amount appears on the control amount and the control measured by the control target. Transmission delay time until the amount is input to the adjustment unit through the transmission unit, the operation delay time until the adjustment signal calculated by the adjustment unit is realized as the operation amount through the operation unit, and the observation data Based on the stored observation data in the data storage unit comprises a learning unit for learning the adjustment portion.

According to a seventeenth aspect of the present invention, in a control device for controlling the controlled variable of the controlled object by changing the manipulated variable input to the controlled object, the controlled variable and the target controlled variable measured by the controlled object are set. Based on the adjustment signal calculated by the adjustment unit, based on the adjustment signal calculated by the adjustment unit, the operation unit for outputting the operation amount to the control target, the observation data having the control amount and the operation amount observed in the past Based on the observation data storage unit that stores the adjustment signal calculated by the adjustment unit and the operation amount output by the operation unit, the adjustment signal calculated by the adjustment unit is realized as the operation amount through the operation unit. Based on the control delay measured by the control target measured by the control target and the control delay measured by the control unit, the transmission delay by the transmission unit. A transmission dead time measuring unit that measures time, an operation delay time measured by the operation delay time measuring unit, a transmission dead time measured by the transmission dead time measuring unit, and an operation amount from the operation unit to the controlled object. A learning unit that learns the adjustment unit based on the response time of the controlled object from the input to the influence of the operation amount on the control amount and the observation data stored in the observation data storage unit. .

According to an eighteenth aspect of the present invention, in the control device according to the sixteenth aspect or the seventeenth aspect, the adjusting unit is based on the control amount and the target control amount measured by the control target,
A virtual operation amount calculation unit that calculates a virtual operation amount, a control amount prediction unit that predicts the control amount after the response time based on the control amount and the virtual operation amount measured in the control target, and the control amount prediction unit Based on the predicted control amount and the target control amount after the response time predicted by the operation amount correction unit that corrects the virtual operation amount, the control amount prediction by the control amount prediction unit, and the virtual operation amount by the operation amount correction unit. Of the virtual operation amount obtained by performing the correction of 1. at least once, and inputting the virtual operation amount to the operation unit as an adjustment signal.

According to a nineteenth aspect of the present invention, in the control device according to the eighteenth aspect, the control amount predicting unit is composed of a neural circuit model, and the calculation of the operation amount correcting unit is performed by an error back propagation calculation of the neural circuit model. Realized, the learning unit
Based on the response time, the operation delay time, the transmission dead time, and the observation data stored in the observation data storage unit, learning data is created to learn the neural circuit model that constitutes the control amount prediction unit. To do.

According to a twentieth aspect of the present invention, in the control device according to the nineteenth aspect, the learning section transmits the learning data based on the observation data stored in the observation data storage section. Learning is performed by using the observation data having the control amount before the dead time and the operation amount after the operation delay time as an input signal and the control amount after the response time as a teacher signal.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram for explaining the principle of the present invention and shows a signal flow at time nΔt.

In this control device, as shown by the thick line in FIG. 1, the controlled object 1, the adjusting section 2 and the operating section 3 constitute a feedback control loop. The learning unit 4 and the observation data storage unit 5 adjust the parameters of the adjustment unit 2 by learning.

A manipulated variable is input to the controlled object 1 every unit time Δt, and the controlled variable is measured. The controlled variable X _n measured by the controlled object 1 is input to the adjustment unit 2. The adjusting unit 2 calculates the adjusting signal C _n based on the control amount Χ _n measured for the controlled object 1 and the preset target control amount Χ ^d . Operation part 3
Then, the manipulated variable U _n is output based on the adjustment signal C _n calculated by the adjustment unit 2. The operation amount U _n output by the operation unit 3 is input to the controlled object 1 and is controlled by changing the operation amount U _n . The subscript n of the symbol indicates the time, for example, X
_n means the control amount measured from the time (n-1) [Delta] t to n [Delta] t, but in this description, it is simply referred to as the control amount at the time n [Delta] t. Hereinafter, the subscript will be used with the same meaning.

The learning unit 4 measures the response time pΔt from the input of the operation amount to the controlled object 1 to the influence of the operation amount on the control amount, and the last 1 unit time stored in the observation data storage unit 5. The controlled _variable (X _nl , X _{n-l + 1} , _...
.., X _n-1 ) and the manipulated variable (U
_nl , U _{n-l + 1} , ..., U _n-1 ) (l is a positive integer) based on the observation data, the parameter W of the adjusting unit 2
Is adjusted by learning and output to the adjusting unit 2. As a result, in the present invention, the adjusting unit 2 can adapt to a gradual change in the characteristic parameter of the controlled object 1 that affects the controlled variable. The response time pΔt may be set by previously measuring or calculating the time from the input of the operation amount to the controlled object 1 until the influence of the operation amount appears in the control amount, as will be described later. You may make it set dynamically according to the change of the controlled object 1.

FIG. 2 is a diagram showing an example of the configuration of the adjusting section 2.

As shown in FIG. 2, the virtual manipulated variable calculator 6 controls the controlled variable X _n and the target controlled variable X ^{d for the} controlled object 1.
The virtual operation amount U (^) _n is calculated based on Based on the virtual operation amount U (^) _n and the control amount Χ _n measured by the controlled object 1, the control amount predicting unit 7 outputs pΔt time after the virtual operation amount U (^) _n is output as the adjustment signal. Predicted controlled variable X
(~) Calculate _{n + p} . pΔt is the response time of the controlled object 1. Manipulated variable correcting section 8, based on the control amount predicting control amount X of pΔt time after that is predicted by the prediction unit 7 (~) n + _p and the target control amount X ^d, the predicted controlled variable X (~) n + _p The virtual manipulated variable U (^) _n is corrected so as to approach the target controlled variable X ^d .
The first switch 9 closes downward when the controlled variable is measured by the controlled object 1, inputs the virtual manipulated variable U (^) _n calculated by the virtual manipulated variable calculating unit 6 into the controlled variable predicting unit 7, and thereafter. , The first switch 9 is closed to the upper side until the next measurement, and the virtual operation amount U (^) _n corrected by the operation amount correction unit 8 is set to the control amount prediction unit 7
For a predetermined number of times or the predicted controlled variable X (~)
until the difference _{n + p} and the target control amount chi ^d is below the predetermined value,
Prediction of the control amount by the control amount prediction unit 7 and correction of the operation amount by the operation amount correction unit 8 are repeated. The control amount prediction unit 7 predicts the control amount and the operation amount correction unit 8 corrects the operation amount up to a predetermined number of times, or the difference between the predicted control amount X (-) _{n + p} and the target control amount ^Xd is a predetermined value. When it is repeated until the following, the second switch 10 is closed and the virtual operation amount U (^) _n
Is output to the operation unit 3 as the adjustment signal C _n . Accordingly, in the present invention, the controlled object 1 that affects the controlled variable X _n
Even if the state of is suddenly changed, the manipulated variable U _n can be corrected accurately and the controlled variable X _n can be controlled to be close to the target controlled variable X ^d .

FIG. 3 is a diagram showing an example of the configuration of the control amount predicting section 7. Here, the control amount prediction unit 7 is composed of a neural circuit model. FIG. 3 shows a virtual operation amount U at time nΔt
(^) _N = (U (^) 1 _n , U (^) 2 _n , ..., U
(^) G _n ) ^T (g represents the number of dimensions of the manipulated variable) and the controlled variable Χ _n = (Χ _{1 n} , Χ 2 _n , ..., Χ h _n ) ^T (h represents the number of dimensions of the controlled variable) Is input, the predicted control amount X (-) _{n + p} = (X (-) 1n _{+ p} , X (-) after _{p [} Delta] t time has elapsed.
2 _{n + p} , ..., Χ (∼) h _{n + p} ) ^T is shown, showing the signal flow of the prediction calculation when the neural circuit model 11 is controlled. The output signal is obtained from the input signal by the forward calculation of the neural circuit model 11. The forward calculation of the neural network model 11 will be described later.

When the control amount predicting unit 7 is composed of the neural circuit model 11, the operation amount correcting unit 8 obtains the correction amount of the virtual operation amount U (^) _n by the error back propagation calculation of the neural circuit model 11. , Correct the virtual operation amount. The error back propagation calculation of the neural network model 11 will be described later together with the forward calculation.

Here, when the control amount predicting unit 7 is composed of the neural circuit model 11, the learning unit 4 is structured as follows.

That is, the learning section 4 has a response time pΔt until the influence of the manipulated variable on the controlled object 1 appears in the controlled variable.
And learning data are created based on the observation data including the past control amount and the manipulated variable stored in the observation data storage unit 5, and the neural circuit model 11 that constitutes the control amount prediction unit 7 is learned. When the neural network model 11 inputs the control amount and the virtual operation amount at the current time as shown in FIG. 3, the response time pΔ
In the case of the configuration that outputs the controlled variable after t, the input signal (U _s ,
X _s), a teacher signal _{(X s + p) (s} = n-l, n-l +
, ..., Np-1) learning data.

Next, the forward calculation, the back propagation calculation, and the learning calculation of the neural network model will be described.

The forward calculation is a calculation for obtaining an output signal when an input signal is given to the neural network model. Here, the forward calculation will be described based on the three-layer neural circuit model shown in FIG. The circles in the figure are nerve cell models called units, and the line segments that connect the units are the connections that transmit signals between the units. The units are arranged in three layers, from the left, an input layer, an intermediate layer, and an output layer. Couplings are provided between the units of the input layer and the units of the intermediate layer, and between the units of the intermediate layer and the units of the output layer. The coupling between the units has a directivity in signal transmission, and in the forward calculation, the signal is transmitted only in one direction from the input layer unit to the intermediate layer unit and from the intermediate layer unit to the output layer unit.

Input signals from the outside are input to the input layer units one by one. In FIG. 3, the virtual operation amount U (^) _n = (U (^) 1 _n , U (^) 2 _n ,
, U (^) g _n ) ^T , controlled variable X _n = (X1 _n , X
2 _n , ..., Xh _n ) ^T are input to each unit, one signal at a time. If xi _i is the input value of the input layer i-th unit,

(Equation 1) The input / output function of the input layer unit is the identity function. The output value yi _i (i = 1, 2, ..., Νi) of the input layer unit is expressed as follows.

[0037]

(Equation 2) Νi is the number of input layer units (Νi = g + h).

