JP3039154B2 - How to adjust the controller - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a controller adjusting method, and more particularly to a controller adjusting method suitable for adjusting a control parameter based on a time response of an input / output variable of a control system.

[0002]

2. Description of the Related Art When a control target is controlled by a controller, it is necessary to adjust the control parameters of the controller according to the characteristics of the control target. As one of the methods, “adjustment method and adjustment system of controller”
(JP-A-3-233702). Hereinafter, an outline of the conventional adjustment method will be described.

In the conventional adjusting method, a control parameter of a controller is determined by a neural network based on a time series signal of input / output variables of a control system in which a controller and a control target are combined.

[0004]

In the above-mentioned prior art, it is difficult to adjust with a single neural network when responding to various control objects or when the characteristics of the control object greatly change. If multiple types of neural networks must be prepared for various types of control objects, or if the characteristics of the control
There was a problem that the number of networks increased significantly.

According to the present invention, the number of adjustment neural networks to be prepared can be minimized and the control parameters can be adjusted well even when various control objects are supported or when the characteristic change of the control object is particularly large. It is intended to provide a method.

[0006]

In order to achieve the above object, a time series signal of input / output variables of a controlled object or a control system is fetched, and these signals are timed with reference to the dead time of the controlled object or the control system. Normalize to the direction of the axis,
A control parameter of the controller is adjusted by a neural network on the basis of the normalized time series signal based on the normalized time series signal based on the fluctuation width of the time response of the control object or the control system.

[0007]

The present invention takes in time-series signals of input / output variables of a controlled object or a control system, normalizes these signals in the direction of the time axis based on the dead time of the controlled object or the control system. Is normalized in the amplitude direction based on the fluctuation width of the time response, and the control parameters of the controller are adjusted by the neural network based on the normalized time series signal. It is possible to cope with a case where the characteristic change is large. The number of prepared neural networks for adjustment can be minimized, and the control parameters can be satisfactorily adjusted.

[0008]

FIG. 1 shows an embodiment of the present invention.

This embodiment is a time series signal normalization process for taking in and normalizing a time series signal of an input / output variable of a control system in which the controller 2 controls the control target 1 and the control target 1 and the controller 2 in combination. The system comprises a system 3, a parameter adjustment system 4 for adjusting the control parameters of the controller 2 based on the time series signals of the normalized input / output variables and the adjustment rules, and an adjustment rule learning system 5 for constructing the adjustment rules of the parameter adjustment system 4 by learning. You.

The case where a PID (proportional / integral / differential) controller is used as the controller 2 will be described. The transfer function G _c (s) of the PID controller is given by Equation 1.

[0011]

(Equation 1)

Here, K _p : proportional gain T _i : integration time T _d : differentiation time Further, the transfer function G _p (s) as the control target 1 can be approximated by a first-order lag + dead time system shown in Expression 2. The case will be described.

[0013]

(Equation 2)

Here, K: gain T: time constant L: dead time The time-series signal normalization processing system 3 takes in a time-series signal of input / output variables of a control system in which the control target 1 and the controller 2 are combined. These signals are normalized based on the dead time L of the control target and the fluctuation width of the time response of the control system as shown in Expression 3.

[0015]

(Equation 3)

Here, Z: time series signal of input / output variables Z _N : time series signal of normalized input / output variables f _N : time series signal Z of normalized input / output variables and normalized input / output The time series signal Z _N of the variable is expressed by Equations 4 and 5, respectively.

[0017]

(Equation 4)

Here, r (tl): target value at the time of sampling (tl) y (tl): control amount at the time of sampling (tl) M: order

[0019]

(Equation 5)

[0020] _{Here, r N (t N -l)} :( t N -1) sampling points of the normalized target value _{y N (t N -l) :(} t N -l) normalized controlled variable sampling time t _N -l: normalized sampling point M _N: degree Note normalized, the normalization process of the time series signal will be described later in detail.

