JP4825960B2 - Control device, control parameter adjustment device, and control parameter adjustment method - Google Patents

Description

  The present invention relates to a control device for controlling a physical state such as temperature and pressure of a control target, a device and method for adjusting control parameters thereof, and a program used for them.

  In a conventional control device, for example, a PID control device, auto-tuning is performed using a step response method or a limit cycle method, and control is performed by adjusting a PID control parameter. The system is identified and the PID control parameter is changed according to the identification result.

  However, in such a conventional example, particularly in a system with strong nonlinearity, there is a problem that reliability of identification is not sufficient and desired control characteristics may not be obtained.

Therefore, the applicant of the present application has already proposed a control device that can obtain more accurate control characteristics according to the system (see Patent Document 1).
JP 2005-339004 A

  The proposed control device corrects the characteristic parameter such as the control parameter generated corresponding to the new state vector by using the correction coefficient, which is a correction coefficient, so that the control error or the model error is reduced. The obtained characteristic parameters are accumulated in the database, that is, learning is performed so that the control error or the model error becomes small.

  For this learning, the learning coefficient previously adjusted by the user is used. However, the learning coefficient must be adjusted by the user on a trial and error basis according to the control target. There is a problem that it is troublesome.

  The present invention has been made in view of the above-described points, and an object thereof is to eliminate the need for a user to adjust the learning coefficient.

  (1) A control device according to the present invention includes a control unit that controls a control target according to a control parameter, a vector that includes a control system characteristic amount as a state vector, and a database that stores a control system characteristic parameter. Adjusting means for adjusting a control parameter of the control means based on the data, and the adjusting means is provided with a new state vector, so that the new state vector is changed based on the state vector of the database. A characteristic parameter generation unit that generates a corresponding characteristic parameter; and a characteristic parameter correction unit that corrects the characteristic parameter generated by the characteristic parameter generation unit in a direction in which a control error or a model error is reduced using a correction coefficient; The characteristic parameter corrected by the characteristic parameter correction unit is stored in the database. A control device for a product, the characteristic parameter correction unit, the correction coefficient, and calculates in accordance with a predetermined calculation formula.

  Here, the characteristic amount of the control system refers to an amount representing the characteristic of the control system, and refers to an amount representing the characteristic of the control target, the control device, or the environment in the control system including the control target, the control device, and the environment. For example, the control amount, the operation amount, the control error, the ambient temperature, the magnitude of the disturbance, and the like.

  The feature amount preferably includes at least control amount time-series data, or preferably includes at least one of operation amount time-series data and a target value (set value).

  A state vector refers to a vector that represents a control system feature value.

  Further, the characteristic parameter of the control system refers to a parameter that represents the characteristic of the control system, and in the control system configured by the control target, the control device, and the environment, refers to a parameter that represents the characteristic of the control target, the control device, or the environment. For example, it refers to a control parameter, a coefficient of a transfer function to be controlled, a time constant of the control object, a dead time, a steady gain, and the like.

  It is preferable that the control system characteristic amount and the control system characteristic parameter are stored in the database in association with each other.

  The adjustment means selects a state vector close to the new state vector from the database when a new state vector as a request point is given, and corresponds to the new state vector based on the selected state vector. It is preferred to generate characteristic parameters.

  The state vector close to the new state vector refers to a state vector that is close to the new state vector, or a state vector similar to the new state vector. The vector distance is preferably a weighted distance or a Mahalanobis distance.

  Preferably, the characteristic parameter generation unit configures the characteristic parameter as a local model using a weighted linear average method or a weighted local regression method based on the state vector of the parameter adjustment database.

  The model error means a difference between the control amount predicted using the local model and the actual control amount.

  Further, the correction in the characteristic parameter correction unit is preferably corrected in the direction of decreasing the control error or the model error by the steepest gradient method (hill climbing method).

  In this characteristic parameter correction unit, the characteristic parameter is corrected using the correction coefficient. However, if the degree of change (change rate) between the characteristic parameter before correction and the characteristic parameter after correction changes, a control error or The speed of correction in the direction of reducing the model error, that is, the learning speed fluctuates, so that learning progresses too much and becomes unstable, or learning is too slow.

  Therefore, it is preferable that the calculation formula for calculating the correction coefficient can calculate a correction coefficient that does not change the learning speed.

  According to the control device of the present invention, the feature quantity and characteristic parameter of the control system are accumulated in the database, and when a new state vector is given, a new state vector is generated as a local model based on the accumulated state vector of the database. Control parameters are adjusted based on the so-called Just-In-Time (JIT) method or Model-on-Demand (MOD) method, which generates characteristic parameters corresponding to In addition to obtaining high-precision control characteristics, each time a new characteristic parameter is generated, the control error or model error can be corrected in the direction of decreasing and stored in the database, so that the control error or model error is reduced. It is possible to obtain more accurate control characteristics by learning.