Next, the input value of the intermediate layer unit is a value obtained by subtracting the threshold value of the intermediate layer unit from the sum of the output value of the input layer unit weighted by the coupling weight value between the input layer and the intermediate layer.

(Equation 3) xh _j is the input value of the middle layer jth unit, w1 _{j, i} is the combined load value between the input layer ith unit and the middle layer jth unit, thh _j is the threshold value of the middle layer jth unit, Nh Represents the number of units in the intermediate layer.

The input / output function of the middle layer unit is a sigmoid type function.

(Equation 4) Then, the output value yh _j of the j-th unit in the middle layer uses the input value xh _j ,

(Equation 5) Can be expressed as

The input value of the output layer unit is a value obtained by subtracting the threshold value of the output layer unit from the sum of the output value of the intermediate layer unit weighted by the coupling weight value between the intermediate layer and the output layer.

[0041]

(Equation 6) xo _k input value of the k-th unit output layer, w2 k, _j intermediate layer j-th unit - bond load value between the output layer the k-th unit, tho _k is a threshold of the k-th unit output layer, No Represents the number of units in the output layer. The input / output function of the output layer unit is an identity function.

[0042]

(Equation 7) The output signal of this output layer unit is output to the outside. In FIG. 3, the predicted value X (-) of the control amount at time (n + p) Δt
_{n + p} = (X1 (~) _{n + p} , X (~) _{2n + p} ..., X (~)
h _{n + p} ) ^T.

[0043]

(Equation 8) The above is the calculation formula for the forward calculation of the neural circuit model 11.

Next, the error back propagation calculation of the neural network model will be described.

When the control amount X (-) _{n + p} predicted by the control amount predicting unit 7 and the target control amount ^Xd are input to the operation amount correcting unit 8, the operation amount correcting unit 8 calculates the virtual operation amount. U (^)
The correction amount of _n is calculated, and the virtual operation amount U (^) _n is corrected. When the control amount prediction unit 7 is composed of the neural circuit model 11, the correction amount of the virtual operation amount U (^) _n can be calculated by the error back propagation calculation of the neural circuit model 11.

An error function is defined from the output signal of the neural network model and the desired output signal. In the case of FIG. 3, the predicted control amount X (-) _{n + p} calculated by the forward calculation of the neural circuit model 11 of the control amount prediction unit 7 and the target control amount X
^{(~) D = (Xl d} , X2 d, ..., Xh d) define the error between ^T error function E as follows.

[0047]

(Equation 9) Next, the correction amount of the unit output value that reduces this error function is sequentially obtained for each layer from the output layer to the input layer, and finally the correction amount of the input signal is obtained. First, the correction amount of the output value of the output layer unit is calculated by using the positive constant η.

(Equation 10) Is calculated.

Next, the correction amount of the output value of the intermediate layer unit is expressed as follows using the correction amount of the output layer unit.

[0049]

[Equation 11] Similarly, the correction amount of the output value of the input layer unit, using the correction amount of the middle layer unit,

(Equation 12) It is expressed as f ′ (x) represents the derivative of the sigmoid function f (x), and specifically,

(Equation 13) Can be expressed as Since the input / output function of the input layer unit is the identity function, the correction amount of the output value becomes the correction amount of the input signal as it is.

When the neural circuit model is as shown in FIG. 3, the correction amount ΔU _n of the virtual operation amount U (^) _n which is the input signal is

[Equation 14] Is expressed as This correction amount is added to the virtual operation amount U (^) _n .

The above is the error backpropagation calculation and the correction of the virtual operation amount realized by this calculation.

Next, the learning calculation of the neural circuit model will be described with reference to FIG.

In FIG. 4, 5 is an observation data storage unit, 1
1 is a neural network model, 12 is a subtractor, 13 is an error backpropagation learning device, 14 is an arrow indicating that the output of the input layer unit of the neural circuit model 11 is input to the error backpropagation learning device 13, and 15 is a neural network. An arrow indicating that the output of the intermediate layer unit of the circuit model 11 is input to the error backpropagation learning device 13,
Reference numeral 16 denotes an arrow indicating that the coupling weight value between the intermediate layer and the output layer of the neural network model 11 is input to the error backpropagation learning device 13, and 17 indicates between the input layer and the intermediate layer calculated by the error backpropagation learning device 13. An arrow indicating that the combined weight value and the correction amount of the threshold value of the intermediate layer unit are output to each combined load and intermediate layer unit and the combined load value and the threshold value are corrected, 18 is calculated by the error backpropagation learning device 13. An arrow indicating that the combined weight value between the intermediate layer and the output layer and the correction amount of the threshold value of the output layer unit are output to each combined load and output layer unit, and each combined weight value and the threshold value are corrected is shown.

Past time sΔt (s = n−1, n−1 +
1, ..., Np-1), the manipulated variable U at time sΔt
_s and the controlled variable X _s as input signals, the observation data storage unit 5
Is output to the neural circuit model 11. In the neural circuit model 11, forward calculation is performed using the manipulated variable U _s and the controlled variable X _s at time sΔt as input signals, and the output signal X (-) _{s + p} is output. In the subtractor 12, the difference X _{s + p} − between the control amount Χ _{s + p} after the response time output as the teacher signal from the observation data storage unit 5 and the output signal X (-) _{s + p} of the neural circuit model 11
Χ (∼) _{s + p} is calculated and input to the error backpropagation learning device 13. In addition to the difference, each output value of the input layer unit of the neural network model, each output value of the intermediate layer unit, and each coupling weight value between the intermediate layer and one output layer are input to the error back-propagation learning device 13, and the coupling weight value is input. , The threshold modification amount is calculated.

In the error backpropagation learning device 13, the teacher signal X
_The error function E is defined as follows from the difference between _{s + p} and the output signal X _{s + p} of the neural circuit model 11.

[0056]

(Equation 15) According to the error backpropagation learning method, the connection weight value of the neural network model 11 and the correction amount of the threshold value are calculated as follows.

[0057]

(Equation 16) Derutadaburyu2 k, _j is the intermediate layer j-th unit - correction amount of the coupling weight value w2 _{k, j} between the output layer the k-th unit, Derutatho _k is the correction amount of the threshold tho _k of the k-th unit output layer, .DELTA.W1
_{j, i} is the correction amount of the coupling load value w1 _{j, i} between the input layer i-th unit and the intermediate layer j-th unit, Δthh _j is the correction amount of the threshold thh _j of the intermediate layer j-th unit, and ε is the learning value. It is a constant. yo ^s _k time output layer output values of the k-th unit when data is input in Esuderutati, the input values of the output layer the k th unit when data xo ^s _k time Esuderutati is input,
yh ^s _j is the output value of the middle-layer j-th unit when the data at time sΔt is input, and xh ^s _k is the input value of the middle-layer j-th unit when the data at time sΔt is input, y
i ^s _i is the output value of the i-th unit in the middle layer when the data at time sΔt is input. Based on these corrections
The connection weight value and the threshold value are modified.

Next, another embodiment of the present invention will be described with reference to FIG.

In the control device shown in FIG. 5, a response time calculating section 19 is added to the control device shown in FIG. The response time calculation unit 19 calculates the response time of the controlled object 1. The response time calculation unit 19 is a time series U = (U _nm , U _{n-m + 1} , ..., U of the manipulated variables input to the controlled object 1 in the past m unit time.
_n-1 ) and the time series of the controlled variable X =
Based on (X _nm , X _{n-m + 1} , ..., X _n-1 ), the response time of the controlled object 1 until the influence of the manipulated variable input to the controlled object 1 appears in the controlled variable is calculated and learned. Output to the unit 4 (m is a positive integer).

Next, another embodiment of the present invention will be described with reference to FIG.

The control device shown in FIG. 6 takes into consideration that the operating section 2 has a time delay. As shown in the figure, the controlled variable X _n measured by the controlled object 1 is input to the adjustment unit 2. The adjusting unit 2 controls the controlled variable Χ measured by the controlled object 1.
_The adjustment signal C _n is calculated based on _n and the target control amount X ^d . The operation unit 3 outputs the operation amount U _n based on the adjustment signal C _n calculated by the adjustment unit 2. However, the operation unit 3 has an operation delay time, and the input adjustment signal delays the operation delay time and the operation amount is delayed. Is output as. The operation amount U _n output by the operation unit 3 is input to the controlled object 1 and is controlled by changing the operation amount U _n . The learning unit 4 realizes the response time pΔt from when the operation amount is input to the controlled object 1 until the influence of the operation amount appears in the control amount and the adjustment signal calculated by the adjustment unit 2 as the operation amount through the operation unit 3. Operation delay time qΔt until the operation is performed, and the control amount (X _nl , X) measured in the past 1 unit time stored in the observation data storage unit 5.
_{n-l + 1} , ..., X _n-1 ) and the manipulated _variables (U _nl , U _{n-l + 1} , ..., U _n-1 ) input to the controlled object 1 (l is a positive integer) Based on the observation data, the parameter W of the adjusting unit 2
Is adjusted by learning and output to the adjusting unit 2. The operation delay time qΔt is the same as the response time pΔt.
The time until the adjustment signal calculated in step 3 is realized as an operation amount through the operation unit 3 may be measured or calculated in advance, and may be set in accordance with a change in the adjustment unit 2 as described later. It may be set to.

FIG. 7 is a diagram showing an example of the configuration of the adjusting section 2 shown in FIG.

As shown in FIG. 7, the virtual manipulated variable calculator 6 controls the controlled variable X _n and the target controlled variable X ^d measured for the controlled object 1.
The virtual operation amount U (^) _n is calculated based on Based on the virtual operation amount U (^) _n and the control amount X _n measured by the controlled object 1, the control amount predicting unit 7 calculates the virtual operation amount U (^) _n as an adjustment signal after pΔt time. Predicted controlled variable Χ
(~) Calculate _{n + p} . pΔt is the response time of the controlled object 1. Manipulated variable correcting section 8, based on the control amount X _{n + p} and the target control amount X ^d after predicted pΔt time by the control amount prediction unit 7, the predicted controlled variable X (~) n + _p is the target control amount X ^The virtual manipulated variable U (^) _n is modified so as to approach ^d . The first switch 9 closes downward when the controlled variable is measured by the controlled object 1,
Virtual operation amount U (^) _n calculated by the virtual operation amount calculation unit 6
Is input to the control amount predicting unit 7, and then the first switch 9 is closed upward until the next measurement, and the virtual operation amount U (^) _n corrected by the operation amount correcting unit 8 is input to the control amount predicting unit 7 Input, and until the predetermined number of times, or until the difference between the predicted control amount X (-) _{n + p} and the target control amount ^Xd becomes a predetermined value or less, the control amount prediction unit 7 predicts the control amount and corrects the manipulated variable. The correction of the operation amount by the unit 8 is repeated. The control amount prediction unit 7 predicts the control amount and the operation amount correcting unit 8 corrects the operation amount up to a predetermined number of times, or the difference between the predicted control amount Χ (~) _{n + p} and the target control amount X ^d is a predetermined value. When it is repeated until the following, the second switch 10 is closed and the virtual operation amount U (^) _n is output to the operation unit 3 as the adjustment signal C _n .