The parameter adjustment system 4 adjusts the control parameters of the controller 2 based on the time series signal Z _N of the input / output variables normalized by the time series signal normalization processing system 3. This adjustment process is represented by Equation 6.

[0022]

(Equation 6)

Here, C: control parameter F: adjustment function The control parameter C is expressed by the following equation (7) in the case of a PID controller.
Becomes

[0024]

(Equation 7)

The adjustment function F (Z _N ) is constituted by a multilayer (m-layer) neural network shown in FIG.

FIG. 3 shows the configuration of a unit which is a component of this neural network.

The input / output relationship of this unit is expressed by the following equations (8) and (9).
Is represented by

[0028]

(Equation 8)

[0029]

(Equation 9)

Here, u _j (k): the sum of inputs to the j-th unit in the k-th layer v _j (k): the output of the j-th unit in the k-th layer w _ij (k-1, k): (K-1) weighting factor of the connection from the i-th unit of the layer to the j-th unit of the k-th layer f: a function that gives the input / output relationship of each unit The first layer of the neural network is an input layer,
The output of the first layer unit is the input signal to the neural network. In the embodiment of the present invention, the input signal to the neural network is a time series signal Z _N of the normalized input / output variables of the controlled object 1, and the correspondence is shown in Expression 10.

[0031]

(Equation 10)

The last layer (the m-th layer in the embodiment of the present invention) of the neural network is an output layer, and the output of a unit in this layer is an output signal of the neural network. In an embodiment of the present invention, a neural network
The output signal of the network is a control parameter C of the controller 2, and the correspondence is shown in Expression 11.

[0033]

[Equation 11]

The adjustment function F (Z _N ) shown in Equation 6 changes as the input / output relationship of the units shown in Equations 8 and 9 changes. That is, when the number of layers in 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 _N ) 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 relationship of each unit, the adjustment function F ( Z _N ) can be constructed.

The adjustment rule learning system 5 constructs the adjustment rules of the parameter adjustment system 4 by learning. Next, the learning algorithm will be described.

First, an input / output pair (Z
Given _N , C), the square of the error shown in Equation 12 is defined as a loss function R.

[0037]

(Equation 12)

Here, w: a sum of all weighting factors of the connection of the neural network v _j (m) (w, Z _N ): an m-th layer obtained comprehensively from the input Z _N and the weighting factor w The correction amount Δw of the output w of the j-th unit in the output layer) is the gradient (gr) of the loss function R with respect to w.
adient) and is expressed by Equation 13.

[0039]

(Equation 13)

Here, Δw _ij (k−1, k): (k−
1) Correction amount of weighting coefficient w _ij (k-1, k) for coupling from the i-th unit of the layer to the j-th unit of the k-th layer ε: positive constant ∂R / ∂u _j (k) = _dj (k), Equation 13 becomes Equation 1
4, Equation 15 and Equation 16.

[0041]

[Equation 14]

[0042]

(Equation 15)

[0043]

(Equation 16)

Here, f ′: using the derivatives of the function f that gives the input / output relationship of each unit, Expressions 14, 15, and 16, gives the weighting coefficient w of the combination.
The modification of _ij (k-1, k) is calculated recursively from k = m to k = 2. That is, an error between the ideal output C (j) in the output layer and the actual output v _j (m) (w, Z _N ) is input and the direction from the output layer to the input layer is opposite to the direction of signal propagation. _Is propagated while taking the sum weighted by w _il (k, k + 1). This is the error back propagation learning algorithm.

It is assumed that a function f for giving an input / output relation of each unit is common to all units and is expressed by Expression 17.

[0046]

[Equation 17]

From Expression 17, Expression 18 is obtained.

[0048]

(Equation 18)

Expression 19 is derived from Expressions 8 and 18.

[0050]

[Equation 19]

In order to make learning converge smoothly and quickly, Equation 14 can be modified as Equation 20.