Moreover, since correction in the direction of reducing the control error or model error is performed using a correction coefficient calculated by a predetermined calculation formula, the user adjusts the correction coefficient in a trial and error manner according to the control target. There is no need to do.
(2) In one embodiment of the control device of the present invention, the characteristic parameter of the control system is the control parameter of the control means.
The control parameter is preferably a PID control parameter of the control means.
According to this embodiment, a control parameter can be generated as a characteristic parameter corresponding to a required point, and the control parameter of the control means can be adjusted.

  (3) In the embodiment of (2) above, the calculation formula may include the term of the control parameter and at least one of the term of the target value and the term of the control amount.

  It is preferable that the term of the control parameter included in the calculation formula is a term of the control parameter before correction in the characteristic parameter correction unit.

  According to this embodiment, the correction coefficient can be changed according to the control parameter, and can be changed according to the target value and the control amount.

  (4) In the embodiment of (3) above, the calculation formula is such that the correction coefficient is proportional to the control parameter, and the square of the target value, the square of the maximum value of the target value, the control It is inversely proportional to either the square of the amount, the square of the maximum value of the control amount, or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount. It may be an expression.

  According to this embodiment, it is possible to calculate a correction coefficient that suppresses fluctuations in the correction speed in the direction of reducing the control error or the model error, that is, the learning speed.

  (5) In another embodiment of the present invention, the characteristic parameter correction unit may correct the characteristic parameter at a period equal to or longer than the dead time of the control target.

  According to this embodiment, it is possible to suppress the occurrence of overshoot and oscillation due to the speed of correction using the correction coefficient, that is, the learning speed is too fast.

  (6) A control parameter adjusting device according to the present invention is a device that adjusts the control parameter of a control device that controls an object to be controlled in accordance with the control parameter, and stores a vector including a feature quantity of the control system as a state vector. And adjusting means for adjusting the control parameters of the control device based on the data of the database in which the characteristic parameters of the control system are stored, and the adjusting means is provided with a new state vector, so that Based on the state vector, a characteristic parameter generation unit that generates a characteristic parameter corresponding to the new state vector, and a characteristic parameter generated by the characteristic parameter generation unit uses a correction coefficient to reduce a control error or a model error. A characteristic parameter correction unit for correcting the direction, and the characteristic parameter The characteristic parameters fixed in correcting unit, there is accumulating in the database, the characteristic parameter correction unit, the correction coefficient, and calculates in accordance with a predetermined calculation formula.

  According to the control parameter adjusting apparatus of the present invention, the control system feature quantity and characteristic parameters are accumulated in the database, and when a new state vector is given, a new local model is created based on the accumulated state vector of the database. Because the characteristic parameters corresponding to the state vector are generated and the control parameters of the control device are adjusted, highly accurate control characteristics can be obtained and new characteristic parameters are generated even in a highly nonlinear system. In addition, since the control error or the model error can be corrected in the direction of decreasing and stored in the database, it is possible to obtain control characteristics with higher accuracy by learning to reduce the control error or the model error.

  Moreover, since correction in the direction of reducing the control error or model error is performed using a correction coefficient calculated by a predetermined calculation formula, the user adjusts the correction coefficient in a trial and error manner according to the control target. There is no need to do.

  (7) In one embodiment of the control parameter adjusting device of the present invention, the characteristic parameter of the control system is the control parameter of the control device.

  According to this embodiment, a control parameter can be generated as a characteristic parameter corresponding to a required point, and the control parameter of the control device can be adjusted.

  (8) In the embodiment of the above (7), the calculation formula may include a term of the control parameter and at least one of a term of the target value and a term of the control amount.

  According to this embodiment, the correction coefficient can be changed according to the control parameter, and can be changed according to the target value and the control amount.

  (9) In the embodiment of (8), the calculation formula is such that the correction coefficient is proportional to the control parameter, and the square of the target value, the square of the maximum value of the target value, the control It is inversely proportional to either the square of the amount, the square of the maximum value of the control amount, or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount. It may be an expression.

  According to this embodiment, it is possible to calculate a correction coefficient that suppresses fluctuations in the correction speed in the direction of reducing the control error or the model error, that is, the learning speed.

  (10) In another embodiment of the control parameter adjusting device of the present invention, the characteristic parameter correction unit corrects the characteristic parameter in a cycle equal to or longer than the dead time of the control target.

  According to this embodiment, it is possible to suppress the occurrence of overshoot and oscillation due to the speed of correction using the correction coefficient, that is, the learning speed is too fast.

  (11) A method for adjusting a control parameter according to the present invention is a method for adjusting the control parameter of a control device that controls an object to be controlled in accordance with the control parameter, and stores a vector including a control system feature amount as a state vector in a database And storing a control system characteristic parameter in the database and generating a characteristic parameter corresponding to the new state vector based on the state vector of the database given a new state vector. A correction step for correcting the generated characteristic parameter in a direction in which a control error or a model error is reduced using a correction coefficient, an accumulation step for storing the corrected characteristic parameter in the database, and the characteristic parameter, Step for adjusting the control parameter Wherein the door, in the correction step, the correction coefficient, and calculates in accordance with a predetermined calculation formula.