FIG. 8 is a diagram showing an example of the configuration of the control amount predicting section 7 shown in FIG. In this case as well, the control amount prediction unit 7 is composed of a neural circuit model. In FIG. 8, time nΔt
Virtual operation amount U (^) _n = (U (^) _1n , U (^) 2 of
_n , ..., U (^) g _n ) ^T (g represents the number of dimensions of the manipulated variable) and the controlled variable Χ _n = (Χ _{1 n} , Χ 2 _n , ..., Χ h _n ) ^T
When (h represents the number of dimensions of the controlled variable) is input, the predicted controlled variable X (~) _{n + p} = (X (~) 1 after the response time pΔt is reached.
_{n + p} , Χ (-) _{2n + p} , ..., Χ (-) h _{n + p} ) ^T is output, showing the signal flow of the prediction calculation during control of the neural circuit model 11. The output signal is obtained from the input signal by the forward calculation of the neural circuit model 11 described above.

When the control amount predicting unit 7 is composed of the neural circuit model 11, the operation amount correcting unit 8 calculates the virtual operation amount U by the error backpropagation calculation of the neural circuit model 11 described above.
(^) The virtual operation amount is corrected by _obtaining the correction amount of _n .

Here, when the control amount predicting unit 7 is composed of the neural circuit model 11, the learning unit 4 is structured as follows.

That is, the learning unit 4 learns data based on the response time pΔt of the controlled object 1, the operation delay time qΔt, and the observation data that is stored in the observation data storage unit 5 and includes the past control amount and operation amount. Is created, and the neural circuit model 11 that constitutes the control amount prediction unit 7 is learned.

[0068] an operation amount U _s and the controlled variable chi _s operation amount input to the time sΔt to the controlled object 1 and the time control amount measured by the control target 1 to sΔt time sΔt observation data.
Therefore, when the operation unit 3 has a time delay, the input signal (U _s , X _s ) and the teacher signal (X _{s + p} ) (s = n-1, n)
When learning is performed using −l + 1, ..., Np−1) as learning data, the neural circuit model 11 causes the controlled variable X _s measured at time sΔt and the manipulated variable U _s input to the controlled object 1 at time sΔt. Then, the control amount X _{s + p} after pΔt time is output. When this neural circuit model 11 is controlled as the control amount predictor 7, the controller 2 causes the predictive control amount X (-) _{n + p}
There will be a demand value of the operating amount U _n for approaching the target control amount X ^d, since the actual value is input to the control object 1 is after the operation delay time Qderutati, can not be accurately controlled .

Therefore, in the present invention, the neural circuit model 11
In addition, the controlled variable X _n and the adjustment signal C _n measured in the controlled object 1
, The predicted controlled variable X (~) _{n + p} after pΔt time
In order to perform the learning of the neural circuit model 11 so as to output the control amount Xs = (X1 _s , X
2 _s , ..., Xh _s ) and the manipulated variable U _{s + q} = (U1 _{s + q} , U2 _{s + q} , which is input to the controlled object 1 after the operation delay time qΔt.
,, Ug _{s + q} ) as the input signal, and the control amount after the response time pΔt Χ _{s + p} = (Χ 1 _{s + p} , Χ 2 _{s + p} ,… Xh _{s + p} ) is used as the teacher signal to create learning data. (S = n-1, n-1 +
1, ..., Np-1). If learning can be performed with this learning data, the control signal C _n obtained by the adjusting unit 2 during control is the target control amount X _{n + p} after the response time pΔt when input to the control target 1 after the operation delay time qΔt. Since the operation amount approaches the amount ^Xd , accurate control can be performed.

Next, another embodiment of the present invention will be described with reference to FIG.

In the control device shown in FIG. 10, an operation delay time measuring section 20 is added to the control device shown in FIG.
The operation delay time measuring unit 20 has a time series C = (C _nm , C _{n-m + 1} , of the adjustment signal calculated by the adjusting unit 2 in the past m unit time,
, C _n-1 ) and the time series U = (U _nm , U _{n-m + 1} , ..., U _n-1 ) of the manipulated variables output from the operating unit 3, the adjustment signal The operation delay time qΔt from the input to the output to the output as the operation amount is calculated and output to the learning unit 4. m is a positive integer. By adding the operation delay time measuring unit 20, even if the operation delay time qΔt changes due to the characteristic change of the operation unit 3, the operation delay time measuring unit 20 calculates the changed operation delay time qΔt, and the learning unit 4 is output. In the learning unit 4, according to the operation delay time qΔt input from the operation delay time measuring unit 20,
Since learning data is created and learning is performed, correct control can be continued even if the operation delay time changes.

Next, another embodiment of the present invention will be described with reference to FIG.

The control device shown in FIG. 11 takes into consideration that there is a transmission section having a dead time for transmitting the control amount measured by the controlled object 1 to the adjustment section 2.

As shown in the figure, the control amount _n measured at the controlled object 1 at the time nΔt is input to the adjustment unit 2 through the transmission unit 21, but the transmission unit 21 has a transmission dead time and control is performed. control amount chi _n measured by the subject 1 is input to the modulating portion 2 delayed transmission dead time. The control amount input to the adjusting unit 2 at time nΔt is represented by X ′ _n . The adjustment unit 2 calculates the adjustment signal C _n based on the control amount X ′ _n and the target control amount X ^d . In the operation unit 3, the adjustment signal C _n calculated by the adjustment unit 2
The manipulated variable U _n is output based on The controlled object 1 is input operating amount U _n of the operation unit 3 and output, the operation amount U _n
Is controlled by changing.

The learning unit 4 receives the response time pΔt from the input of the manipulated variable to the controlled object 1 until the influence of the manipulated variable appears in the controlled variable, and the controlled variable X _n measured by the controlled object 1 as the transmission unit 21. Transmission delay time rΔt until it is input to the adjustment unit 2 through the control data and the control amount (Χ _nl , X _{n-l + 1} ,
, X _n-1 ) and the manipulated variables (U _nl , U _{n-l + 1} , ..., U _n-1 ) input to the controlled object 1, based on the observation data, the parameter W of the adjusting unit 2 is set. It is adjusted by learning and output to the adjusting unit 2. The transmission dead time is the same as the operation delay time qΔt and the response time pΔt.
The control amount X _n measured by the controller 2 is transmitted through the transmitter 21.
The time until it is input may be measured or calculated and set in advance, or may be dynamically set according to the variation of the transmission unit 21 as described later.

FIG. 12 is a diagram showing an example of the configuration of the adjusting section 2 shown in FIG.

[0077] adjusting section 2 shown in FIG. 12 relative to adjustment portion 2 differs from the control amount control amount _X'n inputted to modulating section 2 is measured by the control target 1 is X _n shown in FIG. 7 That is, the transmission dead time rΔt is delayed. The virtual operation amount calculation unit 6 calculates the virtual operation amount U (^) _n based on the control amount _X'n and the target control amount ^Xd input through the transmission unit 21.
The control amount predicting unit 7 calculates the virtual operation amount U (^) _n and the control amount Χ ' _n.
Based on, the predicted control amount X (-) _{n + p} after pΔt time when the virtual operation amount U (^) _n is output as the adjustment signal is calculated. pΔt is the response time of the controlled object 1. Manipulated variable correcting section 8, based on the control amount X _{n + p} and the target control amount X ^d after predicted pΔt time by the control amount prediction unit 7, the predicted controlled variable X (~) n + _p is the target control amount X ^The virtual manipulated variable U (^) _n is modified so as to approach ^d . The first switch 9 closes downward when the controlled variable is measured by the controlled object 1, inputs the virtual manipulated variable U (^) _n calculated by the virtual manipulated variable calculating unit 6 into the controlled variable predicting unit 7, and thereafter. , The first switch 9 is closed to the upper side until the next measurement, and the virtual operation amount U (^) _n corrected by the operation amount correction unit 8 is input to the control amount prediction unit 7, and the predetermined number of times or the prediction control is performed. Quantity Χ (~) _{n + p} and target control quantity Χ ^d
The control amount prediction by the control amount prediction unit 7 and the correction of the operation amount by the operation amount correction unit 8 are repeated until the difference is less than or equal to a predetermined value. The control amount prediction unit 7 predicts the control amount and the operation amount correcting unit 8 corrects the operation amount up to a predetermined number of times, or the difference between the predicted control amount Χ (~) _{n + p} and the target control amount X ^d is a predetermined value. When repeated until the second switch 10
Is closed, and the virtual operation amount U (^) _n is output to the operation unit 3 as the adjustment signal C _n .

FIG. 13 is a diagram showing an example of the configuration of the control amount predicting section 7 shown in FIG. Again, the control amount predicting unit 7
Is composed of a neural circuit model. In FIG. 13, the virtual manipulated variable U (^) _n = (U (^) 1 _n , U at time nΔt
_{(^) 2 n, ...,} U (^) g n) T (g represents the number of dimensions of the operation amount) and the controlled variable _{Χ'n (X1 'n, X2} ' n, ...,
When Χh ′ _n ) ^T (h represents the number of dimensions of the controlled variable) is input, the predicted controlled variable X (∼) _{n + p} = after the response time pΔt
(X1 _{n + p} , X2 _{n + p} , ..., Xh _{n + p} ) ^T shows the signal flow of the prediction calculation during control of the neural circuit model 11. The output signal is obtained from the input signal by the forward calculation of the neural circuit model 11 described above.

When the control amount predicting unit 7 is composed of the neural circuit model 11, the operation amount correcting unit 8 calculates the virtual operation amount U by the error back propagation calculation of the neural circuit model 11 described above.
(^) _Obtain the correction amount of _n and correct the virtual operation amount.