[0052]

(Equation 20)

Here, α: a positive constant (α = 1−ε may be used) τ: the number of corrections The input Z _N is a learning input / output pair (Z _N , C) of learning data input / output. And the output C is referred to as learning teacher data. 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, they are adjusted according to a certain evaluation criterion. As the evaluation criterion, there is a reference model, and there is an adjustment method for partially matching the transfer function of the control system with the transfer function of the reference model, that is, a partial model matching method. A case where the control target 1 that can be approximated by the first-order delay + dead time system is controlled by the PID controller will be described with reference to FIG.

In FIG. 4, the closed-loop transfer function W (s) of the control amount y (s) with respect to the target value r (s) is expressed by Expression 21.

[0056]

(Equation 21)

By substituting equations (1) and (2) into equation (21), equation (22) is obtained.

[0058]

(Equation 22)

Dead time transfer function

[0060]

[Outside 1]

Is expressed by the following equation (23) when it is subjected to the macro-Lin expansion.

[0062]

(Equation 23)

On the other hand, the transfer function W _r (s) of the reference model is
It is given by Equation 24.

[0064]

(Equation 24)

Here, α _i : coefficient σ: time scale factor The closed-loop transfer function W (s) and the number of the control amount y (s) with respect to the target value r (s) obtained by substituting the equation 23 into the equation 22 2
In order for the transfer function W _r (s) of the reference model shown in FIG. 4 to partially match, Equation 25 must be satisfied.

[0066]

(Equation 25)

From Equation 25, Equations 26, 27, 28, and 29 are obtained, and the control parameters K _p , T _i , T _d and the time scale factor σ of the PID controller can be determined from these equations.

[0068]

(Equation 26)

[0069]

[Equation 27]

[0070]

[Equation 28]

[0071]

(Equation 29)

By solving Equation 29, the minimum true root is determined as the time scale factor σ, and K _p , T _i , and T _d can be obtained from Equations 26, 27, and 28. Alternatively, the reference model is a Kitamori model (α ₂ = 0.5, α ₃ = 0.5.
15, α ₄ = 0.03,...), The minimum positive real root of Expression 29 can be approximated by Expression 30, and using the approximate value of σ, Expression 2
K _p , T _i , and T _d can be obtained from 6, 27, and 28.

[0073]

[Equation 30]

Set of input / output of learning data (Z _N , C)
Is obtained by simulation. Again,
An example will be described in which the control system shown in FIG. 4, that is, a control target 1 that can be approximated by a first-order delay + dead time is controlled by a PID controller. The simulation is performed using a model of the control system, and variously changes the characteristics (K, T, L) of the controlled object 1 and the characteristics (K _p , T _i , T) of the controller.
_d) a with various changes, determine the various time response, collecting a normalized time series signal of the target value r and a controlled variable y from these time response as the learning input data Z _N. In addition, the parameter K of the control target 1 when the learning input data Z _N is obtained,
Using T and L, a time scale factor σ is obtained by Expression 29 or Expression 30, and this σ is further calculated by Expression 26 or Expression 2
7, substituting into Equation 28 for the control parameters K _p , T _i , T _d
And the control parameters K _p , T _i , and T _d are used as learning teacher data C.

Using the input / output pair (Z _N , C) of the learning data obtained by the above method, the neural network is learned by the error back propagation learning algorithm described above, and the adjustment function F (z) To build.

The characteristics (K, T, L) of the control target 1 are variously changed, and the characteristics (K _p , T
_When various time responses are obtained by variously changing _i , T _d ), as shown in FIG. 5, the test signal t _S may be superimposed on the target value, or the target value may be directly changed. Is also good.

As a test signal, a pseudo random signal,
Various signals such as a sine wave signal, a step signal, and a ramp signal can be used. Further, the characteristic control target 1 (K, T, L) and the controller 2 of the characteristic _{(K p, T i, T} d) if variously changed, it may change the values of these characteristics at equal intervals, These characteristic values may be changed using uniform random numbers.