  According to the control parameter adjustment method of the present invention, the control system feature quantity and characteristic parameters are accumulated in the database, and when a new state vector is given, a new local model is created based on the accumulated state vector of the database. Because the characteristic parameters corresponding to the state vector are generated and the control parameters of the control device are adjusted, highly accurate control characteristics can be obtained and new characteristic parameters are generated even in a highly nonlinear system. In addition, since the control error or the model error can be corrected in the direction of decreasing and stored in the database, it is possible to obtain control characteristics with higher accuracy by learning to reduce the control error or the model error.

  Moreover, since correction in the direction of reducing the control error or model error is performed using a correction coefficient calculated by a predetermined calculation formula, the user adjusts the correction coefficient in a trial and error manner according to the control target. There is no need to do.

  (12) In one embodiment of the control parameter adjustment method of the present invention, the characteristic parameter of the control system may be a control parameter of the control device.

  According to this embodiment, a control parameter can be generated as a characteristic parameter corresponding to a required point, and the control parameter of the control device can be adjusted.

  (13) In the embodiment of (12), the calculation formula includes the control parameter term, the control parameter term, and at least one of the target value term and the controlled variable term. Terms may be included.

  According to this embodiment, the correction coefficient can be changed according to the control parameter, and can be changed according to the target value and the control amount.

  (14) In the embodiment of the above (13), the calculation formula is such that the correction coefficient is proportional to the control parameter, the square of the target value, the square of the maximum value of the target value, the control It is inversely proportional to either the square of the amount, the square of the maximum value of the control amount, or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount. It may be an expression.

  According to this embodiment, it is possible to calculate a correction coefficient that suppresses fluctuations in the correction speed in the direction of reducing the control error or the model error, that is, the learning speed.

  (15) In another embodiment of the control parameter adjustment method of the present invention, the characteristic parameter correction period of the correction step may be equal to or longer than the dead time of the control target.

  According to this embodiment, it is possible to suppress the occurrence of overshoot and oscillation due to the speed of correction using the correction coefficient, that is, the learning speed is too fast.

  According to the present invention, feature quantities and characteristic parameters of a control system are accumulated in a database, and when a new state vector is given, a new state vector is supported as a local model based on the accumulated state vector of the database. Control parameters are adjusted and control parameters are adjusted, so that even highly nonlinear systems can obtain highly accurate control characteristics, and each time a new characteristic parameter is generated, control errors or model errors are reduced. Since the correction is made in the direction of decreasing, highly accurate control characteristics can be obtained.

  In addition, since the correction in the direction of reducing the control error or the model error is performed using a correction coefficient calculated according to a predetermined calculation formula, the user adjusts the correction coefficient in a trial and error manner according to the control target. There is no need to do.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(Embodiment 1)
First, prior to a characteristic description of an embodiment of the present invention, a premise configuration of the present invention disclosed in Patent Document 1 will be described.

  FIG. 1 is a block diagram of a main part of a temperature regulator as a control device according to one embodiment of the present invention.

  The temperature controller 1 of this embodiment performs temperature control of the control target 2 such as a heat treatment apparatus, and includes a set temperature (target temperature) r set by a host device or a setting unit (not shown), and the control target 2 Is provided with a PID controller 3 as a control means for calculating and calculating the manipulated variable u based on a deviation e from a detected temperature y which is a controlled variable from a temperature sensor (not shown) disposed in the PID controller 3. The adjusting means 4 for adjusting the PID control parameter is provided.

  In the temperature controller 1 of this embodiment, the following is performed in order to obtain highly accurate control characteristics.

  That is, each time new data is obtained, it is stored in the database 6 and, if necessary, a local model is created by taking out similarities to the requested points from a large number of databases 6 accumulated in the past as neighborhoods. The PID control parameter of the PID controller 3 is adjusted based on a so-called Just-In-Time (JIT) method.

  First, the JIT modeling method will be described.

  Consider a discrete-time nonlinear system expressed by the following equation.

上式は、ＮＡＲＭＡＸ（Ｎonlinear ＡＲＭＡＸ）モデルと呼ばれ、ダイナミクスを持つ大域的な非線形システムを表すモデル表現として広く使用されている。ここで、ｙ（ｔ）はシステム出力、ｆ（・）は非線形関数、ｅ（ｔ）は観測雑音、φ（ｔ）はシステムの時刻ｔでの状態を表しており、状態ベクトルと呼ぶこととする。状態ベクトルφ（ｔ）は次式で定義される。

The above equation is called a NARMAX (Nonlinear ARMAX) model, and is widely used as a model expression representing a global nonlinear system having dynamics. Here, y (t) is a system output, f (•) is a nonlinear function, e (t) is observation noise, φ (t) is a state of the system at time t, and is called a state vector. To do. The state vector φ (t) is defined by the following equation.

