JP2862308B2 - Controller adjustment method and adjustment system - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for adjusting a controller, and more particularly, to adjusting a control parameter without being restricted by a time response form of an input / output variable to be controlled. The present invention relates to a controller adjustment method and an adjustment system suitable for the present invention.

[Related Art] When a control target is controlled by a controller, it is necessary to adjust control parameters of the controller according to characteristics of the control target. As one of the methods, there is an adjustment method described in “Automatic tuning method for PID controller applying fuzzy inference” (13th System Symposium of the Society of Instrument and Control Engineers, '87 -11). Hereinafter, an outline of the conventional adjustment method will be described.

FIG. 16 shows a configuration of an auto tuner that realizes a conventional adjustment method.

The control system shown in FIG. 16 includes a control target 101, a PID controller 102 for giving an operation amount thereto, an adder 103 for calculating a deviation between a control amount y of the control target 101 and a target value r,
An auto tuner 104 for adjusting the control parameters of the PID controller 102 is provided.

The auto-tuner 104 extracts a characteristic amount for each of the target value r and the control amount y.
5, and a storage unit 106 for storing an adjustment rule that stores the empirical knowledge and know-how of a skilled operator related to tuning of control parameters as a qualitative adjustment rule.
A fuzzy inference unit 107 for obtaining control parameters using fuzzy inference based on the adjustment rule for the extracted feature amount.

Such adjustment of the parameters by the conventional auto-tuner is based on the characteristics of the overshoot amount E, damping ratio D, vibration period ratio R, etc. shown in FIG. This is performed by extracting quantities and determining control parameters by fuzzy inference based on these feature quantities.

In the fuzzy inference, for example, an adjustment rule of a qualitative expression as described below is represented by a fuzzy rule, and a control parameter is determined by a fuzzy operation from a feature amount using the fuzzy rule.

If the overshoot amount E and the damping ratio D are large, the proportional gain _Kp and the derivative time _Td are reduced.

The control parameters thus obtained are provided to the PID controller 102, and the controller 102 performs control based on the provided parameters.

[Problem to be Solved by the Invention] In the above-described conventional auto tuner, an adjustment rule is constructed as an element for appropriately adjusting a parameter. As described above, these adjustment rules create IF-THEN format rules for various changes in the process based on the empirical knowledge of a skilled person.
Conventionally, several tens of fuzzy rules are required, and there is a constraint that there must be no contradiction as a whole. Moreover, much of the work is done manually.

Therefore, the above-mentioned conventional technology has a fuzzy rule for adjustment, that is, a fuzzy rule. There is a problem that it takes time to construct an adjustment rule.

Further, the above-mentioned conventional technique uses a characteristic amount of a response waveform of a control amount y with respect to a step change of a target value r as a characteristic amount of a process which is a premise of fuzzy inference. However, in an operation in which the target value is not changed stepwise, such a feature amount cannot be extracted, so that there is a problem that it is difficult to adjust the controller.

An object of the present invention is to provide a controller adjustment method and an adjustment system capable of adjusting a control parameter without shortening the construction time of an adjustment rule and without limiting the time response of input / output variables of a control system. It is in.

[Means for Solving the Problems] In order to achieve the above object, the present invention takes in a time series signal or a characteristic amount of input / output variables of a control system, and uses a neural network based on these signals to control the controller. Adjust the parameters.

In neural network learning, the time response of various characteristics of the control system model is obtained using the control system model, and the time series signal or feature value of the input / output variables of the control system model obtained from the time response is obtained. Optimum control parameters corresponding to the characteristics of the control system model are obtained as learning input data, and the parameters are used as learning teacher data.

In addition, a time-series signal or characteristic amount of the input / output variable of the control system obtained from the time response of a good characteristic in the time response of the control target is stored, and the time-series signal or the characteristic amount of the input / output variable is learned. Control data corresponding to the time response with good characteristics,
The neural network can be learned by using this parameter as learning teacher data.

Further, according to the present invention, a neural network, and a means for capturing a time-series signal or feature amount of an input / output variable of a control system in which a controller and a control target are combined,
A controller tuning system is provided that determines the control parameters of the controller by a neural network based on these signals.

According to the present invention, a means for realizing a control system model, a time response for various characteristics of the control system model are obtained, and a time series signal or characteristic of an input / output variable of the control system model obtained from the time response. A means for extracting the amount and using it as learning input data, a means for obtaining an optimal control parameter corresponding to the characteristics of the control system model and using it as learning teacher data, and using the above learning input data and learning teacher data Means for learning the neural network by learning the adjustment rules of the neural network used for adjusting the control parameters of the controller.

Further, according to the present invention, a means for capturing and storing a time-series signal or a characteristic amount of an input / output variable of an actual control system in which a controller and a control target are combined, and outputting the input / output signal as necessary, A neural network for determining a control parameter of a controller using a time-series signal or a feature value of a variable, and a means for storing a control parameter corresponding to a time response of a good characteristic during adjustment of the controller to obtain training data for learning. And means for learning the neural network using the learning teacher data and the stored data corresponding to the time series signal or feature amount of the input / output variable of the actual control system.
A controller coordination system is provided.

[Operation] The present invention uses a model of a controlled object having various characteristics,
The time response of the control system model is obtained, the time series signal or feature value of the input / output variables of the control system model obtained from the time response is used as input data for learning, and the optimal control corresponding to the characteristics of the model to be controlled Since parameters are obtained and used as training data for learning to train the neural network, adjustment rules can be constructed automatically and in a short time.

In addition, since the time series signals or feature values of the input / output variables of the control system are acquired and the control parameters of the controller are adjusted by the neural network based on these signals, even if the process characteristics change, this change is input. It can be detected as a change in the time series signal of the output variable, and the controller parameters can be adjusted online according to the process characteristic change without being restricted by the time response form of the control system input / output variable. Characteristics can be maintained in a good state.

Embodiment FIG. 1 shows a configuration of an embodiment of a control system to which a parameter adjusting method of the present invention is applied.

In this embodiment, a controller 2 that controls a control target 1
It comprises a parameter adjustment system 3 for adjusting the control parameters of the controller 2 and an adjustment rule learning system 4 for constructing the adjustment rules of the parameter adjustment system 3 by learning.

A case where a PID (proportional / integral / derivative) controller is used as the controller 2 will be described. This PID
The transfer function G _c (s) of the controller is given by the following equation.