Next, normalization of a time-series signal will be described. First, as for the reference model, when the time scale factor σ changes in Equation 24, the shape of the time response of the output to the input expands and contracts in the direction of the time axis. That is,
When the time scale factor σ doubles, the shape of the time response of the output to the input doubles in the direction of the time axis,
The rise time is also doubled. Conversely, when the time scale factor σ is halved, the shape of the time response of the output to the input shrinks by half in the direction of the time axis, and the rise time is also halved.

When the reference model is the Kitamori model, the time scale factor σ is
It can be approximated by Equation 30. By substituting Equation 30 into Equation 24, Equation 31 is obtained.

[0080]

(Equation 31)

The equation (31) is based on the control parameters K _p and T so that the transfer function of the control system partially matches the Kitamori model.
The transfer function of the control system when _i and _Td are adjusted is approximately represented, and is a function of only the dead time L. Therefore, the shape of the time response of the output with respect to the input of the control system expands and contracts in the direction of the time axis corresponding to only the dead time L which is the characteristic parameter of the control target 1. That is, if the dead time L is doubled, the shape of the time response of the output to the input is doubled in the direction of the time axis, and the rise time is also doubled. Conversely, when the dead time L is halved, the shape of the time response of the output to the input shrinks by half in the direction of the time axis,
The rise time is halved.

From the above, the control parameters K _p and K _p are adjusted so that the transfer function of the control system partially matches the Kitamori model.
When T _i and T _d are adjusted, the time series signal of the input / output variable of the control system is expressed by the integral time of the dead time L as shown in Expression 32.
, It can be seen that the time response of the output with respect to the input has almost the same shape.

[0083]

(Equation 32)

Here, Δt: sampling period N: integer That is, as described above, in the control system approximated by the equation 31, the shape of the time response of the output to the input is determined by the characteristic parameter of the controlled object 1. Expands and contracts in the direction of the time axis in response to only the dead time L. However, as shown in Expression 32, when the time-series signal of the input / output variable of the control system is sampled at an integer fraction of the dead time L, The time response of the output will be of approximately the same shape. For example, the dead time L
Is doubled, the shape of the time response of the output with respect to the input extends twice in the direction of the time axis, but the sampling period also doubles, so that the shape of the output with respect to the input becomes almost the same. When the dead time L is reduced by half, the shape of the time response of the output to the input is reduced by half in the direction of the time axis, but the sampling period is also reduced by half. become.

On the other hand, regarding the control parameters K _p , T _i , and T _d , the proportional gain K
_{When p} is increased, the oversist is increased, when the integration time T _i is increased, the rise time is prolonged, and when the differentiation time T _d is increased, vibrations are likely to occur. Further, the dead time delay of the control system is substantially equal to the dead time L of the control target.

From these facts, even when the control system is not optimally adjusted, if the time series signal of the input / output variable of the control system is sampled at an integral fraction of the dead time L as shown in Expression 32, It can be seen that the time response of the output with respect to the input has a shape substantially corresponding to the adjustment state of the control system. Note that this is not limited to the case where the partial model matching method is used, and the same can be said for the case where another adjustment method is used.

Therefore, the time series signals of the input / output variables of the control system are fetched, and these signals are normalized in the direction of the time axis with reference to the dead time L of the control object. If the control parameters of the controller are adjusted by a neural network based on the normalized time series signal, even if the characteristic change of the controlled
Control parameters can be adjusted with a minimum number without increasing the number of networks.

Further, when Expression 22 which is a transfer function of the control system is modified, Expression 33 is obtained.

[0089]

[Equation 33]

From the equation (33), it can be seen that the steady gain of the transfer function of the control system is 1. From this, when the time-series signal of the input / output variables of the control system is normalized by the input variables, the time response of the input and the output to the input is substantially in accordance with the adjustment state of the control system regardless of the fluctuation range of the input. It turns out that it becomes.

Therefore, the time series signals of the input / output variables of the control system are fetched, and these signals are normalized in the direction of the time axis with reference to the dead time L of the control object. And the time series signal of the input / output variable normalized by the dead time L is further normalized by the input variable. For example, as shown in Expression 34, the normalized time-series signal Z is normalized in the amplitude direction by the variation width of the input variable.
If the control parameter C of the controller is adjusted by the neural network based on _N , even if the characteristic change of the control target is large, the control parameter C can be controlled with the minimum number without increasing the number of prepared neural networks for adjustment. Can be adjusted.