ここで、ｕ（ｔ）はシステム入力、ｄはむだ時間、ｎ_ｙ，ｎ_ｕはそれぞれ出力と入力の次数である。ＪＩＴモデリングでは、（２）式の状態ベクトルの形式でデータベースへの蓄積が行われる。また、出力ｙ（ｔ）の予測値を得るために必要な状態ベクトルφ（ｔ）を要求点ｑと呼ぶ。ＪＩＴモデリング法とは、この要求点に類似した状態ベクトルをデータベースから近傍として抽出し、局所モデルを構成する方法である。

Here, u (t) is a system input, d is a dead time, and n _y and n _u are output and input orders, respectively. In JIT modeling, data is stored in a database in the state vector format of equation (2). Further, the state vector φ (t) necessary for obtaining the predicted value of the output y (t) is called a request point q. The JIT modeling method is a method of extracting a state vector similar to this request point as a neighborhood from a database and constructing a local model.

  In this embodiment, the PID control parameter of the PID controller 3 in the block diagram of FIG. 1 is adjusted based on the JIT method as described above.

  Here, a speed type I-PD control law as shown below is applied.

但し、ｅ(ｔ)は制御誤差信号であり、ｒ(ｔ)を目標値とすると、以下で定義される。

However, e (t) is a control error signal, and is defined below, where r (t) is a target value.

また、ｋｃ、Ｔ_Ｉ、およびＴ_Ｄはそれぞれ比例ゲイン、積分時間、微分時間を、Ｔｓはサンプリング時間を表しており、(４)式では、簡単のために、Ｋｐ＝ｋｃ、Ｋ_Ｉ＝ｋｃＴｓ／Ｔ_Ｉ、Ｋ_Ｄ＝ｋｃＴ_Ｄ／Ｔｓと置き換えている。さらに、Δ(：＝１−ｚ^−１)は差分演算子を表している。

Further, kc, T _I , and T _D represent proportional gain, integration time, and derivative time, respectively, and Ts represents sampling time. In the equation (4), for simplicity, Kp = kc, K _I = kcTs / T _I , K _D = kcT _D / Ts Further, Δ (: = 1−z ⁻¹ ) represents a difference operator.

  The basic idea of this method is to obtain PID control parameters as an inverse model such as the following equation by the JIT method.

但し、ｙ_ｒは参照モデルの出力を示しており、次式のように設計される。

However, y _r denotes the output of the reference model is designed as follows.

上式に含まれるＴ(ｚ^−１)については、Ｔ．Ｙamamoto and Ｓ．Ｌ．Ｓhah： Ｄesign and Ｅxperimental Ｅvaluation of ａ Ｍultivariable Ｓelf-Ｔuning ＰＩＤ Ｃontroller,Ｐroc.of ＩＥＥＥ Ｃonference on Ｃontrol Ａpplications,Ｔrieste,pp.1230-1234(1998)に基づいて設計する。

Regarding T (z ⁻¹ ) included in the above equation, T.W. Yamamoto and S. L. Shah: Designed based on Design and Experimental Evaluation of a Multivariable Self-Tuning PID Controller, Proc. Of IEEE Conference on Control Applications, Trieste, pp. 1230-1234 (1998).

The JIT-PID controller algorithm is shown below.
[STEP 1] Creation of initial database In the JIT method, if there is no past accumulated data, it is impossible to identify in principle. Create
[STEP 2] Distance calculation, neighborhood selection A state vector necessary for obtaining a predicted value of the PID parameter K (t) at time t is called a request point q. The distance between the request point q and the state vector stored in the database is obtained by the weighted L1 norm of the following equation.

この距離が小さいものからｋ個のベクトルを近傍として選択する。
［ＳＴＥＰ３］ 局所モデルの構成
続いて、ＳＴＥＰ２において選択された近傍に対して、以下で示される、重みつき局所線形平均法（ＬＷＡ）により局所モデルを構成する。

K vectors having the smallest distance are selected as neighbors.
[STEP 3] Configuration of Local Model Next, a local model is configured for the neighborhood selected in STEP 2 by the weighted local linear average method (LWA) shown below.

ここで、ｗｉは、近傍の第ｉサンプルに対応する重みであり、次式で与える。

Here, wi is a weight corresponding to the i-th sample in the vicinity, and is given by the following equation.