Here, K _p : proportional gain T _i : integration time T _d : differentiation time Also, as the controlled object 1, a case where the transfer function G _p (s) can be approximated by a first-order lag + dead time system represented by the following equation explain.

Here, K: gain T: time constant L: dead time The parameter adjustment system 3 takes in a time-series signal of input / output variables of the control target 1 and adjusts control parameters of the controller 2 based on the time-series signal. This adjustment process is represented by the following equation.

C = F (z) (3) where C: control parameter z: time-series signal of input / output variables F: adjustment function For control parameter C and PID controller, the following equation is obtained.

C = [K _p , T _i , T _d ] ^T (4) where T is a symbol representing transposition.
When approximation can be made by a next-delay + dead-time system, it is expressed by the following equation.

z = [y (t) y (t-1) ... y (tL) x (t) x (t-1) ... x (tL)] ^T (5) where y (tL): ( t-L) Control amount at the sampling time x (t-L): (t-L) Operation amount at the sampling time The adjustment function F (z) is, for example, a multilayer (m) shown in FIG.
Layer) neural network. FIG. 3 shows an example of the configuration of a unit which is a component of the neural network.

 The input / output relationship of this unit is represented by the following equation.

Here, u _j (k): the sum of the inputs to the j-th unit of the k-th layer v _j (k): the output w _ij (k−1, k) of the j-th unit of the k-th layer: the (k− 1) Weighting coefficient of the connection from the ith unit of the layer to the jth unit of the kth layer f: a function giving the input / output relationship of each unit The first layer of the neural network is the input layer, The output of the unit becomes the input signal to the neural network. In the embodiment of the present invention, the input signal to the neural network is the time-series signal Z of the input / output variable of the controlled object 1, and the correspondence is shown in the following equation.

The last layer (the m-th layer in the embodiment of the present invention) of the neural network is an output layer, and an output of a unit in this layer is an output signal of the neural network. In an embodiment of the present invention, the output signal of the neural network is a control parameter C
And the correspondence is shown in the following equation.

The adjustment function F (z) shown in the equation (3) is given by (6) and (7).
When the input / output relationship of the unit shown in the equation changes, it changes accordingly. That is, when the number of layers of the neural network, the number of units in each layer, the weight coefficient w _ij (k−1, k) of each unit, and the function f that gives the input / output relationship of each unit change, the adjustment function F (z ) Changes. Therefore, by adjusting the number of layers, the number of units in each layer, the weight coefficient w _ij (k−1, k) of each unit, and the function f that gives the input / output relation of each unit, the adjustment function F ( z) can be constructed.

The adjustment rule learning system 4 builds the adjustment rules of the parameter adjustment system 3 by learning.

 Next, the learning algorithm will be described.

First, when an input / output pair (z, c) is given as learning data, the square of the error represented by the following equation is defined as a loss function R.

Here, w: the sum of the weighting factors of the connection of the neural network v _j (m) (w, z): the number of the m-th layer (output layer) obtained comprehensively from the input z and the weighting factor w The correction amount Δw of the output w of the j unit is obtained from the gradient of w of the loss function R, and is expressed by the following equation.

Here, Δw _ij (k−1, k): a weighting coefficient w for coupling from the ith unit of the (k−1) th layer to the jth unit of the kth layer
The correction amount of _ij (k−l, k) ε: a positive constant ΔR / Δw _ij (k−l, k) on the right side of equation (11) can be transformed as follows.

Substituting equation (7) into equation (12) and rearranging yields the following equation.

When k ≠ m, ∂R / ∂u _j (k) on the right side of equation (13) is
It is obtained by the following equation.

Substituting equations (6) and (7) into equation (14) and rearranging:
The following equation is obtained.

Here, f ′: a derivative of a function f that gives an input / output relationship of each unit. In other words, equations (11) and (15) are represented by the following equations.

Further, when k = m, ∂R / ∂u _j (m) is obtained by the following equation from equation (10).

Using the above equations (10), (17), and (18), the modification of the weight coefficient w _ij (k−1, k) of the combination is recursively calculated from k = m toward k = 2. . That is, the error between the ideal output c _j and the actual output v _j (m) (w, z) at the output layer is input, and v _{jl is} applied from the output layer to the input layer in the direction opposite to the signal propagation. Propagation is performed while taking the sum weighted by (k, K + 1). This is the error back propagation learning algorithm.

It is assumed that a function f for giving an input / output relationship of each unit is common to all units and is expressed by the following equation.

From equation (19), the following equation is obtained.

 f ′ (u) = f (u) {1−f (u)} (20) From equations (6) and (20), the following equation is derived.

f ′ (u _j (k)) = v _j (k) ｛1−v _j (k)｝ (21) In order to smoothly and quickly converge the learning, the expression (16) is expressed by the following expression. Can be modified.

_{Δw ij (kl, k) (} τ) = - εv i (kl) (τ) · d j (k)
(Τ) + αΔw _ij (k−l, k) (τ−1) (16a) where α: positive constant (may be α = 1−ε) τ: number of corrections Input / output of learning data In the set (z, c), the input z is called input data for learning, and the output c is called teacher data for learning.

Next, a method for acquiring learning data according to the embodiment of the present invention will be described.

When adjusting the control parameters of the controller 2, the adjustment is performed according to a certain evaluation criterion. One of the evaluation criteria is a reference model. There is an adjustment method in which the transfer function of the control system partially matches the transfer function of the reference model, that is, a partial model matching method. Therefore, the first order delay +
A case where the control target 1 that can be approximated by the dead time system is controlled by the PID controller will be described with reference to FIG.

In FIG. 4, the control amount y with respect to the target value r (s)
The closed loop transfer function W (s) of (s) is expressed by the following equation.

By substituting the equations (1) and (2) into the equation (22), the following equation is obtained.

The ^dead time transfer function e- ^Ls can be expressed by the following equation when ^subjected to a ^macrolein expansion.

On the other hand, the transfer function W _r (s) of the reference model is given by the following equation.

Here, α _i : coefficient σ: time scale factor The target value r obtained by substituting equation (24) into equation (23)
Closed-loop transfer function W of control amount y (s) with respect to (s)
(S) and the transfer function W of the reference model shown in equation (25)
_{In order for r} (s) to partially match, the following equation must be satisfied.

From equation (26), the following equation is obtained.
The control parameters K _p , T _i , T _d and the time scale factor σ of the controller 2 can be determined.