Here, when the control amount predicting unit 7 is composed of the neural circuit model 11, the learning unit 4 is structured as follows.

That is, the learning unit 4 learns data based on the response time pΔt of the controlled object 1, the transmission dead time rΔt, and the observation data including the past control amount and operation amount stored in the observation data storage unit 5. Is created, and the neural circuit model 11 that constitutes the control amount prediction unit 7 is learned. Transmission unit 21
If there is a dead time in transmission, the input signal (U _s ,
X _s), a teacher signal _{(X s + p) (s} = n-l, n-l +
1, ..., Np-1) is used as learning data, the neural circuit model 11 controls the controlled variable X _s measured at time sΔt and the manipulated variable U input to the controlled object 1 at time sΔt.
_{From s,} the control amount Χ _{s + p} after the response time pΔt is output. When this neural circuit model 11 is controlled by the control amount predicting unit 7, the control amount predicting unit 7 transmits the transmission dead time rΔ.
Since the control amount measured before t is used as the current control amount and the predicted control amount X (~) _{n + p} after _{p [} Delta] t time is output, accurate control cannot be performed.

[0082] In the present invention, when the control amount X _'n and the adjustment signal C _n inputted to modulating section 2 at time nΔt input to the neural network model 11, the prediction control amount after pΔt time Χ
In order to perform learning of the neural circuit model 11 so that (~) _{n + p} is output, as shown in FIG. 14, the controlled _variable X _sr = (X1 _sr measured by the controlled object 1 before the dead time of the transmission. ，
X2 _sr , ..., Xh _sr ) and the manipulated variable U _s = (U1 _s , U)
2 _s , ..., Ug _s ) as an input signal, and a control amount Χ _{s + p} = (Χ 1 _{s + p} , Χ _{2 + p} , ..., Xh _{s + p} ) after the response time pΔt is used as a teacher signal (s = n -L + r, n-l + r + 1, ...,
n-p-1). If learning can be performed with this learning data, the adjustment signal C _n obtained by the adjustment unit 2 at the time of control is predicted from the control amount X ′ _n and the adjustment signal C _n measured by the controlled object 1 before the transmission dead time rΔt. Control amount Χ after response time pΔt
Since _{n + p} is adjustment signal approaches the target control amount chi ^d, it can be accurately controlled.

Next, another embodiment of the present invention will be described with reference to FIG.

In the control device shown in FIG. 15, a transmission dead time measuring section 22 is added to the control device shown in FIG. The transmission dead time measuring unit 22 determines the time series X = (X _nm , X of the controlled variable measured by the controlled object 1 in the past m unit time.
_{n-m + 1} , ..., Χ _n-1 ) and the time series of control quantities input to the adjusting unit 2 X '= (X _nm , Χ' _{n-m + 1} , ..., Χ ' _n-1 ).
Based on, the transmission dead time rΔt until the controlled variable measured by the controlled object 1 is input to the adjusting unit 2 is calculated and output to the learning unit 4. m is a positive integer. By adding the transmission dead time measuring unit 22, even if the transmission dead time changes due to the characteristic change of the transmission unit 21, the transmission dead time measuring unit 22 calculates the changed transmission dead time rΔt, and the learning unit 4 Is output to. The learning unit 4 creates learning data according to the transmission dead time rΔt input from the transmission dead time measuring unit 22 and performs learning, so that correct control can be continued even when the transmission dead time changes.

Next, another embodiment of the present invention will be described with reference to FIG.

In the control device shown in FIG. 16, the operation unit 3 has a dead time until the adjustment signal calculated by the adjustment unit 2 is input to the control target 1 as an operation amount, and the transmission unit 21 controls the control target 1 This is because it takes a dead time to transmit the control amount measured in step 2 to the adjustment part 2. As shown in the figure, the controlled variable X measured by the controlled object 1 at time nΔt
_n is input to the adjusting unit 2 through the transmitting unit 21, and the transmitting unit 21 has a transmission dead time rΔt, and the control amount X _n measured by the controlled object 1 is input to the adjusting unit 2 with a delay of the transmission dead time rΔt. To be done. The control amount input to the adjusting unit 2 at time nΔt is represented by X ′ _n . The adjustment unit 2 calculates the adjustment signal C _n based on the control amount X ′ _n and the target control amount X ^d . Operation part 3
Then, the operation amount U _n is output based on the adjustment signal C _n calculated by the adjustment unit 2, but the operation unit 3 has a time delay, and the input adjustment signal C _n is delayed as the operation delay time qΔt to be an operation amount. Is output. The operation amount U _n output by the operation unit 3 is input to the controlled object 1 and is controlled by changing the operation amount U _n .

In the learning unit 4, the response time pΔt from the input of the operation amount to the controlled object 1 until the influence of the operation amount appears in the control amount and the adjustment signal calculated by the adjusting unit 2 are transmitted through the operating unit 3. The operation delay time qΔt until it is realized as an operation amount, the transmission dead time rΔt until the control amount measured by the controlled object 1 is input to the adjustment unit 2 through the transmission unit 21, and stored in the observation data storage unit 5. Control amount (X _nl , X _{n-l + 1} , ..., X
_n-1 ) and the manipulated _variables (U _nl , U input to the controlled object ₁ )
_{n-l + 1} , ..., U _n-1 )
The parameter W of the adjusting unit 2 is adjusted by learning and output to the adjusting unit 2.

FIG. 17 is a diagram showing an example of the configuration of the adjusting section 2 shown in FIG.

As shown in FIG. 17, the virtual operation amount calculator 6
Calculates a virtual manipulated variable U (^) _n based on the controlled variable _X'n and the target controlled variable ^Xd input through the transmission unit 21. Control amount prediction unit 7 based on the virtual operation amount U (^) _n and the controlled variable X _'n, the virtual operating amount U (^) predicted controlled amount after pΔt time _{when n} is outputted as the adjustment signal X (~ )
Calculate _{n + p} . pΔt is the response time of the controlled object 1. The manipulated variable correcting unit 8 calculates p calculated by the control amount predicting unit 7.
Based on the Δt time control amount after X _{n + p} and the target control amount X ^d, modify the virtual operation amount U (^) _n, as predicted controlled variable X (~) n + _p approaches the target control amount X ^d . The first switch 9 closes downward when the controlled variable is measured by the controlled object 1, inputs the virtual manipulated variable U (^) _n calculated by the virtual manipulated variable calculating unit 6 into the controlled variable predicting unit 7, and thereafter. , The first switch 9 is closed upward until the next measurement, and the virtual manipulated variable U (^) _n corrected by the manipulated variable correcting unit 8 is input to the control amount predicting unit 7,
Prediction of the control amount by the control amount predicting unit 7 and operation by the operation amount correcting unit 8 until a predetermined number of times or a difference between the predicted control amount X (-) _{n + p} and the target control amount ^Xd becomes a predetermined value or less The amount modification is repeated. The control amount prediction unit 7 predicts the control amount and the operation amount correction unit 8 corrects the operation amount up to a predetermined number of times, or the predicted control amount X (-) _{n + p} and the target control amount Χ ^d
Is repeated until the difference becomes less than or equal to a predetermined value, the second switch 10 is closed, and the virtual manipulated variable U (^) _n becomes the adjustment signal C.
It is output to the operation unit 3 as _n .

FIG. 18 is a diagram showing an example of the configuration of the control amount predicting section 7 shown in FIG. Again, the control amount predicting unit 7
Is composed of a neural circuit model. In FIG. 18, the virtual operation amount U (^) _n = (U (^) 1 _n , U at time nΔt
_{(^) 2 n, ...,} U (^) g n) T (g represents the number of dimensions of the operation amount) and the controlled variable _{X'n = (X1 'n,} Χ2' n,
, Xh ' _n ) ^T (where h represents the number of dimensions of the controlled variable), the predicted controlled variable X (~) _{n + p} = (X
(~) 1 _{n + p} , X (~) 2 _{n + p} , ..., X (~) h _{n + p} )
The signal flow of the prediction calculation at the time of control of the neural circuit model 11 which outputs ^T is shown. The output signal is obtained from this input signal by the forward calculation of the neural circuit model 11 described above.

When the control amount predicting unit 7 is composed of the neural circuit model 11, the operation amount correcting unit 8 calculates the virtual operation amount U by the error backpropagation calculation of the neural circuit model 11 described above.
(^) _Obtain the correction amount of _n and correct the virtual operation amount.

Here, when the control amount predicting section 7 is composed of the neural circuit model 11, the learning section 4 is structured as follows.

The learning section 4 determines the response time pΔt of the controlled object 1.
The learning amount is created based on the operation delay time qΔt, the transmission dead time rΔt, and the observation data including the past control amount and the operation amount stored in the observation data storage unit 5, and the control amount prediction unit 7 is configured. Learning of the neural circuit model 11 is performed. When the operation unit 3 has an operation delay time and the transmission unit 21 has a transmission dead time, an input signal (U _s , Χ _s ) and an output signal (X _{s + p} ) (s = n−1, n−1 + 1) , ..., np
When learning is performed using -1) as learning data, the neural network model 11 causes the controlled variable X _s measured at the time sΔt and the time sΔ.
The control amount Χ _{s + p} after pΔt time is output from the operation amount U _s input to the controlled object 1 at t. When control is performed using this neural circuit model 11 in the control amount predicting unit 7, the control unit 2 controls the control amount X observed before the transmission dead time rΔt.
_nr is regarded as the current controlled _variable X _n, and the predicted controlled variable Χ
(-) _{N + p} is the manipulated variable U _n for approaching the target controlled variable Χ ^d
Since the value of is calculated and the value is actually input to the controlled object 1 after the operation delay time qΔt, accurate control cannot be performed.

Therefore, in the present invention, the neural circuit model 11
Is the control amount X ′ input to the adjustment unit 2 at time nΔt
_{When n} and the control signal C _n output at time nΔt are input, the neural circuit model 11 is trained so as to output the predicted control amount Χ (∼) _{n + p} after pΔt time. , And the controlled variable Χ _sr = (Χ1 _sr , X2 _sr , ..., Xh measured by the controlled object 1 before the transmission dead time rΔt.
_sr ) and the manipulated variable U _{s + q} = (U1 _{s + q} , U2 _{s + q} , ..., U input to the controlled object 1 after the operation delay time qΔt.
g _{s + q} ) is the input signal, and the controlled variable after pΔt time Χ _{s + p} =
Create learning data using (Χ1 _{s + p} , Χ2 _{s + p} , ..., Χ _{s + p} ) as a teacher signal (s = n−1 + r, n−1 + r).
+1, ..., Np-1). If learning can be performed with this learning data, the adjustment signal C _n obtained by the adjustment unit 2 during control is
When the input is made to the controlled object 1 after the operation delay time qΔt, p
Since the control amount _{An + p} after Δt time is the operation amount approaching the target control amount ^Xd , accurate control can be performed.