［ＳＴＥＰ４］ データ修正
ＳＴＥＰ３において、局所モデルとして得られるＰＩＤ制御パラメータＫ^ｏｌｄは、制御誤差（ε：＝ｙ_ｒ−ｙ）を考慮していないため、ＳＴＥＰ２、ＳＴＥＰ３を繰り返し行い、ＰＩＤ制御パラメータを逐次抽出したとしても、システムの状態に応じてＰＩＤ制御パラメータが適切に調整されない恐れがある。そこで、ＳＴＥＰ２で得られたＰＩＤ制御パラメータＫ^ｏｌｄに対して制御誤差εにより修正を行ない、その修正されたデータＫ^ｎｅｗをデータベースに蓄えるものとする。

[STEP 4] Data correction In STEP 3, the PID control parameter K ^old obtained as a local model does not consider the control error (ε: = y _r −y). Therefore, STEP 2 and STEP 3 are repeated, and the PID control parameter is sequentially set. Even if it is extracted, the PID control parameter may not be adjusted appropriately according to the state of the system. Accordingly, it is assumed that the PID control parameter K ^old obtained in STEP 2 is corrected by the control error ε, and the corrected data K ^new is stored in the database.

  The correction method is given by the following equation.

ここで、ηは修正係数としての学習係数、Ｊは以下で定義される誤差の評価規範を表している。

Here, η represents a learning coefficient as a correction coefficient, and J represents an error evaluation standard defined below.

また、（１３）式、右辺第２項のそれぞれの偏微分は、以下のように展開される。

Further, the partial differentials of the expression (13) and the second term on the right side are expanded as follows.

この実施の形態では、上述のようにＪＩＴ法に基づいて、ＰＩＤ制御器３のＰＩＤ制御パラメータを調整するために、調整手段４は、参照モデル５と、パラメータ調整用のデータベース６と、ＰＩＤチューニング部７とを備えている。

In this embodiment, in order to adjust the PID control parameters of the PID controller 3 based on the JIT method as described above, the adjustment means 4 includes a reference model 5, a parameter adjustment database 6, and PID tuning. Part 7.

  In this embodiment, as shown in FIG. 2, the database 6 stores data consisting of a state vector, which is a vector including a control system feature quantity, and a PID control parameter corresponding to this state vector. The That is, a set of PID control parameters corresponds to a set of state vectors.

In this embodiment, as shown in FIG. 3, the state vector is composed of the output yr of the reference model 5 and the time series data of the detected temperature y of the controlled object 2 that is the controlled variable, and the PID control parameter is Kp = kc, K _I = kcTs / T _I , and K _D = kcT _D / Ts.

  The PID tuning unit 7 of this embodiment includes a neighborhood selection unit 8, a local model generation unit 9 as a PID control parameter generation unit, and a data correction unit 10, as shown in the functional block diagram of FIG. ing.

  In the neighborhood selection unit 8, every time the above-mentioned request point is given, that is, every time the state vector at the latest time t is given, the distance between this state vector and the state vector stored in the database 6 is It is obtained by the weighted L1 norm of the above-described equation (10), and k state vectors having the smallest distance are selected as neighborhoods.

The local model generation unit 9 obtains the PID control parameter K ^old by the weighted local linear average method for the selected nearby state vector. That is, a weight is determined according to the above equation (12) using the k neighboring state vectors selected, and this weight is determined as a PID control parameter corresponding to the k state vectors according to the above equation (11). Is multiplied (by weighted average) to obtain the PID control parameter K ^old .

The PID control parameter K ^old generated by the local model generation unit 9 is set in the PID controller 3.

The data correction unit 10 corrects the PID control parameter K ^old generated by the local model generation unit 9 in such a direction that the control error is reduced as shown in the above-described equations (13) to (17). The corrected PID control parameter K ^new is stored in the database 6 in association with the state vector.

  In this data correction unit 10, in order to correct based on the control error, the output yr of the reference model 5 and the detected temperature y that is a control amount are given.

  In this embodiment, the PID controller 3 and the adjusting means 4 are constituted by, for example, a microcomputer, and based on the JIT method by executing a program according to the present invention stored in the ROM of the microcomputer. The PID control parameter is adjusted, and the database 6 described above may be configured on a RAM of a microcomputer, for example, or an external storage device may be used.

  In this embodiment, the amount of data that can be stored in the database 6 is limited, and when the control object 2 changes with time, it is preferable to delete the old data. Is trying to delete from the old data.

  As described above, numerical examples are shown below to evaluate the effectiveness of the temperature controller 1 for adjusting the PID control parameter of the PID controller 3 based on the JIT method.

  Here, a polystyrene polymerization reactor is considered as a control target. The relational expression between the jacket temperature u (t) and the reaction temperature y (t) in the reactor is given by the following equation (Nakanishi, Hanakuma: Process Control Basics and Practice, Asakura Publishing, (1992) ).

ここで、Ｅａ＝２４０、Ｒ＝０．０１９８６である。このシステムは、目標値が高くなる(約７５度以上)につれ非線形性が著しく強くなるという特徴を持っている。

Here, Ea = 240 and R = 0.01986. This system has a feature that the nonlinearity becomes remarkably stronger as the target value becomes higher (about 75 degrees or more).

  A PID control parameter adjustment method based on the JIT method is applied to this system.