By solving equation (30), the minimum true root is determined as the time scale factor σ, and according to equations (27), (28) and (29),
K _p , T _i , and T _d can be obtained. Alternatively, the reference model is the Kitamori model (α ₂ = 0.5, α ₃ = 0.15, α ₄ = 0.03, ...)
In the case of, the minimum positive real root of the equation (30) can be approximated by the following equation. Using the approximate value of σ, K _p , T _i , T _d can be calculated by the equations (27), (28), and (29). You can ask.

σ ≒ 1.37L (31) Next, an example of adjustment rule learning and an example of parameter adjustment will be described more specifically.

An adjustment rule learning system 4 constructs an adjustment rule (weight coefficient) of the parameter adjustment system 3 by learning.

The input / output pair (z, c) of the learning data is obtained by simulation offline. For example, a description will be given of the construction of an adjustment rule in a case where a control target that can be approximated by a first-order delay + dead time system is controlled by a PID controller.

FIG. 18 shows a functional configuration of one embodiment of the adjustment rule learning system 4.

The adjustment rule learning system 4 of the present embodiment has simulators 41 and 42 functioning as means for realizing the control target 1 and the controller 2 shown in FIG. 4 as a model.

The learning system 4 includes a simulator 41 and
42, a time-series signal capturing unit 43 that captures an operation amount and a control amount by simulation with respect to a target value and outputs it as a time-series signal, and an input for storing and extracting the time-series signal as needed and outputting the same. A time-series signal generating unit 44, a neural network 45 for estimating control parameters based on weighting coefficients for the time-series signal from the input time-series signal input unit 44, and a simulator.
Based on the characteristics (K, T, L) of the control target model realized by the 41 simulations, the teacher control parameters K _p ^* ,
Calculate T _i ^* , T _d ^* , store the results, and take them out as necessary. The teacher control parameter generator 46, the teacher control parameters and the neural network 45.
An error calculating unit 47a for obtaining an error from the parameter estimated by the above, and a weighting factor correcting unit for correcting the weighting factor based on the error and setting the weighting factor in the neural network 45
47.

In such an adjustment rule learning system 4, the simulation is performed using simulators 41 and 42 which are control system models in which a control target that can be approximated by a first-order lag + dead time system is controlled by a PID controller. At this time, the input signal is collected by the time-series signal acquisition unit 43 and the input time-series signal generation unit 44. That is, the time-series signal acquisition unit 43 variously changes the characteristics (K, T, L) of the model to be controlled and variously changes the characteristics (K _p , T _i , T _d ) of the controller model. Various time responses are obtained, and time series signals of the operation amount x and the control amount y are collected from these time responses as learning input data z.

Further, the teacher parameter generation unit 46 uses the parameters K, T, and L of the model to be controlled when the learning input data is obtained, and calculates the time scale and the time scale according to the equation (30) or (31), for example. The factor σ is found, and this σ is
7), (28), are substituted into (29), the control parameter K _p
^* , _Ti ^* , _Td ^* , and the control parameters _Kp ^* , _Ti ^* , _Td
^* Is the learning teacher data.

A set of input and output (z,
c) The weighting factor correction unit 47 uses the error back propagation learning algorithm to learn the neural network 45, and adjusts the adjustment function F (z) or the adjustment rule (weighting factor).
To build.

In the simulation, the characteristics (K, T, L) of the model to be controlled are variously changed, and the characteristics ( _Kp , _Ti , _Td ) of the model of the controller (simulator 42) are variously changed. When the time response is obtained, the test signal ts may be superimposed on the operation amount as shown in FIG. Various signals such as a pseudo-random signal, a sine wave signal, and a ramp signal can be used as the test signal ts.

Further, when variously changing the characteristics (K, T, L) of the model to be controlled and the characteristics (K _p , T _i , T _d ) of the controller, even if these characteristics are changed using uniform random numbers, Good.

Next, adjustment of the control parameters of the controller 2 by the parameter adjustment system 3 will be described.

As shown in FIG. 19, the parameter adjustment system 3
A time-series signal capturing unit 31 that captures an operation amount and a control amount with respect to a target value from a control target 1 and a controller 2 that constitute an actual control system and outputs the time-series signal; 4
Based on the adjustment rules (weighting factors) set by,
It has a neural network 32 for estimating control parameters ( _p , _i , _d ) as its main function.

The neural network 32 may be the same as the neural network 45 used in the adjustment rule learning system 4.

Parameter adjustment system 3 configured as described above
The time series signal acquisition unit 31 fetches a time series signal of input / output variables of the control target 1 of the real process online, and controls the controller 2 using the neural network 32 based on the time series signal. Adjust the parameters. For example, in the case of a PID controller, the control parameters are K _p , T _i , and T _d , and these are adjusted by using an adjustment rule (weight coefficient) constructed by offline learning.
Adjust by the network 32.

Next, an embodiment of an online adjustment rule learning method will be described.

In the above offline learning, a model of the control system is used,
An adjustment function F (z) or an adjustment rule (weight coefficient) is constructed using data obtained by the simulation. In this case, it is assumed that the characteristics of the control target are known to some extent. For example, the adjustment function F (z) is constructed such that the characteristics of the control target can be approximated by a first-order lag + dead time system.

For this reason, if the characteristics of the control target deviate from the assumed characteristics, the adjustment function also deviates, and it takes a little time to adjust to a desired control response.

To solve this problem, online learning of the adjustment function is required. This online learning method will be described below.

FIG. 20 shows a configuration of a parameter adjustment system with an online learning function, and FIG. 21 is a diagram for explaining an online tuning process.

The parameter adjustment system 5 with the online learning function shown in FIG. 20 is similar to the one shown in FIG. 19, except that the operation amount and the control amount and the target value A time-series signal capturing unit 51 that captures and outputs the time-series signal, and stores the time-series signal, and an input time-series signal generation unit 52 that outputs the time-series signal as necessary. A neural network 53 for estimating the control parameters ( _p , _i , _d ) based on the control parameters and a control parameter giving a desired response in the process of tuning the parameters of the controller are stored and output as teacher control parameters as necessary. Teacher control parameter generator 54
And an error calculating unit 55a for calculating an error between the teacher control parameter and the parameter estimated by the neural network 53; and a weighting factor for correcting the weighting factor based on the error and setting the weighting factor in the neural network 53 And a correction unit 55.

The system 5 has switches SW1 and SW2 for switching between a learning state and a tuning state. The switches SW1 and SW2 are in the online learning state when switched to the a side, and are in the online tuning state when switched to the b side.

The neural network 53 of this system 5 is learned off-line by using the simulators 41 and 42 as shown in FIG. This learning is
It may be performed by another system, or the system may be provided with a simulator.