Next, another embodiment of the present invention will be described with reference to FIG.

In the control device shown in FIG. 20, an operation delay time measuring unit 20 and a transmission dead time measuring unit 22 are added to the control device shown in FIG. The operation delay time measuring unit 20 uses the time series C = (C _nm , C _{n-m + 1} , ..., C _n-1 ) of the adjustment signal calculated by the adjusting unit 2 in the past m unit time, and the operation unit 3
Of the manipulated variable output from U = (U _nm ,
U _{n-m + 1} , ..., U _n-1 ) based on the adjustment signal
The operation delay time qΔt from the input to the output to the output as the operation amount is calculated and output to the learning unit 4. Further, the transmission dead time measurement unit 22 has a time series X = (Χ _nm , X _{n-m + 1} , ..., Χ) of the controlled variable measured by the controlled object 1 in the past m unit time.
_n-1 ) and the time series of the control amount input to the adjusting unit 2 Χ '=
Based on (Χ ' _nm , X'n _{-m + 1} , ..., Χ'n _-1 ), the transmission dead time rΔt until the controlled variable measured by the controlled object 1 is input to the adjusting unit 2 is calculated. , To the learning unit 4. m is a positive integer. By adding the operation delay time measuring unit 20, even when the operation delay time changes due to the characteristic change of the operation unit 3, the operation delay time measuring unit 20
The operation delay time qΔt after change is calculated by and is output to the learning unit 4. In the learning unit 4, the operation delay time measuring unit 2
Since learning data is created and learning is performed according to the operation delay time qΔt input from 0, correct control can be continued even when the operation delay time changes. Further, by adding the transmission dead time measurement unit 22, the transmission unit 2
Even if the transmission dead time changes due to the characteristic change of 1, the transmission dead time measuring unit 22 calculates the changed transmission dead time rΔt and outputs it to the learning unit 4. The learning unit 4 creates learning data according to the transmission dead time rΔt input from the transmission dead time measuring unit 22 and performs learning, so that correct control can be continued even when the transmission dead time changes.

[0097]

EXAMPLE An example in which the present invention is applied to a device for controlling the dissolved oxygen concentration in a sewage treatment process for controlling the dissolved oxygen concentration in the sewage treatment process will be described.

FIG. 21 is a block diagram showing the configuration of the control device according to an embodiment of the present invention.

As shown in the figure, the sewage treatment process 23
Then, the dissolved oxygen concentration which is the control amount and the sewage inflow amount are measured for each unit time Δt, and the aeration air amount which is the operation amount is input. Dissolved oxygen concentration DO _n , sewage inflow amount Qinf _n measured at time nΔt and sewage treatment process 23 before unit time
The aeration air amount Qaer _n-1 input to the control unit 25 is input to the adjustment unit 25 through the transmission unit 24. However, since the transmission unit 24 has a transmission dead time, it is input to the adjustment unit 25 with a delay in the transmission dead time. The aeration air volume input to the control unit 25 at time nΔt is represented by Qaer ′ _n , the dissolved oxygen concentration is represented by DO ′ _n , and the sewage inflow volume is represented by Qinf ′ _n−1 . The control unit 25 controls the target dissolved oxygen concentration DO ^d , the aeration air amount Qaer ′ _n−1 , and the dissolved oxygen concentration DO ′.
_The control signal C _n is calculated based on _n and the sewage inflow amount Qinf ′ _n . In the blower 26, the control signal C _n calculated by the control unit 25 is converted into the aeration air amount Qaer _n and output to the sewage treatment process 23. Aeration amount QAER _n the blower 26 has output is input to the wastewater treatment process 23.

The operation delay time measuring unit 27 uses the time series C = (C of the adjustment signal calculated by the adjusting unit 25 in the past m unit time.
_nm , C _{n-m + 1} , ..., C _n-1 ) and the time series Qaer = (Qaer _nm , Q) of the aeration air volume output from the blower 26.
based on aer _nm-1 , ..., Qaer _n-1 ), the operation delay time qΔt until the control signal is realized as the aeration air volume.
Is calculated. m is a positive integer.

The transmission dead time measuring unit 28 uses the time series DO = (D Ο ₁ , D Ο ₂ , ..., DO _n-1 ) of the dissolved oxygen concentration observed in the sewage treatment process 23 in the past m unit time, and the adjusting unit 25. Time series of dissolved oxygen concentration DO '= (D
Based on O ′ ₁ , DO ′ ₂ , ..., DO ′ _n−1 ), the transmission dead time rΔt from when the dissolved oxygen concentration is observed in the sewage treatment process 23 to when it is input to the adjusting unit 25 is calculated.

The learning unit 29 receives the response time pΔt from the input of the aeration air volume to the sewage treatment process 23 until the influence of the aeration air volume appears in the dissolved oxygen concentration, and the operation delay time calculated by the operation delay time measurement unit 27. qΔt,
The transmission dead time rΔt calculated by the transmission dead time measuring unit 28 and the past aeration air amount (Qaer _nl , Qaer _{n-l + 1} , ..., Qae) stored in the observation data storage unit 30.
r _n-1 ) and the dissolved oxygen concentration (DO _nl , DO _{n-l + 1} , ...,
DO _n-1 ) and the inflow amount (Qinf _nl , Qinf _{n-l + 1} ,
, Qinf _n-1 ) based on the observed data, the parameter W of the adjusting unit 25 is adjusted by learning and output to the adjusting unit 25.

Next, the adjusting unit 25 is a virtual aeration air volume calculating unit,
The case of being composed of the dissolved oxygen concentration prediction unit and the aeration air volume correction unit will be described with reference to FIG. The virtual aeration air flow rate calculation unit 31 calculates the virtual aeration air flow rate Qaer based on the aeration air flow rate Qaer ′ _n−1 before the unit time, the dissolved oxygen concentration DO ′ _n and the target dissolved oxygen concentration DO ^d measured in the sewage treatment process 23.
(^) Calculate _n . A specific method of this calculation is, for example, when K is a positive constant,

[Equation 17] May be calculated as

(Equation 18) As described above, the amount of aeration air before a unit time may be used as it is as the virtual amount of aeration air.

The dissolved oxygen concentration predicting unit 32 uses the virtual aeration air quantity Q.
aer (^) _n and aeration amount Qaer before unit time
_'N-1, the dissolved oxygen concentration DO' _n measured in sewage treatment process 23, based on the sewage inlet flow Qinf' _n, pΔt time when the virtual aeration amount Qaer _{(^) n} is outputted as the adjustment signal _{C n} Predicted dissolved oxygen concentration after DO (-)
Calculate _{n + p} . p is the response time of the sewage treatment process 23.

The aeration air flow rate correction unit 33 calculates the predicted dissolved oxygen concentration DO based on the dissolved oxygen concentration DO (−) _{n + p} and the target dissolved oxygen concentration DO ^d predicted by the dissolved oxygen concentration prediction unit 32.
(-) _{n + p} to modify the virtual aeration air quantity Qaer _{(^) n} so as to approach the target dissolved oxygen concentration DO ^d.

The first switch 34 is for the sewage treatment process 23.
When the dissolved oxygen concentration is measured at, it closes to the lower side, and the virtual aeration air volume Qaer (^) calculated by the virtual aeration air volume calculation unit 31.
_n is input to the dissolved oxygen concentration predicting unit 32, and then the first switch 34 is closed upward until the next measurement, and the virtual aeration air amount Qaer (^) _n corrected by the aeration air amount correction unit 33 is used as the dissolved oxygen concentration. Dissolved oxygen by the dissolved oxygen concentration predicting unit 32 is input to the predicting unit 32 until a predetermined number of times or until the difference between the predicted dissolved oxygen concentration DO (−) _{n + p} and the target dissolved oxygen concentration DO ^d becomes a predetermined value or less. The prediction of the concentration and the correction of the aeration air volume by the aeration air volume correction unit 33 are repeated. The dissolved oxygen concentration prediction unit 32 predicts the dissolved oxygen concentration and the aeration air amount correction unit 33 corrects the aeration air amount up to a predetermined number of times, or the difference between the predicted dissolved oxygen concentration DO (−) _{n + p} and the target dissolved oxygen concentration DO ^d . Is repeated until the value becomes less than or equal to a predetermined value,
The second switch 35 is closed, and the virtual aeration air amount Qaer (^)
_n is output to the blower 26 as a control signal C _n .

Next, the case where the dissolved oxygen concentration predicting unit 32 is composed of a multi-layer neural circuit model will be described. FIG.
3 is an explanatory diagram of an example of the dissolved oxygen concentration predicting unit 32. The dissolved oxygen concentration predicting unit 32 has a virtual aeration air amount Qaer (^).
_n , dissolved oxygen concentration DO ′ _n , sewage inflow amount Qinf ′ _n ,
The aeration air amount Qaer ' _n-1 before the unit time is input. A time delay occurs due to the delay time elements 37 arranged in multiple _stages , and the time (n− _maer + 1) Δt to the time (n−1) Δ.
Aeration amount up to t (Qaer ' _{n-maer + 1} , Qaer'
_{n-maer + 2} , ..., Qaer ' _n-1 ), time (n-m _do +
1) Dissolved oxygen concentration (DO ′) from Δt to time nΔt
_{n-mdo + 1} , DO ' _{n-mdo + 2} , ..., DO' _n ), time (n
−m _inf +1) Δt to time nΔt sewage inflow (Qinf ′ _{n- minf + 1} , Qinf ′ _{n-minf + 2} , ...,
Qinf ′ _n ) and the virtual aeration amount Qaer (^) _n are input to the neural circuit model 36. m _aer , m _do , and m _inf are positive integers. The neural circuit model 36 calculates the predicted dissolved oxygen concentration DO (-) _{n + p} at time (n + p) [Delta] t from these input signals by forward calculation.