Here, the design polynomial T (z ⁻¹ ) included in the reference model is designed as follows.

ここで、Ｔ(ｚ^−１)は、上述の文献Ｔ．Ｙamamoto and Ｓ．Ｌ．Ｓhah： Ｄesign and Ｅxperimental Ｅvaluation of ａ Ｍultivariable Ｓelf-Ｔuning ＰＩＤ Ｃontroller,Ｐroc.of ＩＥＥＥ Ｃonference on Ｃontrol Ａpplications,Ｔrieste,pp.1230-1234(1998)を参考にして設計した。

Here, T (z ⁻¹ ) is the above-mentioned document T.I. Yamamoto and S. L. Shah: Designed with reference to Design and Expert Evaluation of a Multivariable Self-Tuning PID Controller, Proc. Of IEEE Conference on Control Applications, Trieste, pp. 1230-1234 (1998).

  FIG. 5 shows the control result at this time, and FIG. 6 shows the temporal change of the PID control parameter.

  As shown in these figures, it can be seen that the PID control parameters are appropriately adjusted online according to the nonlinearity of the system, and a good control result is obtained.

  In such a temperature controller, as described above, the PID control parameter generated corresponding to the new state vector is corrected using the learning coefficient η, which is a correction coefficient, in a direction in which the control error or the model error is reduced. However, the learning coefficient η is a fixed value set by the user. For this reason, as shown in the above equation (13), the correction range of the PID control parameter changes depending on the value of the PID control parameter K (t), and the learning becomes slow or the learning progresses too much. It becomes easy to oscillate.

  For this reason, the user has to make trial and error adjustments according to the control target so as to obtain an appropriate learning coefficient η, and the trial run using the actual control target must be repeated under the same conditions as during actual operation. There is a problem that it may be necessary.

  Therefore, in this embodiment, the learning coefficient η can be calculated according to a previously calculated calculation formula.

That is, in this embodiment, the learning coefficient η is calculated according to one of the following formulas.
η = [η _P η _I η _D ] = A · K ^old (t) / r (t) ²
η = [η _P η _I η _D ] = A · K ^old (t) / y (t) ²
η = [η _P η _I η _D ] = A · K ^old (t) / [r (t) · y (t)]
Here, A is a constant, for example, preferably A = 0.1 to 10, K ^old is the above-mentioned PID control parameter generated by the local model generation unit 9, and r (t) is the above setting. The value (target temperature) and y (t) are the above-mentioned output (detected temperature) which is a controlled variable.

  Next, the process of deriving this calculation formula will be described.

  A condition for maintaining a constant learning speed without deriving too much or slowing down learning is derived.

  First, the learning speed is defined as follows.

Learning speed = PID gain change rate = K ^new / K ^old (20)
Here, the (15), (16) Evaluation of the error from the equation norms J and above (13), (14), ^{K new new} from equation (17) is expressed as follows.

上記（２０）式と、Ｋｐについての上記（２２）式である
Ｋｐ^ｎｅｗ（ｔ）＝Ｋｐ^ｏｌｄ（ｔ）＋η_Ｐ・ε（ｔ＋１）{ｙ（ｔ−１）−ｙ（ｔ）}とにより、
学習速度＝Ｋｐ^ｎｅｗ／Ｋｐ^ｏｌｄ
＝［Ｋｐ^ｏｌｄ＋η_Ｐ・ε（ｔ＋１）{ｙ（ｔ−１）−ｙ（ｔ）}］／Ｋｐ^ｏｌｄ
＝１＋｛η_Ｐ・ε（ｔ＋１）{ｙ（ｔ−１）−ｙ（ｔ）}｝／Ｋｐ^ｏｌｄ
となる。

According to the above equation (20) and the above equation (22) for Kp, Kp ^new (t) = Kp ^old (t) + η _P · ε (t + 1) {y (t−1) −y (t)} ,
Learning speed = Kp ^new / Kp ^old
= [Kp ^old + η _P · ε (t + 1) {y (t−1) −y (t)}] / Kp ^old
= 1 + {η _P · ε (t + 1) {y (t−1) −y (t)}} / Kp ^old
It becomes.

Therefore, it can be seen that the learning speed of Kp is approximately proportional to ε (t + 1) and {y (t−1) −y (t)} and inversely proportional to Kp ^old .

  Furthermore, ε (t + 1) is approximately proportional to r (t), and {y (t−1) −y (t)} is approximately proportional to y (t).

As a result, the learning speed of Kp is approximately proportional to r (t) and y (t) and inversely proportional to Kp ^old .

That is,
Kp learning speed ∝r (t) · y (t) / Kp ^old
Here, in order to keep the learning speed constant regardless of these variables, the influence of the variables may be canceled by setting the inverse of these variables to η.

In particular,
η _P ∝Kp ^old / [r (t) · y (t)]
Therefore, it can be expressed as follows using the constant A.