 Next, the operation of this system will be described.

First, when tuning the controller 2, the switches SW1 and SW2 are switched to the b side, and the time series signal from the control system is passed through the time series signal acquisition unit 51.
The input is made to the neural network 53 online. In the neural network 53, the control parameters are estimated using the learning results up to that time. The result is sent to the controller 2 for tuning. The result is also sent to the teacher control parameter generator 54 at the same time, where it is temporarily stored.

Here, it is assumed that as a result of adjusting the control parameters based on the time series signal in the online tuning state, a tuning process as shown in FIG. 21 is obtained. FIG. 21 (a) shows a case where the initial control response has an overshoot, and FIG. 21 (b) shows a case where the initial control response is overdamping. Both because the characteristics of the controlled object is slightly shifted from the assumed characteristics, desirable control response (control parameter _{^{K p *, T i *,}} T d * to the corresponding) is taking time to adjust to .

The parameter adjustment system 5 with online learning function
The data of the tuning process is stored, and the adjustment function is corrected by online learning using the data.
That is, the control parameters _Kp ^* , _Ti ^* , and _Td ^* when the desired control response is obtained are used as teacher control parameters (teacher data for learning), and the input / output variables of the control system taken in the tuning process are obtained. Learning is performed using the time-series signal as an input time-series signal (input data for learning). At this time, the changeover switches SW1 and SW2 in FIG. 20 are in the state of the “a” side.

In the online learning in the above embodiment, the adjustment function F (z) or the adjustment rule (weighting factor) obtained by the offline learning is directly corrected by the data obtained in the online tuning process. .
On the other hand, as shown in FIG. 22, the adjustment function F obtained by the offline learning of the neural network 6a is used.
The correction function ΔF (z) of (z) is calculated by another neural network using the data obtained in the online tuning process.
It may be constructed using the network 6b.

That is, differences ΔK _p ^* , Δ between the control parameters K _p ^* , T _i ^* , T _d ^* when a desired control response is obtained and the control parameters _p , _i , _d estimated offline.
T _i ^* and ΔT _d ^* are used as learning teacher data, learning is performed using time series signals of input and output variables of the control system taken in the tuning process as learning input data, and a correction function ΔF is used.
Construct (z). In this case, in the online tuning after learning, a control parameter is obtained by a modified adjustment function F '(z) shown in the following equation.

F ′ (z) = F (z) + ΔF (z) When the characteristics of the control target greatly deviate from the assumed characteristics, the adjustment function F constructed by the offline learning is used.
Even if adjustment is performed using (z), it may be difficult to approach a desired control response. In this case, the control parameter K _p, T _i, the _{T d, ΔK p, ΔT i} , modifying by [Delta] T _d, using the hill-climbing method, gradually desired control response (control parameters
_Kp ^* , _Ti ^* , _Td ^* ).

Also in this case, the parameter adjustment system with online learning function 5 stores the data of this tuning process,
The adjustment function F (z) is corrected by online learning using these data. That, K _p ^* control parameters when desired control responses were ^obtained, T _i ^*, a T _d ^* and learning teacher data, the time series signal of the input and output variables of the control system taken-tuning process, Learning is performed as learning input data.

In addition, about the above-mentioned online tuning,
FIG. 23 shows the result of a simulation performed using a simulator as shown in FIG. 18 instead of the actual control system.

As shown in the figure, the online tuning of the above-described embodiment can be performed in any of (a) when the initial control response has a large overshoot, and (b) when the initial control response is in an overdamping state. After one or two trials, tuning has been achieved to achieve the desired response.

In each of the above embodiments and other embodiments to be described later, an example of a system configuration of hardware that can configure each system of the adjustment rule learning system 4, the parameter adjustment system 3, and the parameter adjustment system 5 with the online learning function will be described. See Figure 24.

The hardware system shown in FIG. 24 includes an arithmetic unit CPU that implements the functions of each system, a read-only memory ROM that stores fixed data such as programs and coefficients executed by the CPU, and various arithmetic results executed by the CPU. And a read / write memory RAM for temporarily storing data for constructing a neural network, an analog input unit AI and a digital input unit DI for taking in data from outside, and an analog output unit AO for outputting data to outside. It comprises a digital output unit DO and a communication interface CI / O for communicating with other systems.

The units such as the CPU, ROM, and RAM are connected by an internal bus BUS. Then, a signal from the control system is taken in via the analog input unit AI or the digital unit DI. The control parameters are output to the controller via the analog output unit AO or the digital output unit DO.

Although not shown, the hardware system includes an input device such as a keyboard for inputting instructions from an operator, a display for displaying processing details, a printer for printing out processing results, and the like as necessary. Connected.

The display can display, for example, a response waveform as shown in FIGS. 21 and 23, a graphic display of a learning state of the neural network, and the like.

Next, various embodiments to which the present invention can be applied other than the above-described embodiments will be described. In each of the embodiments described below,
Each of them can be applied independently, or they can be appropriately combined. Further, it can be combined with the embodiment described above.

The specific configuration and operation of each embodiment described later can be easily understood by referring to the description of the above-described embodiment, and thus detailed description is omitted to avoid duplication.

In the above embodiment, an example was described in which a PID controller having the transfer function shown in equation (1) was used as the controller. However, the present invention is not limited to this, and can be applied to a case where a controller of a different type is used. Hereinafter, these examples are given.

First, the present invention provides a PI which shows a transfer function G _c (s) in the following equation:
Applicable when using a (proportional / integral) controller.

Secondly, the present invention can be applied to the case where an I-PD controller or an IP controller which exhibits the characteristics in the following equation is used.

Here, x (s): operation quantity r (s): the target value y (s): controlled variable K _i: integral gain K _p: proportional gain K _d: differential gain Third, in the above embodiment, the case where the position type PID controller shown in the expression (1) is used as the controller 2 has been described. However, as shown in FIGS. It can also be applied when using a speed PI controller.

Fourth, in the above-described embodiment, a case has been described in which a continuous-time PID controller is used as the controller 2 as shown in the equation (1). The present invention is also applicable when a discrete-time controller such as a PI controller is used.