Other structures of the dissolved oxygen concentration predicting section 32 are possible. For example, as shown in FIG. 24, time (n + p) Δ
t prediction dissolved oxygen concentration DO (-) n + _p of the time variation of the dissolved oxygen concentration DO' _n of n.DELTA.t Delta] DO (-) of n + _p is calculated by the neural network model, the calculated amount of change and time n
Dissolved oxygen concentration Diomikuron' _n of Δt is added, the time (n +
p) The predicted dissolved oxygen concentration DO (−) _{n + p} of Δt may be output.

The aeration air flow rate correction unit 33 calculates the predicted dissolved oxygen concentration D based on the predicted dissolved oxygen concentration DO (-) _{n + p} calculated by the dissolved oxygen concentration prediction unit 32 and the target dissolved oxygen concentration DO ^d.
O (-) _{n + p} modifies the virtual aeration amount Qaer _{(^) n} to approach the target dissolved oxygen concentration DO ^d. When the dissolved oxygen concentration prediction unit 32 is composed of the neural circuit model 36, the aeration air volume correction unit 33 has the same configuration as the neural circuit model 36 of the dissolved oxygen concentration prediction unit 32, and the error inverse of the neural circuit model 36 is reversed. Virtual aeration air volume Qaer by propagation calculation
(^) _Obtain the correction amount of _n and make corrections.

Next, the learning section 29 will be described. When the dissolved oxygen concentration predicting unit 32 is composed of the neural circuit model 36, the learning unit 29 causes the response time p of the sewage treatment process 23.
Δt, the operation delay time qΔt measured by the operation delay time measuring unit 27, the transmission dead time rΔt measured by the transmission dead time measuring unit 28, and the observation data storage unit 3
Learning data is created on the basis of the observation data composed of the past dissolved oxygen concentration, the sewage inflow amount, and the aeration air amount stored in 0, and the neural network model 36 constituting the dissolved oxygen concentration prediction unit 32 is learned.

The observation data stored in the observation data storage unit 30 are the aeration air volume input to the sewage treatment process 23 in the past and the dissolved oxygen concentration measured in the sewage treatment process 23.
It is composed of sewage inflow. Use the observation data to the learning of the neural circuit model 36, the input signal _(Qaer s, Qaer
_s-1 , ..., Qaer _{s-maer + 1} , DO _s , DO _s-1 ,
…, DO _{s-mdo + 1} , Qinf _s , Qinf _s-1 ,…, Q
inf _s-m inf _{+ 1} ), output signal DO _{s + p} (S = n-1 +
m _max , n-l + m _max +1, ..., n-p-1) (m
_{If max} is the maximum value of m _aer , m _do , and m _inf ) and the learning data is used as the learning data, the dissolved oxygen concentration predicting unit 32 is controlled.
Then, the dissolved oxygen concentration observed before the transmission dead time rΔt,
The dissolved oxygen concentration after the response time pΔt is predicted by using the sewage inflow and the aeration air flow as the currently observed values, and further, the predicted dissolved oxygen concentration DO _{n + p} approaches the target dissolved oxygen concentration DO ^d. Although aeration amount QAER _n obtained is aeration amount input at time nΔt sewage treatment process 23, since the inputted actual sewage treatment process 23 is after the operation delay time Qderutati, it can not be accurately controlled .

In the present invention, the dissolved oxygen concentration DO ′ _n and the sewage inflow amount Qinf ′ input to the control unit 25 at time nΔt.
_{When n} , the aeration air amount Qaer ′ _n−1, and the control signal C _n output from the control unit 25 are input to the dissolved oxygen concentration prediction unit 32, the predicted dissolved oxygen concentration DO _{n +} after the response time pΔt of the sewage treatment process 23 is input. Neural circuit model 3 to output _p
6, the operation delay time measuring unit 27 calculates the operation delay time qΔt, the transmission dead time measuring unit 28 calculates the transmission dead time rΔt, and the operation delay time qΔ is calculated.
Aeration air volume Qa input to the sewage treatment process 23 after t
er _{s + q} and (r + 1) Δt before (r + _maer −1) Δ
The amount of aeration air (Qaer _sr-1 , Qaer _sr-2 , ..., Qaer input to the sewage treatment process 23 until t before
_sr-maer +1) and rΔt before (r + m _do -1) Δt
Dissolved oxygen concentration (DO _sr , DO _sr-1 , ..., D _{sr-mdo + 1} ) and sewage inflow (Qinf _sr , Qinf _sr-1 , ..., Qinf) measured in the sewage treatment process 23 until before.
_{sr-minf + 1} ) is used as an input signal, and learning data using the dissolved oxygen concentration DO _{s + p} after the response time pΔt as an output signal is created (s = n−1 + r + _maer , ... _Np −1). If learning can be performed with this learning data, the control signal C _n obtained by the control unit 25 at the time of control is such that the dissolved oxygen concentration approaches the target dissolved oxygen concentration when input to the sewage treatment process 23 after the operation delay time by the blower 26. Because of the air volume, accurate control is possible.

Next, the calculation procedure of the entire control device will be described with reference to the flowchart shown in FIG. In the flow chart of FIG. 25,
Although written so as to control from time 0 to time ΔΔt, this calculation procedure repeats the same procedure every unit time Δt, and therefore the procedure for obtaining the aeration air amount at time nΔt is shown here.

(A) The target dissolved oxygen concentration DO ^d set in advance, the observed dissolved oxygen concentration DO ' _n , and the aeration air amount Qaer' _n-1 before the unit time are the virtual aeration air amount calculation unit 3
1 and the virtual aeration air amount Qaer (^) _n is output.

(B) The first switch 34 is closed downward,
The virtual aeration air amount Qaer (^) _n obtained in (a),
Aeration amount Qaer ' _n-1 before unit time, dissolved oxygen concentration D
O ′ _n and the sewage inflow amount Qinf ′ _n are input to the dissolved oxygen concentration prediction unit 32, and the predicted dissolved oxygen concentration DO (−) after the response time pΔt is calculated by the forward calculation of the neural circuit model 36.
_{n + p} is output.

The predicted dissolved oxygen concentration DO (−) _{n + p} and the target dissolved oxygen concentration DO ^d obtained in (c) and (b) are input to the aeration air volume correction unit 33, and the error back propagation calculation of the neural circuit model 36 is performed. The correction amount of the virtual aeration air amount is obtained by the above, and the corrected virtual aeration air amount Qaer (^) _n is output.

(D) The first switch 34 is closed to the upper side for a predetermined number of times or until the difference between the target dissolved oxygen concentration DO ^d and the predicted dissolved oxygen concentration DO (-) _{n + p} is within a predetermined range, Fixed virtual aeration amount QAER _n is input to the dissolved oxygen concentration predicting section again are repeated procedures (b) (c).

When the conditions (e) and (d) are satisfied, the second switch 35 is closed, and the virtual aeration air amount Qaer (^) _n is input to the blower 26 as the control signal C _n , and the blower 26 informs the sewage treatment process 23. aeration air quantity Qaer _n is input.

The procedure from (a) to (e) is repeated every unit time Δt.

Next, the forward calculation, the error backpropagation calculation, and the learning calculation of the neural network model 36 will be described. When the dissolved oxygen concentration predictor 32 is composed of a multilayer neural network model, the predicted value of the dissolved oxygen concentration is obtained by the forward calculation of the multilayer neural network model, and the correction amount of the aeration air volume is the error back propagation calculation of the multilayer neural network model. Required by.

First, the forward calculation will be described.
Here, the forward calculation will be described with reference to FIG.

The unit of the input layer has a virtual aeration air quantity Qae.
r (^) _n , from time (n-1) [Delta] t to time (n- _maer)
Aeration volume up to +1) Δt (Qaer ' _n-1 , Qaer
_{'N- 2, ..., Qaer' n} -maer + 1), dissolved oxygen concentration from time nΔt time (n-m _do +1) to Δt (DO'
_n , DO ' _n-1 , ..., DO' _{n-mdo + 1} ), from the time nΔt to the time (n-m _inf +1) Δt, the sewage inflow (Qi
nf ' _n , Qinf' _n-1 , ..., Qinf '
_{n −minf +1} ) is input to one unit one signal at a time,
The predicted value DO _{n + p} of the dissolved oxygen concentration at time (n + p) Δt is output. m _do , m _inf and m _aer each represent a positive integer. The formula shows the forward calculation for calculating the output signal from this input signal.

If xi _i is an input value of the i-th unit of the input layer,

[Equation 19] The input / output function of the input layer unit is the identity function. The output yi _i (i = 1, ..., N i) of the input layer unit is expressed as follows.

[0124]

(Equation 20) Νi is the number of input layer units (Ni = m _do + m _aer +
m _inf ). Next, the input value xh of the j-th unit in the middle layer
_j is the threshold value t of the intermediate layer unit from the sum of the output value of the input layer unit weighted by the coupling weight value between the input layer and the intermediate layer
It is a value obtained by subtracting hh _j .

[0125]

(Equation 21) w1 _{j, i} is the coupling load value between the input layer i-th unit and the intermediate layer j-th unit, thh _j is the threshold value of the intermediate layer j-th unit, and Νh is the number of units of the intermediate layer.

The input / output function of the intermediate layer unit is a sigmoid type function.

(Equation 22) Then, the output force value yh _j of the intermediate layer unit is the input value x
use h _j ,

(Equation 23) Can be expressed as yh _j is the output value of the j-th unit in the middle layer. Input value xo _k (k = 1) of the kth unit in the output layer
Is the value obtained by subtracting the threshold value of the output layer unit from the sum of the output value of the intermediate layer unit weighted by the coupling weight value between the intermediate layer and the output layer.

[0127]

(Equation 24) w2 k, _j is the intermediate layer j-th unit - bond load value between the output layer the k-th unit, tho _k is the threshold of the k-th unit output layer. The input / output function of the output layer unit is an identity function.

[0128]

(Equation 25) The output value of this output layer unit is the predicted dissolved oxygen concentration DO
(−) _{N + p} .

[0129]

(Equation 26) The above is the calculation formula for the forward calculation of the three-layer neural circuit model.

Next, the error back propagation calculation will be described.
When the dissolved oxygen concentration DO (−) _{n + p} predicted by the dissolved oxygen concentration prediction unit 32 and the target dissolved oxygen concentration DO ^d are input to the aeration air amount correction unit 33, the aeration air amount correction unit 33 causes the virtual aeration air amount. Calculate the correction amount of and correct it. When the dissolved oxygen concentration predicting unit 32 is composed of the neural circuit model 36, the correction amount of the virtual aeration volume is the neural circuit model 36.
It can be calculated by the error back-propagation calculation. The calculation formula of the aeration air volume correction unit 33 is shown below.