η _P = A · Kp ^old / [r (t) · y (t)]
If r (t) ≈y (t),
η _P = A · Kp ^old / [r (t) · y (t)] ≈A · Kp ^old / r (t) ²
≒ A ・ Kp ^old / y (t) ²
K _I, for even _{K D,} likewise, can be derived to calculate equation as follows.

η _I = A · K _I ^old / [r (t) · y (t)] ≈A · K _I ^old / r (t) ²
≒ A ・ K _I ^old / y (t) ²
η _D = A · K _D ^old / [r (t) · y (t)] ≈A · K _D ^old / r (t) ²
≒ A ・ K _D ^old / y (t) ²
In summary,
η = [η _P η _I η _D ] = A · K ^old (t) / r (t) ² (23)
η = [η _P η _I η _D ] = A · K ^old (t) / y (t) ² (24)
η = [η _P η _I η _D ] = A · K ^old (t) / [r (t) · y (t)] (25)
The constant A may be different for each of the proportional, integral, and differential learning coefficients η _P , η _I , and η _D.

  Further, when the maximum value rmax of the target temperature r (t) or the maximum value ymax of the control amount y (t) is known in advance, in the above equations (23) to (25), r Rmax may be used instead of (t), or ymax may be used instead of y (t).

  In this embodiment, for example, the learning coefficient η is calculated according to the calculation formula (23), and the PID control parameter is corrected using the learning coefficient η.

That is, in the data correction unit 10 described above, the learning coefficient η is calculated from the equation (23) using the PID control parameter K ^old (t) and the target value r (t) generated by the local model generation unit 9, The PID control parameter K ^old (t) is corrected using the calculated learning coefficient η, and the corrected PID control parameter K ^new is stored in the database 6 in association with the state vector. The constant A in the calculation formula may be set as a default value, for example, so that the user can adjust it as necessary.

  Here, the constant A will be described. FIG. 7 shows a simulation result of a change in output y (t) when the constant A is set to A = 0.1, 1, 10 using the above equation (23) as a calculation formula for the learning coefficient η. The solid line indicates A = 1, the broken line indicates A = 0.1, the alternate long and short dash line indicates A = 10, and the target value is also indicated by a virtual line.

In the simulation of FIG. 7, the characteristics of the controlled object are the characteristics of the dead time L and the first order delay {k / (τs + 1)}, the time constant τ = 100 seconds, the dead time L = 1 second, and the steady state gain (system gain). ) When K = 100 to 300 ° C. and the target value is 10 ° C., K = 300 ° C., and when the target value is 30 ° C., K = 100 ° C. For other target values, linear interpolation is used. It is. The gains are Kp = 0.33, K _I = 0.011, and K _D = 0.118.

  8 and 9 show the sections of time t = 200 to 600 and 1400 to 1800 in FIG. 7 in an enlarged manner.

  In any case of A = 0.1, 1 and 10, the deviation from the target value is improved as the learning progresses. However, at A = 0.1 shown by the broken line, the learning speed is slow, and the dashed line At A = 10 shown, learning progresses too much and overshoot occurs, whereas when A = 1, overshoot does not occur and the learning speed is not slow, which is an appropriate value. I understand that.

  FIG. 10 shows a change in ISE (Integral of Squared Error) when the constant A is set to A = 0.1, 1, 10, and the horizontal axis corresponds to the number of learnings. . The solid line indicates A = 1, the broken line indicates A = 0.1, and the alternate long and short dash line indicates A = 10. As shown in FIG. 10, when A = 10 indicated by the alternate long and short dash line, the learning speed is large, and even if A is larger than 10, no improvement in the learning speed can be expected. When A = 0.1 indicated by a broken line, the learning speed is slightly slow, and if A is smaller than 0.1, learning takes time.

  Therefore, the constant A is preferably in the range of A = 0.1-10.

  Next, the effect by this embodiment is demonstrated based on simulation.

  FIG. 11 shows the output y when the ratio (L / τ) between the time constant τ and the dead time L is 0.01, the target value is 10 to 30, and the steady gain K to be controlled is 100 to 300. The change of (t) is shown, the solid line shows the output of the embodiment, the broken line shows the output of the comparative example in which the learning coefficient η is fixed, and the target value is also shown by a virtual line.

  FIG. 12 shows a case where the ratio (L / τ) between the time constant τ and the dead time L is 0.01, the target value is 1000 to 3000, and the steady gain K to be controlled is 10000 to 30000. The change of the output y (t) is shown, the solid line indicates the output of the embodiment, and the broken line indicates the output of the comparative example in which the learning coefficient η is fixed.

The fixed value of the learning coefficient η in the comparative example is that the ratio (L / τ) between the time constant τ and the dead time L is 0.01, the target value is 100 to 300, and the steady gain K to be controlled. In the case of 1000 to 3000, the value was adjusted by trial and error.
The constant A is A = 1.

  In FIG. 11, both are improved by learning, but it can be seen that the embodiment is faster and better in learning.