Where K _p : proportional gain K _i : integral gain K _d : differential gain u (t + 1) = u (t) + K _p [{r (t + 1) -y (t + 1)}-{ r (t) -y (t)}] + K _i {r (t + 1) -y (t + 1)} + K _d [{r (t + 1) -y (t + 1)}-2 {r (t) -y (t)} + {r (t-1) -y (t-1)}] (velocity type PID)… (36) u (t + 1) = u (t) + _Kp [{r (t + 1) -y (t + 1)}-{r (t) -y (t)}] + _Ki {r (t +1) -y (t + 1)} (velocity type PI) (38) That is, the present invention uses various types of controllers as the controller 2 without being limited to the continuous time position type PID controller. Also applicable to cases.

Next, in the above embodiment, the case where the transfer function of the controlled object can be approximated by a first-order lag + dead time system has been described. However, the present invention is not limited to this, and the controlled object having a different transfer function is not limited to this. Can also be applied. Hereinafter, these examples are given.

First, according to the present invention, the controlled object has a transfer function G
_The present invention can also be applied to a case where approximation can be made by a secondary delay + dead time system indicating _p (s).

Second, according to the present invention, the controlled object has a transfer function G
_{If p} (s) can be approximated by the integral + dead time system,
Alternatively, the present invention can also be applied to a case where approximation can be made by an integration + 1 order delay + dead time system.

That is, the present invention is not limited to the case where the transfer function of the controlled object can be approximated by a first-order lag + dead time system,
The present invention can also be applied to a case where various types of transfer functions can be used for approximation.

In the above embodiment, the case where the time series signal Z of the absolute value of the input / output variable to be controlled is used as the input signal to the neural network as shown in the equation (8) has been described. The present invention is not limited to this, and can be applied to a case where a time-series signal of another format is used.

First, the present invention can be applied to a case where a time-series signal corresponding to a change in an input / output variable to be controlled is used as shown in the following equation.

Secondly, the present invention can be applied to a case where a time series signal of a target value and an input / output variable to be controlled is used as an input signal to the neural network as shown in the following equation.

Third, the present invention can be applied to a case where a time series signal of a control deviation and an input / output variable to be controlled is used as an input signal to the neural network as shown in the following equation.

That is, the present invention is also applicable to a case where variables are appropriately selected from input / output variables of the control system as input signals to the neural network, and time series signals of their absolute values or changes are used.

Fourth, as an input signal to the neural network, the control parameters of the controller
The present invention can also be applied to a case where time series signals of K _p , T _i , T _d and input / output variables to be controlled are used.

Fifth, when input signals to the neural network are time-series signals of control parameters K _p , T _i , T _d , target values, and input / output variables to be controlled, as shown in the following equation. Also applicable to

Sixth, as shown in the following equation, the time series signal of the control parameters Kp, Ti, Td of the controller 2, the control deviation, and the input / output variable of the control object is used as the input signal to the neural network. Applicable.

That is, the present invention selects appropriate variables from input / output variables of the control system as input signals to the neural network, and calculates their absolute values or time-series signals of changes and control parameters of the controller 2. It can also be applied when used.

In the above embodiment, the case where the absolute values of the control parameters K _p , T _i , and T _d of the controller 2 are used as the output signal of the neural network as shown in Expression (9) has been described. However, the present invention is not limited to this, and can be applied to other output formats. Hereinafter, these examples are given.

First, the present invention provides a controller 2 as shown in the following equation:
Can be applied to the case where the variation of the control parameters K _p , T _i , and T _d is used.

In this case, the control parameters K _p , T _i , and T _d are obtained by the following equations.

K _p (t) = K _p (t−1) + ΔK _p T _i (t) = T _i (t−1) + ΔT _i (49) T _d (t) = T _i (t−1) + ΔT _d 7, the present invention is also applicable to the case where the correction ratios C _p , C _i , and C _d of the control parameters K _p , T _i , and T _d of the controller 2 are used as the output signals of the neural network as shown in the following equation. Applicable.

In this case, the control parameters K _p , T _i , and T _d are obtained by the following equations.

K _p (t) = C _p · K _p (t−1) T _i (t) = C _i · T _i (t−1) (51) T _d (t) = C _d · T _d (t− 1) In the above embodiment, the case where the control system is in a closed loop as shown in FIG. 3 has been described.
In the parameter adjustment in the present invention, as shown in FIG. 8, the control system can be brought into an open loop state by opening the switch 2a.

In this case, as the input / output signal to the neural network, as shown in the equation (8) or (42), a time series signal of the absolute value or the change of the input / output variable of the control target 1 is used. Thus, the control parameters can be adjusted even during the test operation of the process.

When the control system is an open loop, the learning input data z is obtained by simulation using an open loop control system model shown in FIG. Similarly to the closed loop, the learning teacher data includes the parameters K, T, and K of the control target 1 when the learning input data z is obtained.
Using L, the time scale factor σ is obtained by the equation (30) or (31), and this σ is further converted to the equation (27),
(28), (29) are substituted into equation control parameter K _p, T _i, determine the T _d, and the control parameter K _p, T _i, the teacher data for learning the T _d.

In the above embodiment, the case where the test signal t _s is superimposed on the manipulated variable as shown in FIG. 5 when learning input data is obtained by simulation has been described. However, the present invention is shown in FIGS. 10 and 11. As described above, the present invention can be applied to a case where the test signal t _s is superimposed on the target value or the control amount. As the test signal t _s, the pseudo-random signal, a sine wave signal, a step signal, various signals such as the ramp signal can be utilized.

In the above embodiment, the case where the operation data of the controlled object 1 is used as shown in FIG. 3 when adjusting the control parameters of the controller 2 has been described. As shown in the above, the present invention can also be applied to a case where the test signal t _s is superimposed on an operation amount, a target value, or a control amount. The test signal includes a pseudo random signal,
Various signals such as a sine wave signal and a step signal can be used.

In the above embodiment, when learning the adjustment rule of the control parameter, the case where the partial model matching method is used for creating the teacher data and the Kitamori model is used as the reference model has been mainly described. The present invention is not limited to the Kitamori model as a reference model, and can be applied to a case where other various reference models are used. FIG. 15 shows an example of another reference model.

In the above-described embodiment, the case where the partial model matching method is used for creating the teacher data when learning the adjustment rule of the control parameter has been described. However, the present invention is not limited to this method, for example,
(1) Ziegler-Nichols method, (2) Chien-Hrones-Res
The present invention can also be applied to the case where various adjustment methods such as the wick method and the like (refer to Masami Masubuchi, "Revised Basic Theory of Automatic Control", Corona Corporation, June 1977) are used.