[0131] First, the dissolved oxygen concentration predicting section 32 by the predicted dissolved oxygen concentration DO (-) and n _{+ p,} the error function E from the error between the target dissolved oxygen concentration DO ^d is defined as follows.

[0132]

[Equation 27] Next, the correction amount of the unit output value that reduces this error function is sequentially obtained for each layer from the output layer to the input layer, and finally the correction amount of the input signal is obtained. First, the correction amount of the output value of the output layer unit is calculated by using the positive constant η.

[Equation 28] Is calculated. Next, the correction amount of the output value of the intermediate layer unit is expressed as follows using the correction amount of the output layer unit.

[0133]

(Equation 29) Similarly, the correction amount of the output value of the input layer unit, using the correction amount of the middle layer unit,

[Equation 30] It is expressed as f ′ (x) represents the derivative of the sigmoid function f (x), and specifically,

(Equation 31) Can be expressed as Since the input / output function of the input layer unit is the identity function, the correction amount of the output value becomes the correction amount of the input value as it is, and the correction amount ΔQae of the virtual aeration air amount Qaer (^) _n.
r and _n are

(Equation 32) Is expressed as This correction amount is the virtual aeration air amount Qaer
(^) It is added to _n . The above is the error backpropagation calculation and the virtual aeration air volume stop achieved by the calculation.

Next, learning of the multilayer neural network model will be described with reference to FIG. The learning is performed because the dissolved oxygen concentration prediction unit 32 adapts to the gradual change in characteristics of the sewage treatment process.

In FIG. 26, 30 is an observation data storage unit, 38 is a subtractor, 39 is an error backpropagation learning device, and 40 is an output value of the input layer unit of the neural network model 36, which is input to the error backpropagation learning device 39. Arrow 41 indicating that the output value of the intermediate layer unit of the neural network model 36 is input to the error backpropagation learning device 39, and 42 is the coupling load between the intermediate layer and the output layer of the neural circuit model 36. An arrow indicates that the value is input to the error backpropagation learning device 39, and 43 is a connection weight value between the input layer and the intermediate layer calculated by the error backpropagation learning device 39 and a correction amount of the threshold value of the intermediate layer unit, which are combined. The weights are output to the intermediate layer unit, and the arrows indicate that the weight values and threshold values are corrected, and 44 is the error backpropagation learning device 39.
Outputs the combined load value between the middle layer and the output layer and the correction amount of the threshold value of the output layer unit, which are calculated in, to each combined load and output layer unit, and represents an arrow indicating that each combined load value and the threshold value are corrected. .

Past time sΔt (s = n-1 + _maex ,
..., for the n-p-1), time (s + q) aeration amount of Δt Qaer _{s + q} and time _{(s-r-m aer +1} ) time from Δt (s-r-1) aeration air flow rate of up to Δt (Qaer
_{sr-maer + 1, Qaer sr-} maer + 2, ..., Qaer
_sr-1 ), time (s-r-m _do +1) Δt to time (s
-R) Dissolved oxygen concentration up to Δt (D _{sr-mdo + 1} , DO
_{sr-mdo + 2} , ..., DO _{s- r} ) Time (s-r-m _inf +
1) Sewage inflow (Q) from Δt to time (s−r) Δt
inf _{sr-minf + 1,} Qinf _{sr-minf + 2} , ..., Qi
nf _sr ) is output from the observation data storage unit 30 to the neural circuit model 36 as an input signal. Neural circuit model 3
At 6, the forward calculation is performed on the input signal, and the output signal DO (-) _{s + p} is output. In the subtractor 38, the difference DO _{s + p} − between the dissolved oxygen concentration DO _{s + p} after the response time output from the observation data storage unit 30 as a teacher signal and the output signal DO (−) _{s + p} of the neural circuit model 36. DΟ (−) _{s + p} is calculated and input to the error backpropagation learning device 39. In addition to the difference, the output value of the input layer unit of the neural network model 36, each output value of the intermediate layer unit, and each coupling weight value between the intermediate layer and one output layer are input to the error backpropagation learning unit 39, and the coupling weight is input. The value and the correction amount of the threshold value are calculated.

In the subtractor 38, the error function E is defined as follows from the difference between the teacher signal DO _{s + p} and the output signal DO (−) _{s + p} of the neural circuit model 36.

[0138]

[Equation 33] According to the error backpropagation learning method, the connection weight value of the neural network model 36 and the correction amount of the threshold value are calculated as follows.

[0139]

(Equation 34) Derutadaburyu2 k, _j is the intermediate layer j-th unit - correction amount of the coupling weight value w2 _{k, j} between the output layer the k-th unit, Derutatho _k is the correction amount of the threshold tho _k of the k-th unit output layer, .DELTA.W1
_{j, i} is the correction amount of the coupling load value w1 _{j, i} between the input layer i-th unit and the intermediate layer j-th unit, Δthh _j is the correction amount of the threshold thh _j of the intermediate layer j-th unit, and ε is the learning value. It is a constant. yo ^s _k is the output value of the output layer the k th unit when the data in time sΔt is input, the input value of the output layer the k th unit when data xo ^s _k time sΔt is input, yh ^s _j output values of the intermediate layer j-th unit of time in which data of time sΔt is input, the input value of the intermediate layer j-th unit when xh ^s _k is the data of the time sΔt is input, yi ^s _i at time It is the output value of the i-th unit in the middle layer when the data of sΔt is input. The coupling load value and the threshold value are corrected based on these correction amounts.

The present invention is not limited to the above-mentioned embodiments, and it goes without saying that various modifications can be made without departing from the gist of the present invention. For example, in the above-mentioned embodiment, the three-layer type neural circuit model is given as an example of the prediction model.
It can be similarly applied to a multilayered neural circuit model, and can be similarly applied to a function approximation model having a learning function. Regarding the input / output signal of the dissolved oxygen concentration predicting unit, in the embodiment, the dissolved oxygen concentration after pΔt time is predicted from the dissolved oxygen concentration, the sewage inflow amount, and the aeration air amount measured in the past and the present, Within the predictable range,
It is also possible to reduce the input signal or input another observation data. As for the output signal, not only the dissolved oxygen concentration after pΔt time is predicted, but also the dissolved oxygen concentration at multiple times are predicted, and the virtual aeration volume is corrected from the error between the predicted dissolved oxygen concentration and the target dissolved oxygen concentration. You can also find the quantity.

[0141]

As described above, according to the present invention,
Since the time factor existing in the feedback control loop is taken into consideration, more accurate control can be performed. That is, the learning unit performs learning in consideration of the response time from the input of the operation amount to the control target until the influence of the operation amount appears in the control amount, so that more accurate control can be performed.

When the operation section has an operation delay time, the observation data does not include the adjustment signal, and the operation amount input to the control target and the control amount measured by the control target are included,
By measuring the operation delay time of the operation unit and using the operation amount input to the control target after this operation delay time as the input signal of the operation amount of the learning data, the control amount prediction unit operates the input operation amount. Since the control amount when it is input to the controlled object after the delay time is predicted, it is possible to perform accurate control without being affected by the delay time by the operation unit during control.

Further, when there is a time delay in transmitting the measured value measured by the controlled object to the adjusting section, the transmission dead time associated with this transmission is measured and measured before this transmission dead time. By using the control amount as the input signal of the control amount of the learning data, the control amount prediction unit measures the control amount and the operation amount measured before the transmission dead time and input to the control amount prediction unit after the delay time accompanying the transmission. Since the control amount is predicted from this, accurate control can be performed without being affected by the transmission time delay associated with transmission during control.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining the principle of the present invention.

FIG. 2 is a diagram showing in detail one configuration example of an adjusting unit in FIG.

FIG. 3 is a diagram showing a configuration example in which a controlled variable predicting unit in FIG. 2 is composed of a neural circuit model.

FIG. 4 is a diagram for explaining learning of a neural circuit model.

FIG. 5 is a block diagram of a case where a response time calculation unit that calculates a response time of a controlled object is provided.

FIG. 6 is a block diagram for explaining the principle of the present invention when the operation delay of the operation unit is considered.

FIG. 7 is a diagram showing in detail one configuration example of an adjusting unit in FIG.

FIG. 8 is a diagram showing a configuration example in the case where the control amount prediction unit in FIG. 7 is composed of a neural circuit model.

9 is a diagram for explaining learning of the neural circuit model of FIG.

FIG. 10 is a block diagram for explaining the principle of the present invention when the operation delay time measuring unit for measuring the operation delay time of the operation unit is equipped.

FIG. 11 is a block diagram for explaining the principle of the present invention in consideration of a transmission time delay of a transmission unit.

FIG. 12 is a diagram showing in detail one configuration example of an adjusting unit in FIG.

FIG. 13 is a diagram showing a configuration example in the case where the control amount prediction unit in FIG. 12 is composed of a neural circuit model.

FIG. 14 is a diagram for explaining learning of the neural circuit model of FIG.

FIG. 15 is a block diagram for explaining the principle of the present invention when the transmission dead time measuring unit for measuring the transmission dead time of the transmission unit is installed.

FIG. 16 is a block diagram for explaining the principle of the present invention in consideration of the operation delay time of the operation unit and the transmission dead time of the transmission unit.

FIG. 17 is a diagram showing in detail one configuration example of an adjusting unit in FIG.

FIG. 18 is a diagram showing a configuration example in the case where the control amount prediction unit in FIG. 17 is composed of a neural circuit model.

19 is a diagram for explaining learning of the neural circuit model of FIG.

FIG. 20 is a view for explaining the principle of the present invention when the operation delay time measuring unit for measuring the operation delay time of the operation unit and the transmission dead time measuring unit for measuring the transmission dead time of the transmission unit are provided. It is a block diagram.

FIG. 21 is a block diagram of an example in which the present invention is applied to control of dissolved oxygen concentration in a sewage treatment process.

22 is a diagram showing in detail an example of the configuration of the adjusting section in FIG. 21. FIG.

23 is a diagram showing a configuration example in the case where the control amount prediction unit in FIG. 22 is composed of a neural circuit model.

FIG. 24 is a diagram showing a configuration example in the case where the control amount prediction unit in FIG. 22 is composed of a neural circuit model.

FIG. 25 is a flow chart for explaining a calculation procedure of the control device.

FIG. 26 is a diagram for explaining learning of the neural circuit model of FIG. 23.