[0092]

(Equation 34)

Here, Δr (t _N− 1): the fluctuation value of the target value r normalized in the direction of the time axis Δy (t _N− 1): the fluctuation value of the control amount y normalized in the direction of the time axis Δr _max : the fluctuation range of the target value r As the fluctuation range of the input variable, either a temporally dynamic fluctuation range or a temporally static fluctuation range may be used.

In the embodiment of the present invention, the case where the PID controller shown in Expression 1 is used as the controller 2 has been described. However, in the present invention, the transfer function G _C (s) is expressed by Expression 35.
The present invention can also be applied to a case where a PI (proportional / integral) controller is used.

[0095]

(Equation 35)

The present invention can also be applied to the case where an I-PD controller or an IP controller having the characteristics shown in Expressions 36 and 37 is used.

[0097]

[Equation 36]

Here, x (s): manipulated variable r (s): target value y (s): controlled variable K _p : proportional gain T _i : integration time T _d : differentiation time

[0099]

(37)

The reference model is generalized and the time scale factors σ of various control systems are summarized and shown in Table 1.

[0101]

[Table 1]

From this table, PID, PI, I-PD, I
Σ / L of various control systems using the P controller is n
(= L / T). Also,
Table 1 also shows various relationships between n and σ / L when the reference model is the Kitamori model. That is,
The PID controller has a wide range of n
Σ / L can be approximated by 1.37 with respect to PI, IP
In various control systems of D and IP, σ / L slightly changes in accordance with the change of n, but since the change is not so large, σ / L
Even if L is approximated by a constant value, there is no problem, and the present invention can be applied to various control systems using the PI, I-PD, and IP controllers as described above. Alternatively, when the range of the characteristic change of the control target is known, σ /
L may be approximated to a constant value, and the present invention is applicable. In this case, the approximation accuracy is improved as compared with the case where the entire range is approximated by a constant value.

Further, in the embodiment of the present invention, the case where the position type PID controller shown in Formula 1 is used as the controller 2 has been described. However, the present invention is applied to the case where the speed type PID controller is used as shown in FIGS. The present invention can be applied to a case where a speed controller such as a controller or a speed PI controller is used.

In the embodiment of the present invention, the case where a continuous-time PID controller is used as the controller 2 as shown in Equation 1 has been described. The present invention can be applied to a case where a discrete-time controller such as a PID controller or a discrete-time PI controller is used.

[0105]

(38)

Here, K _p : proportional gain _Ki : integral gain K _d : differential gain

[0107]

[Equation 39]

[0108]

(Equation 40)

[0109]

[Equation 41]

That is, the present invention is not limited to the continuous time position type PID controller as the controller 2 but can be applied to a case where various types of controllers are used.

In the embodiment of the present invention, the case where the transfer function of the controlled object 1 can be approximated by a first-order lag + dead time system has been described.
This can also be applied to the case where approximation can be made with a second-order delay + dead time system shown in FIG.

[0112]

(Equation 42)

Here, T ₁ and T _{2 are} time constants. In the present invention, when the transfer function of the controlled object 1 can be approximated by the integral + dead time system shown in Equation 43, or the integral + 1 order delay shown in Equation 44 + It can also be applied to the case where approximation can be made with a dead time system.

[0114]

[Equation 43]

[0115]

[Equation 44]

That is, the present invention is not limited to the case where the transfer function of the controlled object 1 can be approximated by a first-order lag + dead time system, and includes various types that can be approximated by a term including no dead time + dead time system. The process can be applied.

In the embodiment of the present invention, the case where the normalized time series signal Z _N of the input / output variable of the control system is used as the input signal to the neural network as shown in Expression 10 has been described. Can also be applied to the case where the time change of the normalized time-series signal of the input / output variable of the control system is used as shown in Expression 45.