In the above embodiment, the case where one kind of the PID controller shown in the expression (1) is used as the controller 2 has been described. However, according to the present invention, various controllers such as a PI controller, an I-PD controller, and an IP controller and adjustment rules of these controllers are prepared, and one controller and a corresponding adjustment rule are prepared from these controllers and the adjustment rules. It can also be applied when selecting and using rules.

At this time, there are (1) a method of preparing adjustment rules of various controllers as separate weight coefficients corresponding to the controllers, and (2) a method of adjusting the adjustment rules of various controllers by one kind of weight coefficient. In the latter case,
As an input signal to the neural network, as shown in the following equation, using the time-series signal of the variable C _T and the control target of the input and output variables indicating the type of controller.

In the above embodiment, when learning the adjustment rule of the control parameter, the partial model
Using the matching method, as the reference model, the overshoot amount and the damping ratio, which are the characteristic amounts of the response waveform, are fixed.
The case where the Kitamori model is used has been mainly described. However, the present invention is not limited to this, and uses a reference model in which the characteristic amount of the response waveform such as the amount of overshoot and the damping ratio is variable,
The present invention is also applicable to a case where a characteristic amount of a response waveform such as an overshoot amount and a damping ratio can be designated in adjusting a control parameter.

At this time, (1) a method of preparing controller adjustment rules as different weighting factors corresponding to various overshoot amounts, damping ratios, and the like; (2) a method of preparing controller adjustment rules with various overshoot amounts, damping ratios, and the like. For example, there is a method of using one type of weight coefficient. In the latter method, as input signals to the neural network, variables O and A that specify the amount of overshoot, damping ratio, etc.
And time-series signals of input / output variables to be controlled.

In the above embodiment, the one-input one-output system in which the control target 1 can be approximated by a first-order lag + dead time system has been described. However, the present invention can also be applied to a case where the control target 1 is a multi-input multi-output system. In this case, adjustment rules are prepared corresponding to controllers of a multi-input multi-output system such as various chemical plants and boilers, and an appropriate adjustment rule is selected from these adjustment rules.

In the above-described embodiment, the case where learning input data is obtained by simulation has been described. However, the present invention provides a time response (operation data) of a plant control system having various characteristics.
The time response of the good characteristic is recorded, and the time series signal of the input / output variable of the control system obtained from the time response is used as the input data for learning, and the control parameter corresponding to the time response of the good characteristic Is recorded, and this parameter is used as training data for learning, so that the present invention can be applied to a case where a neural network is trained.

In the above embodiment, the time series signals of the input / output variables of the control system are fetched, and the parameters of the controller are adjusted by the neural network based on these signals. In the following, an embodiment will be described in which a feature amount of a control system is extracted and parameters of the controller are adjusted by a neural network based on these signals.

The basic configuration of this embodiment is the same as that shown in FIG. 1 described above, but differs in the internal functions of the adjustment rule learning system 4 and the parameter adjustment system 3. Therefore, the present embodiment will be described focusing on this difference.

FIG. 25 shows an example of a functional configuration of the adjustment rule learning system 4 used in the present embodiment. As a hardware system for constructing such a system including the system shown in FIG. 27 described later, for example, the same system as that shown in FIG. 24 can be used.

The adjustment rule learning system 4 of the present embodiment has simulators 41 and 42 modeled on the control target 1 and the controller 2 shown in FIG. The models of these simulators 41 and 42 are not limited to those shown in FIG. 4, but may be other different models described above.

In addition, the learning system 4 uses these simulators.
From 41 and 42, an operation amount and a control amount with respect to a target value are taken in, a feature amount extraction unit 48 for extracting a feature amount about them, and the extracted feature amounts are stored and extracted and output as necessary. A neural network 45 including a feature storage / extraction unit 49 for estimating a control parameter of the feature from the feature storage / extraction unit 49 based on a weighting coefficient; ) Are calculated on the basis of the characteristics (K, T, L) of the teacher parameters _Kp ^* , _Ti ^* , _Td ^* , and the results are stored and taken out if necessary. And an error calculator 47a for calculating an error between the teacher control parameter and the parameter estimated by the neural network 45. A weighting coefficient is corrected based on the error. And a weighting coefficient modification unit 47 sets the Raru network 45.

This embodiment has the same configuration as the learning system 4 shown in FIG. 18 except that the input signal to the neural network 45 is not a time-series signal of input / output variables but a feature amount. . Therefore, the learning operation of the adjustment rule is the same as that shown in FIG.
The same applies to the modification examples such as the superposition of t _s , and the description will not be repeated here.

The characteristic amounts used here include, for example, an overshoot amount E, a damping ratio D and a settling time ratio R as shown in FIG.

Next, FIG. 27 shows an example of a functional configuration of the parameter adjustment system 3 used in the present embodiment.

As shown in FIG. 27, the parameter adjustment system 3
A feature amount extraction unit 33 that takes in an operation amount and a control amount with respect to a target value from a control target 1 and a controller 2 that constitute an actual control system and outputs them as a feature amount, and sets the feature amount by the adjustment rule learning system 4 Neural network for estimating control parameters ( _p , _i , _d ) based on the adjusted rules (weighting factors)
32 as a main function.

The neural network 32 may be the same as the neural network 45 used in the adjustment rule learning system 4.

The present embodiment has the same configuration as the parameter adjustment system 3 shown in FIG. 19 described above, except that the input signal to the neural network 32 is not a time-series signal of input / output variables but a feature amount. is there.

The parameter adjustment operation is the same as that in FIG. 19 except for using the feature quantity, and the modification can be similarly applied to the corresponding portions.

The feature amounts handled by the feature amount extraction unit 33 are the same as those in the embodiment of FIG. 25 described above, and for example, those shown in FIG. 26 are used.

Next, FIG. 28 shows a system that includes a system similar to the above-described parameter adjustment system with online learning function shown in FIG. 20 and adjusts control parameters based on actual process data and learns adjustment rules. An example will be described.

In this embodiment, as shown in FIG. 28, a parameter adjustment system 5 with an online learning function is provided for the control target 1 and the controller 2 of the actual process.

The parameter adjustment system 5 with an online learning function shown in FIG. 28 fetches an operation amount and a control amount with respect to a target value from the control target 1 and the controller 2 which are actual control systems, similarly to the system shown in FIG. A feature value extracting unit 56 for extracting the feature values, and a feature value storing / extracting unit for storing the feature values and outputting the feature values as necessary.
57, a neural network 53 for estimating a control parameter ( _p , _i , _d ) based on a weighting factor for the feature quantity, and a control parameter that gives a desired response in a process of tuning the parameter of the controller 2; A teacher control parameter generation unit 54 that outputs as a teacher control parameter as necessary, and an error calculation unit 55a that calculates an error between the teacher control parameter and the parameter estimated by the neural network 53.
And a weight coefficient correction unit 55 for correcting the weight coefficient based on the error and setting the weight coefficient in the neural network 53.