[Explanation of symbols]

 1 ... Control target 2 ... Adjusting unit 3 ... Operating unit 4 ... Learning unit 5 ... Observation data storage unit 6 ... Virtual operation amount calculation unit 7 ... Control amount prediction unit 8 ... Operation amount correction unit 9 ...... First switch 10 ...... Second switch 12 ...... Subtractor 13 ...... Error back propagation learning device 19 ...... Response time calculation unit 20 ...... Operation delay time measurement unit 21 ...... Transmission unit 22 ...... Transmission dead time Measuring unit

Claims (20)

[Claims] 1. A control device for controlling a controlled variable of a controlled object by changing an operation amount input to the controlled object, wherein an adjustment signal is calculated based on a controlled variable and a target controlled variable measured by the controlled object. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Section, based on the observation time stored in the observation data storage unit and the response time of the control object from the input of the operation amount from the operation unit to the control object until the influence of the operation amount appears in the control amount And a learning unit that learns the adjusting unit. 2. A control device for controlling a controlled variable of a controlled object by changing an operation amount input to the controlled object, wherein an adjustment signal is calculated based on the controlled variable and the target controlled variable measured by the controlled object. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Section, based on the operation amount output from the operation unit and the control amount measured by the control target, after the operation amount is input from the operation unit to the control target, the influence of the operation amount appears in the control amount. Up to the response time calculation unit for calculating the response time of the controlled object, based on the response time calculated by the response time calculation unit and the observation data stored in the observation data storage unit,
A control unit comprising: a learning unit that learns the adjustment unit. 3. The control device according to claim 1, wherein the adjustment unit calculates a virtual operation amount based on the control amount and the target control amount measured by the control target, and a virtual operation amount calculation unit. A control amount prediction unit that predicts a control amount after the response time based on a control amount and a virtual operation amount measured by the control target; and a predicted control amount after a response time predicted by the control amount prediction unit and An operation amount correction unit that corrects the virtual operation amount based on the target control amount, a control amount prediction performed by the control amount prediction unit, and a virtual operation amount correction performed by the operation amount correction unit at least once. A control device, comprising means for inputting an operation amount as an adjustment signal to the operation unit. 4. The control device according to claim 3, wherein the control amount predicting unit is composed of a neural circuit model, and the calculation of the operation amount correcting unit is realized by an error backpropagation calculation of the neural circuit model, the learning unit. Is a learning device that creates learning data based on the response time and the observation data stored in the observation data storage unit, and learns a neural circuit model that constitutes the control amount prediction unit. . 5. The control device according to claim 4, wherein, when the learning unit creates learning data based on the observation data stored in the observation data storage unit, the control amount and the operation amount at the same time. The control device is characterized in that learning is performed by using observation data having the following as an input signal and using a control amount after the response time as a teacher signal. 6. A control device for controlling a control amount of a control target by changing an operation amount input to the control target, wherein an adjustment signal is calculated based on a control amount and a target control amount measured by the control target. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Section, the response time of the control target from the input of the operation amount from the operation unit to the control target until the influence of the operation amount appears in the control amount, and the adjustment signal calculated by the adjustment unit through the operation unit. A control device comprising: a learning unit that learns the adjustment unit based on an operation delay time until it is realized as an operation amount and the observation data stored in the observation data storage unit. 7. A control device for controlling a control amount of a control target by changing an operation amount input to the control target, wherein an adjustment signal is calculated based on the control amount and the target control amount measured by the control target. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Unit, the operation delay time until the adjustment signal calculated by the adjusting unit is realized as the operation amount through the operating unit, based on the adjusting signal calculated by the adjusting unit and the operation amount output by the operating unit. The operation delay time measuring unit for measuring the operation delay time, the operation sending time measured by the operation delay time measuring unit, and the operation amount input to the control target from the operation unit, and then the influence of the operation amount is the control amount. Based on the stored observation data in the observation data memory unit and the response time of the control object to appear, the control apparatus characterized by comprising a learning section for learning the adjustment portion. 8. The control device according to claim 6, wherein the adjustment unit calculates a virtual operation amount based on a control amount and a target control amount measured by the control target, and a virtual operation amount calculation unit. A control amount prediction unit that predicts a control amount after the response time based on a control amount and a virtual operation amount measured by the control target; and a predicted control amount after a response time predicted by the control amount prediction unit and An operation amount correction unit that corrects the virtual operation amount based on the target control amount, a control amount prediction performed by the control amount prediction unit, and a virtual operation amount correction performed by the operation amount correction unit at least once. A control device, comprising means for inputting an operation amount as an adjustment signal to the operation unit. 9. The control device according to claim 8, wherein the control amount prediction unit is composed of a neural circuit model, and the calculation of the operation amount correction unit is realized by an error backpropagation calculation of the neural circuit model, the learning unit. Based on the response time, the operation delay time, and the observation data stored in the observation data storage unit to create learning data and perform learning of the neural circuit model that constitutes the control amount prediction unit. Characteristic control device. 10. The control device according to claim 9, wherein, when the learning unit creates learning data based on the observation data stored in the observation data storage unit, the control amount at the time and the time A control device for performing learning by using as input signals observation data having an operation amount after an operation delay time and using a control amount after the response time as a teacher signal. 11. A control device for controlling a control amount of a control target by changing an operation amount input to the control target, wherein an adjustment signal is calculated based on the control amount and the target control amount measured by the control target. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Section, the response time of the control target from the input of the operation amount from the operation unit to the control target until the influence of the operation amount appears in the control amount and the control amount measured by the control target are transmitted through the transmission unit. A control device comprising: a learning unit that learns the adjustment unit based on a transmission dead time until input to the adjustment unit and the observation data stored in the observation data storage unit. 12. A control device for controlling a controlled variable of a controlled object by changing an operation amount input to the controlled object, wherein an adjustment signal is calculated based on the controlled variable and the target controlled variable measured by the controlled object. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Unit, a transmission dead time measuring unit that measures a transmission dead time by the transmission unit based on a control amount measured by the control target and a control amount input to the adjustment unit through the transmission unit, and the transmission dead time The transmission dead time measured by the measuring unit, the response time of the controlled object from the input of the manipulated variable from the operation unit to the controlled object until the influence of the manipulated variable appears on the controlled variable, and the observation data Based on the stored observation data in the data storage unit, the control apparatus characterized by comprising a learning section for learning the adjustment portion. 13. The control device according to claim 11 or 12, wherein the adjusting unit calculates a virtual operation amount based on the control amount and the target control amount measured by the control target. A control amount prediction unit that predicts a control amount after the response time based on a control amount and a virtual operation amount measured by the control target; and a predicted control amount after a response time predicted by the control amount prediction unit and An operation amount correction unit that corrects the virtual operation amount based on the target control amount, a control amount prediction performed by the control amount prediction unit, and a virtual operation amount correction performed by the operation amount correction unit at least once. A control device, comprising means for inputting an operation amount as an adjustment signal to the operation unit. 14. The control device according to claim 13, wherein the control amount predicting unit is composed of a neural circuit model, the calculation of the operation amount correcting unit is realized by an error backpropagation calculation of the neural circuit model, and the learning unit. Based on the response time, the transmission time delay, and the observation data stored in the observation data storage unit to create learning data and perform learning of the neural circuit model that constitutes the control amount prediction unit. Characteristic control device. 15. The control device according to claim 14, wherein when the learning unit creates learning data based on the observation data stored in the observation data storage unit, the control amount before the transmission dead time is set. A control device which performs learning by using as input signals observation data having an operation amount at the time and a control amount after the response time as a teacher signal. 16. A control device for controlling a controlled variable of a controlled object by changing an operation amount input to the controlled object, wherein an adjustment signal is calculated based on the controlled variable and the target controlled variable measured by the controlled object. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Section, the response time of the control target from the input of the operation amount from the operation unit to the control target until the influence of the operation amount appears in the control amount and the control amount measured by the control target are transmitted through the transmission unit. It is stored in the observation data storage unit and the operation delay time until the control signal calculated by the control unit is realized as a manipulated variable through the operation unit. Observation data and on the basis the control apparatus characterized by comprising a learning section for learning the adjustment portion. 17. A control device for controlling a control amount of a control target by changing an operation amount input to the control target, wherein an adjustment signal is calculated based on the control amount and the target control amount measured by the control target. And an operation unit that outputs an operation amount to the controlled object based on an adjustment signal calculated by the adjustment unit, and an observation data storage that stores observation data having a control amount and an operation amount observed in the past. Unit, the operation delay time until the adjustment signal calculated by the adjusting unit is realized as the operation amount through the operating unit, based on the adjusting signal calculated by the adjusting unit and the operation amount output by the operating unit. And an operation delay time measuring unit for measuring the transmission delay time by the transmission unit based on the control amount measured by the controlled object and the control amount input to the adjustment unit through the transmission unit. A dead time measuring unit, an operation delay time measured by the operation delay time measuring unit, a transmission dead time measured by the transmission dead time measuring unit, and an operation amount input from the operation unit to the control target A learning unit for learning the adjusting unit based on the response time of the controlled object until the influence of the operation amount appears in the control amount and the observation data stored in the observation data storage unit. Control device. 18. The control device according to claim 16 or 17, wherein the adjusting unit calculates a virtual operation amount based on the control amount and the target control amount measured by the control target, and A control amount prediction unit that predicts a control amount after the response time based on a control amount and a virtual operation amount measured by the control target; and a predicted control amount after a response time predicted by the control amount prediction unit and An operation amount correction unit that corrects the virtual operation amount based on the target control amount, a control amount prediction performed by the control amount prediction unit, and a virtual operation amount correction performed by the operation amount correction unit at least once. A control device, comprising means for inputting an operation amount as an adjustment signal to the operation unit. 19. The control device according to claim 18, wherein the control amount predicting unit is composed of a neural circuit model, the calculation of the operation amount correcting unit is realized by an error backpropagation calculation of the neural circuit model, and the learning unit. Is based on the response time, the operation delay time, the transmission dead time, and the observation data stored in the observation data storage unit, creates learning data, and the neural network model that constitutes the control amount prediction unit. A control device characterized by performing learning. 20. The control device according to claim 19, wherein when the learning unit creates learning data based on the observation data stored in the observation data storage unit, the control amount before the transmission dead time is set. And a control amount after the operation delay time as an input signal, and a control amount after the response time as a teacher signal to perform learning.