[0118]

[Equation 45]

In the embodiment of the present invention, the case where the normalized time series signal Z _N of the input / output variable of the control system is used as the input signal to the neural network as shown in Expression 10 has been described. Can also be applied to the case where the normalized time-series signal of the input / output variables and the target value of the control target 1 is used as shown in Expressions 46 and 47.

[0120]

[Equation 46]

[0121]

[Equation 47]

The present invention can also be applied to a case in which a control deviation and a normalized time-series signal of input / output variables of the controlled object 1 are used as input signals to the neural network as shown in Expressions 48 and 49. .

[0123]

[Equation 48]

[0124]

[Equation 49]

That is, according to the present invention, as input signals to the neural network, variables are appropriately selected from input / output variables of the control system, and addition, subtraction, multiplication and division of these normalized time-series signals, time variation, etc. It can be applied to the case where the processing value is used.

In the embodiment of the present invention, the case where the normalized time series signal of the input / output variable of the control system is used as the input signal of the neural network as shown in Expression 10 has been described. As shown by reference numeral 50, the present invention can be applied to the case where the control parameters K _p , T _i , T _d of the controller 2 and the normalized time series signal of the input / output variables of the control system are used.

[0127]

[Equation 50]

Here, K _p *, T _i *, T _d *: K _p , T _i ,
Present value of T _{d Further} , according to the present invention, as input signals to the neural network, the control parameters K _p , T _i , T _{d of} the controller 2 and the input / output of The present invention can also be applied to a case where a normalized time-series signal of a variable and a target value is used.

[0129]

(Equation 51)

[0130]

(Equation 52)

Further, according to the present invention, as input signals to the neural network,
The present invention can also be applied to the case where the control parameters K _p , T _i , T _{d of} the controller 2, the control deviation, and the normalized time series signal of the input / output variables of the control target 1 are used.

[0132]

(Equation 53)

[0133]

(Equation 54)

That is, according to the present invention, as input signals to the neural network, appropriate variables are selected from input / output variables of the control system, and addition / subtraction / multiplication / division of these normalized time-series signals, time variation, etc. The present invention can be applied to the case where the processing value and the control parameter value of the controller 2 are used.

In the embodiment of the present invention, the case where the absolute values of the control parameters K _p , T _i , and T _d of the controller 2 are used as output signals of the neural network as shown in Expression 11 has been described. According to the present invention, the control parameters K _p , T _i ,
The present invention can also be applied to the case where the corrected amount of _Td is used.

[0136]

[Equation 55]

Here, ΔK _p , ΔT _i , ΔT _d : K _p ,
T _i, modify content in this case the T _d, the control parameter K _p, T _i, T _d, the number 56
Required by

[0138]

[Equation 56]

Further, according to the present invention, as the output signal of the neural network, the correction ratio C _p of the control parameters K _p , T _i , T _d of the controller 2 as shown in Expression 57,
The present invention can be applied to the case where C _i and C _d are used.

[0140]

[Equation 57]

In this case, the control parameters K _p , T _i , T
_d is obtained by Expression 58.

[0142]

[Equation 58]

In the embodiment of the present invention, the case where the test signal t _S is superimposed on the target value as shown in FIG. 5 when learning input data is obtained by simulation has been described. As shown in FIG. 9, the present invention can be applied to a case where the test signal t _S is superimposed on the operation amount or the control amount. Note that various signals such as a pseudo random signal, a sine wave signal, a step signal, and a ramp signal can be used as the test signal t _S.

In one embodiment of the present invention, the case where the operation data is used as shown in FIG. 1 when adjusting the control parameters of the controller 2 has been described.
As shown in FIGS. 10, 11, and 12, the present invention can be applied to a case where a test signal is superimposed on a target value, an operation amount, or a control amount. Note that various signals such as a pseudo random signal, a sine wave signal, and a step signal can be used as the test signal.