The system 5 has switches SW1 and SW2 for switching between a learning state and a tuning state. The switches SW1 and SW2 are in the online learning state when switched to the a side, and are in the online tuning state when switched to the b side.

The neural network 53 of the system 5 is learned offline using the simulators 41 and 42 as shown in FIG. 25 described above. This learning is
It may be performed by another system, or the system may be provided with a simulator.

This embodiment is the same as the system shown in FIG. 20 except that the feature amount is used for adjusting the control parameters. It is similar to that shown in the figure.
Therefore, the modified example of FIG. 20 can be similarly applied to the corresponding portions.

Next, an embodiment in which the present invention is applied to process control will be described with reference to FIG.

Note that, as described above, the present invention can be widely applied to controlled objects and controllers exhibiting various transfer functions, and the example shown in FIG. 29 is only an example. For example, the present invention can be applied to the control of a power plant, a manufacturing plant, a chemical plant, and other various automatic driving devices.

FIG. 29 is a block diagram conceptually showing a control system for controlling the steam temperature of the boiler.

The embodiment shown in FIG. 1 includes a control target 100, a steam temperature controller 200 for controlling the control target 100, a parameter adjustment system 3 for adjusting the control parameters of the controller 200, an adjustment rule learning system 4, a steam temperature sensor 12
0 is provided.

The control target 100 includes a boiler 111, a fuel flow control valve 112 for controlling a fuel flow to a burner portion of the boiler 111,
A flow meter 113 for detecting the fuel flow and a fuel flow controller 114 for controlling the opening of the fuel flow control valve 112 are arranged.

The steam temperature controller 200 executes the PID control as described above. The steam temperature controller 200 executes a PID control calculation on a deviation between the steam temperature target value and the steam temperature (control amount) from the sensor 120 according to a control parameter, and Outputs the required flow rate (operating variable).

This required fuel flow value is sent to the fuel flow controller 114. The fuel flow controller 114 calculates and outputs an opening command to the fuel flow control valve 112 based on the fuel flow from the flow meter 113 and the required fuel flow value. In response to this, the opening of the fuel control valve 112 is changed, and the fuel flow rate is adjusted. Accordingly, the combustion state of the burner of the boiler 111 changes, and the steam temperature changes.

Further, the steam temperature (control amount) and the required fuel flow rate (operating amount) are input to the parameter adjustment system 3 and, as described above, as a time-series signal or by extracting a characteristic amount, It is input to the network and the control parameters are estimated by the neural network. This control parameter is controlled by the steam temperature controller 20
Sent to 0. Therefore, the control parameters of the steam temperature controller 200 are adjusted online.

The adjustment rule (weight coefficient) used in the parameter adjustment system 3 is determined by the adjustment rule learning system 4 by offline learning using a neural network as described above.

As described above, according to the present embodiment, the control parameters of the steam temperature controller 200 can be adjusted on-line in accordance with the operation amount and the control amount, so that optimal control can be performed in response to a change in the control system. As described above, the adjustment rule used in the parameter adjustment system 3 can be automatically executed by the adjustment rule learning system 4 using a neural network. Therefore,
Unlike a conventional system, there is no need to manually and extensively create an adjustment rule.

In the embodiment shown in FIG. 29, the parameter adjustment system 3 and the adjustment rule learning system 4 are used. However, the parameter adjustment system 4 with the online learning function shown in FIG. 20 or FIG. May be adjusted and learning of adjustment rules may be performed.

As described above, according to each embodiment of the present invention, the time response of the control system model is obtained using the models of the control target having various characteristics, and the input / output of the control system model obtained from the time response is obtained. Learning the neural network by using the time-series signal or feature value of the variable as input data for learning, finding the optimal control parameters corresponding to the characteristics of the model to be controlled, and using this as training data for learning. Can be.

Therefore, the construction of the adjustment rule can be achieved automatically and in a short time.

Further, according to each embodiment of the present invention, a time series signal or a feature amount of an input / output variable of a control system is acquired, and a control parameter of a controller is adjusted by a neural network based on these signals. Changes, the change can be detected as a change in the time-series signal or characteristic value of the input / output variables, and can be adapted to process characteristic changes without being limited by the time response of the control system input / output variables. Thus, the parameters of the controller can be adjusted online, and the characteristics of the control system can be maintained in a good state.

Further, according to the embodiment capable of performing online learning of adjustment rules, in an actual control system, a control parameter in a case where appropriate control can be performed is used as teacher data, and a time-series signal of actual input / output variables at that time or The neural network can be trained with the input data for learning the feature amount. Therefore, even when the characteristics of the control target deviate from the assumed characteristics of the model, the parameters can be adjusted appropriately and promptly.

[Effects of the Invention] As described above, the present invention has the effect of shortening the construction time of the adjustment rule and adjusting the control parameters without restricting the time response of input / output variables of the control system. is there.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a control system to which the control parameter adjusting method of the present invention is applied, FIG. 2 is an explanatory diagram schematically showing the configuration of a multilayer neural network, and FIG. FIG. 4 is a block diagram schematically showing a functional configuration of a unit having an intermediate layer in the neural network, FIG. 4 is a block diagram showing a configuration of a control system model to which the above-described embodiment is applied, and FIG. 6 and 7 are block diagrams showing examples of other control system models to which the present invention is applied, and FIG. 8 is a block diagram showing a state in which a test signal is superimposed on the operation amount of the control system. FIG. 9 is a block diagram showing the configuration of an embodiment in which the control system is in an open loop, FIG. 9 is a block diagram showing the configuration of a control system model in an open loop, and FIG. Show FIG. 11 is a block diagram showing a state in which a signal is superimposed on a test on a target value of a control system model. FIG. 11 is a block diagram showing a state in which a signal is superimposed on a test on a control amount of the control system shown in FIG. FIG. 13, FIG. 13 and FIG. 14 are block diagrams each showing an embodiment of a system for superimposing test signals, FIG. 15 is a table showing an example of a reference model for creating teacher data in the present invention, FIG. Is a block diagram showing adjustment of control parameters using a conventional auto tuner, and FIG.
FIG. 17 is a waveform diagram showing a response waveform of a feature amount when the target value is step-changed, FIG. 18 is a block diagram showing a configuration of one embodiment of an adjustment rule learning system used in the present invention, and FIG. FIG. 20 is a block diagram illustrating a configuration of an embodiment of a parameter adjustment system used in the present invention. FIG. 20 is a block diagram illustrating a configuration of an embodiment of a parameter adjustment system with an online learning function used in the present invention. Waveform diagram showing the online tuning process. Fig. 22 is a block diagram showing an example of creating a correction function for the adjustment function obtained in the offline learning using data obtained in the online tuning process using another neural network. FIG. 23 is a waveform diagram showing the results of a simulation performed using the simulator of the adjustment rule learning system, and FIG. FIG. 25 is a block diagram showing a system configuration of hardware capable of configuring the system of each embodiment of the present invention. FIG. 25 is a functional configuration diagram of an adjustment rule learning system used when adjusting a control parameter using a feature quantity. FIG. 26 is a block diagram showing an example, FIG. 26 is a waveform diagram showing an example of a response waveform used for extracting a feature used in the adjustment rule learning system, and FIG. 27 is a diagram showing a case where control parameters are adjusted using the feature. FIG. 28 is a block diagram showing an example of a functional configuration of a parameter adjustment system used in the present invention. FIG. 28 is a block diagram showing a configuration of another embodiment of a parameter adjustment system with an online learning function used in the present invention. FIG. 4 is a block diagram showing a configuration of an embodiment in which is applied to process control. 1, 100 controlled objects, 2 controllers, 3 parameter adjustment systems, 4 adjustment rule learning systems, 5 parameter adjustment systems with online learning functions, 31, 43, 51 ... time-series signal acquisition Department, 32, 45, 53 ...
... Neural network, 33, 48, 56 ... Feature extraction unit, 41, 42 ... Simulator, 44, 52 ... Input time-series signal generation unit, 46, 54 ... Teacher control parameter generation unit, 47 55, a weight coefficient correction unit, 49, 57, a feature amount storage / extraction unit, 111, a boiler, 112, a fuel flow control valve,
113… Flow meter, 114… Fuel flow controller, 120…
... Steam temperature sensor, 200 ... Steam temperature controller.

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Makoto Shimoda 4026 Kuji-cho, Hitachi City, Ibaraki Prefecture Inside Hitachi, Ltd.Hitachi Research Laboratory (72) Inventor Masakazu Kondo 4026 Kuji-cho, Hitachi City, Ibaraki Prefecture Hitachi Research, Ltd. In-house (56) References JP-A-1-116869 (JP, A) JP-A-3-118606 (JP, A) JP-A-3-201008 (JP, A) JP-A-3-77101 (JP, A) Masatoshi Tokita and 3 others, "Robot Force Control by Neural Network Model (2nd Report: Adaptive Control by Perceptron)", Proceedings of the 6th Annual Conference of the Robotics Society of Japan, Robotics Society of Japan, October 1988 March 20, p.143-144 Masahide Nomura and three others, "How to tune a PI controller using partial knowledge of the step response waveform." Law, 15th System Symposium, 10th Knowledge Engineering Symposium, Joint Symposium Materials, The Society of Instrument and Control Engineers, October 19, 1989, p. 251-256 (58) Fields surveyed (Int.Cl. ^6, DB name) G05B 13/00 - 13/04 G06F 15/18 550 JICST file (JOIS)

Claims (7)

(57) [Claims] 1. A method for adjusting a controller for controlling a control target so that a control amount matches a target value, comprising: a time-series signal or a feature amount of input / output variables of a control system in which the controller and the control target are combined; And a controller parameter newly determined by a neural network based on these signals. 2. A method for adjusting a controller for controlling a control target so that a control amount matches a target value, comprising: a time-series signal or characteristic amount of input / output variables of a control system in which the controller and the control target are combined; And a control parameter of the controller is determined by a neural network based on these signals. 3. A method of adjusting a controller for controlling a control target so that a control amount matches a target value, the control method comprising: controlling a control system in which a controller and a control target are combined; A method for adjusting a controller, comprising taking variables specifying the desired response of the system and determining control parameters of the controller by a neural network based on these signals. 4. A method for adjusting a controller for controlling a control target so that a control amount matches a target value, wherein a time response is obtained for various characteristics of the control system model using a control system model. A time series signal or characteristic amount of an input / output variable of the control system model obtained from the control parameter of the controller at that time, a variable indicating the type of the controller, and a variable specifying a target response of the control system Is used as learning input data, an optimum control parameter corresponding to the characteristics of the control system model is obtained, and the neural network is trained by using the parameter as learning training data. The data corresponding to the input data for learning is input to the neural network that has been Adjusting method of a controller and determining the meter. 5. A controller adjustment system for controlling a control target so that a control amount matches a target value, comprising: a neural network; and a time-series signal of input / output variables of a control system in which the controller and the control target are combined. Or a means for taking in a feature quantity, the taken-in time-series signal or feature quantity, a control parameter of the controller at that time, a variable indicating the type of the controller, and a variable designating a target response of the control target. Means for inputting any one of the following to the neural network, wherein the neural network determines a control parameter of the controller based on the input signal. 6. A neural network used for adjusting control parameters of a controller of a control system,
A system for learning adjustment rules, a means for realizing a control system model, a time response for various characteristics of the control system model, and a time series of input / output variables of the control system model obtained from the time response. A signal or a feature quantity, a control parameter of the controller at that time, a variable indicating the type of the controller, and a variable that specifies one of the responses to be controlled by the control target, A means for obtaining an optimal control parameter corresponding to the characteristics of the model of the control system and using it as learning teacher data; and a means for learning a neural network using the learning input data and the learning teacher data. An adjustment rule learning system characterized by the following. 7. A controller for controlling an object to be controlled in accordance with a control parameter, comprising: a time-series signal or characteristic amount of an operation amount and a control amount of the object to be controlled; a control parameter of the controller at that time; and a type of the controller. , And
A parameter adjustment system that fetches any of the variables that specify the desired response of the controlled object and executes the adjustment of the control parameters using a neural network; and realizes a model of the controlled object and the controller by simulation. A time response is obtained for various characteristics of the control system model, and a time-series signal or feature amount of an operation amount and a control amount of the control system model obtained from the time response,
Learning the neural network so that the control parameters of the controller at that time, a variable indicating the type of the controller, and a variable designating a target response of the control object are optimal control parameters. An adjustment rule learning system that sets an adjustment rule of the parameter adjustment system.