In one embodiment of the present invention, when learning the adjustment rule of the control parameter, a case where a partial model matching method is used for generating teacher data and a Kitamori model is used as a reference model thereof. Although mainly described, the present invention is not limited to the reference model of the Kitamori model, and can be applied to a case where various other reference models are used. Table 2 shows examples of other reference models.

[0146]

[Table 2]

In one embodiment of the present invention, a case has been described in which a partial model matching method is used to create teacher data when learning the adjustment rules for the control parameters. Without being limited to the model matching method, (1) the ziegler-Nichols method,
(2) Various adjustment methods such as the chien-Hrones-Roswick method (Revised Automatic Control Basic Theory by Masami Masubuchi, Corona, Showa 52)
This is also applicable to the case of using

In the embodiment of the present invention, the case where one kind of the PID controller shown in Formula 1 is used as the controller 2 has been described.
Various controllers such as an I-PD controller and an IP controller and adjustment rules for these controllers are prepared, and one controller and a corresponding adjustment rule are selected from these controllers and adjustment rules for use. Applicable. At this time, there are (1) a method in which adjustment rules of various controllers are prepared as separate weight coefficients corresponding to the controllers, and (2) a method in which adjustment rules of various controllers are handled by one type of weight coefficient. In the latter method, as input signals to the neural network, as shown in Expressions 59 and 60, a variable C _T indicating the type of the controller, a target value, and a normalized time-series signal of the input / output variables of the controlled object 1 are represented. Used.

[0149]

[Equation 59]

[0150]

[Equation 60]

In one embodiment of the present invention, when learning the adjustment rule of the control parameter, a partial model matching method is used for generating teacher data, and the characteristic amount of the response waveform is used as a reference model. Although the case where the Kitamori model with a fixed overshoot amount and damping ratio is used has been mainly described, the present invention uses a reference model in which the characteristic amount of the response waveform such as the overshoot amount and damping ratio is variable, It is also applicable to the case where the characteristic amount of the response waveform such as the amount of overshoot and the damping ratio can be designated in the adjustment of. At this time, (1) a method of preparing controller adjustment rules as separate weighting factors corresponding to various overshoot amounts, damping ratios, etc., and (2) a method of preparing controller adjustment rules, various overshoot amounts, damping ratios, etc. There is a method corresponding to 121 kinds of weight coefficients. In the latter method, Equations 61 and 6 are used as input signals to the neural network.
As shown in FIG. 2, variables 0 and A for specifying an overshoot amount, a damping ratio, and the like, a target value, and normalized time-series signals of input / output variables of the controlled object 1 are used.

[0152]

[Equation 61]

[0153]

(Equation 62)

In one embodiment of the present invention, a 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 relates to a case where the control target is a multi-input multi-output system. Can also be applied. In this case, adjustment rules are prepared corresponding to controllers of a multi-input multi-output system such as various chemical plants and thermal power plants, and an appropriate adjustment rule is selected from these adjustment rules.

In the embodiment of the present invention, the case where the learning input data is obtained by simulation has been described. However, the present invention is applicable to the time response (operation data) of the control system of the plant having various characteristics. The time response is recorded, and the normalized time-series signal of the input / output variables of the control system obtained from the time response is used as input data for learning, and the control parameters corresponding to the time response with good characteristics are recorded. In addition, the present invention can be applied to a case where a neural network is trained by using this parameter as learning teacher data.

In the embodiment of the present invention, the case where the input / output signals of the control system are sequentially taken and the adjustment of the control parameter is carried out sequentially has been mainly explained. The present invention can also be applied to a case where input / output signals are sequentially acquired and control parameters are adjusted in a batch manner. That is, the present invention can be applied to a case where control parameters are adjusted in batches at a point in time when input and output signals of a control system which are sequentially captured are gathered to some extent.

In one embodiment of the present invention, the time series signals of the input / output variables of the control system are normalized in the direction of the time axis based on the dead time of the control object. The time series signal of the input / output variable of the control system may be normalized in the direction of the time axis with reference to the dead time of the control system.

In one embodiment of the present invention, the time-series signals of the input / output variables of the control system are normalized in the direction of the time axis by sampling the dead time L of the control object by an integer fraction of an integer. However, as shown in Equation 63, the time series signal of the input / output variable of the control system may be normalized in the direction of the time axis by sampling at a period Δt proportional to the dead time L.

[0159]

[Equation 63]

Here, a: constant In one embodiment of the present invention, the time series signal of the input / output variable of the control system is normalized in the amplitude direction by the variation width of the input variable. Alternatively, the time series signal of the input / output variable of the control system may be normalized in the amplitude direction by the variation width of each variable.

[0161]

[Equation 64]

Here, Δy _max : the fluctuation amount of the control amount y It is to be noted that the fluctuation width of the input / output variable can be either a dynamic fluctuation width in time or a static fluctuation width in time. Good.

In the embodiment of the present invention, the method of measuring the dead time of the control object or the control system has not been specifically described. However, the step response method, the least square estimation method, and the like are used for measuring the dead time. Yes (see Kajiro Watanabe et al., "Control Engineering Using a Personal Computer", Kaibundo, November 1990), and these methods can be used.

[0164]

According to the present invention, the time series signals of the input / output variables of the control target or the control system are fetched, and these signals are normalized in the direction of the time axis based on the dead time of the control target or the control system. In addition, the control parameters of the controller are adjusted by the neural network based on the normalized time series signal based on the normalized time series signal based on the fluctuation width of the time response of the control target or the control system. It is possible to cope with a case where the characteristic change of the control target or the control target is large. The number of prepared neural networks for adjustment can be minimized, and the control parameters can be satisfactorily adjusted.

[Brief description of the drawings]

FIG. 1 is a control configuration diagram showing one embodiment of the present invention.

FIG. 2 is a configuration diagram of a neural network.

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

FIG. 4 is a configuration diagram illustrating an example in which a control target is controlled by a PID controller.

FIG. 5 is a control configuration diagram showing one embodiment of the present invention.

FIG. 6 is a control configuration diagram showing one embodiment of the present invention.

FIG. 7 is a control configuration diagram showing one embodiment of the present invention.

FIG. 8 is a control configuration diagram showing one embodiment of the present invention.

FIG. 9 is a control configuration diagram showing one embodiment of the present invention.

FIG. 10 is a control configuration diagram showing one embodiment of the present invention.

FIG. 11 is a control configuration diagram showing one embodiment of the present invention.

FIG. 12 is a control configuration diagram showing one embodiment of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Control target, 2 ... Controller, 3 ... Time series signal normalization processing system, 4 ... Parameter adjustment system, 5 ... Adjustment rule learning system.

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Eiji Toyama 5-2-1 Omikacho, Hitachi City, Ibaraki Prefecture Inside the Hitachi, Ltd. Omika Plant (72) Inventor Atsushi Yokokawa 5-chome Omikacho, Hitachi City, Ibaraki Prefecture No. 1 In the Omika factory, Hitachi, Ltd. (56) References JP-A-2-8903 (JP, A) JP-A-3-233702 (JP, A) JP-A-2-108179 (JP, A) 3-129401 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G05B 11/00-13/04

Claims (2)

(57) [Claims] 1. A method for adjusting a controller for controlling a control target so that a control amount matches a target value, wherein a time series signal of input / output variables of a control system or a control system in which the controller and the control target are combined is acquired. These signals are normalized in the direction of the time axis with reference to the dead time of the control object or control system, and the control parameters of the controller are determined by a neural network based on the normalized time series signals. How to adjust the controller. 2. A method for adjusting a controller for controlling a control target so that a control amount matches a target value, wherein a time series signal of an input / output variable of a control target or a control system in which the controller and the control target are combined is acquired. These signals are normalized in the direction of the time axis based on the dead time of the control target or the control system, and normalized in the amplitude direction based on the fluctuation width of the time response of the control target or the control system. A controller adjustment method comprising: determining a control parameter of a controller by a neural network based on the obtained time-series signal.