  In FIG. 12, in the comparative example of the broken line in which the learning coefficient η is fixed, the learning progresses too much and oscillates, whereas in the solid line embodiment in which the learning coefficient η is calculated according to the calculation formula, there is no oscillation and learning. The improvement effect by is also recognized.

(Embodiment 2)
It is conceivable to match the learning cycle with the control cycle, but in this case, when the learning cycle is shorter than the dead time, learning becomes difficult and oscillation may occur.

  Therefore, in another embodiment of the present invention, the learning cycle is not matched with the control cycle, but is matched with the dead time.

  That is, the period of data correction by the data correction unit 10 described above is made to coincide with the dead time of the control object 2 measured in advance.

  The simulation results of the embodiment in which the learning period is made to coincide with the dead time and the comparative example in which the learning period is made to coincide with the control period (1 second) will be described.

  FIG. 13 shows that the ratio (L / τ) between the time constant τ (= 100) and the dead time L is 0.1, the target value is 100 to 300, and the steady gain K to be controlled is 1000 to 3000. FIG. 4A shows the change in the output y (t) in the case where FIG. 4A shows the output of the embodiment in which the learning cycle Ts is matched with the dead time L, and FIG. 4B shows the learning cycle Ts. The output of the comparative example matched with the fixed value (1 second) which is a control period is shown, respectively, and the target value is also shown with the virtual line.

  FIG. 14 shows a case where the target value and the steady gain K to be controlled are the same as those in FIG. 13, and the ratio (L / τ) between the time constant τ (= 100) and the dead time L is 0.5. The output y (t) of FIG. 4 shows the change of the output of the embodiment in which the learning period Ts coincides with the dead time L, and FIG. 4B controls the learning period Ts. The output of the comparative example matched with the fixed value (1 second) which is a period is shown, respectively.

  In the comparative example in which the learning cycle Ts is made to coincide with the control cycle, the learning cycle Ts becomes shorter than the dead time L and oscillation occurs, as shown in FIGS. 13B and 14B. On the other hand, in the embodiment in which the learning cycle Ts is made to coincide with the dead time L, the learning effect is recognized as shown in FIGS. 13 (a) and 14 (a).

(Embodiment 3)
FIG. 15 is a configuration diagram corresponding to FIG. 1 of a system including a control parameter adjusting device according to another embodiment of the present invention, and parts corresponding to the above-described embodiment are denoted by the same reference numerals. Attached.

  In the above-described embodiment, the adjusting means 4 for adjusting the PID control parameter is incorporated in the temperature controller 1 based on the JIT method. However, in this embodiment, the above-described adjusting means is connected to an external device such as a PLC or a personal computer. The control parameter adjusting device 11 is built in the apparatus, and the adjusting device 11 adjusts the PID control parameter of the general-purpose temperature controller 1 a including the PID controller 3.

  The adjustment device 11 is sequentially provided with data of the set value r from the temperature regulator 1a and the detected temperature y from the controlled object 2 by communication or A / D conversion, while the adjustment device 11 is supplied to the temperature regulator 1a. PID control parameters are sequentially given and adjusted by communication.

  The adjustment operation of the PID control parameter based on the JIT method by the adjustment device 11 is the same as that in the above-described embodiment.

(Other embodiments)
As another embodiment of the present invention, the reference model may be omitted, and the set value r may be used as the state vector instead of the output of the reference model 5.

  In each of the above-described embodiments, description has been made by applying to PID control. However, the present invention is not limited to PID control, but can be applied to adjustment of other control parameters such as GPC (Generalized Predictive Control). Of course.

  In the above-described embodiment, the present invention has been described by applying to a temperature controller that performs temperature control. However, the present invention is not limited to a temperature controller, and various pressures, flow rates, speeds, liquid levels, and the like to be controlled The present invention can be applied to a control device that controls a physical state.

  In the above-described embodiment, the program stored in the ROM of the microcomputer is read and executed. However, as another embodiment of the present invention, for example, a flexible disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic It may be recorded on a tape or the like, and read from these recording media and executed as necessary.

  In each of the above embodiments, a state vector close to the request point has been selected by the neighborhood selection unit, but as another embodiment of the present invention, the neighborhood selection unit is omitted and based on all the state vectors in the database, A characteristic parameter may be generated.

  The present invention is useful for a control device such as a temperature controller.

It is a block diagram of the temperature regulator which concerns on one embodiment of this invention. It is a figure which shows the structure of the data accumulate | stored in the parameter adjustment database of FIG. It is a figure which shows the specific example of the data of FIG. FIG. 2 is a functional block diagram of a PID tuning unit in FIG. 1. It is a figure which shows the time change of reaction temperature. It is a figure which shows the time change of a PID control parameter. It is a figure which shows the change of the output about different constant A. FIG. It is a figure which expands and shows the one part area of FIG. It is a figure which expands and shows the one part area of FIG. It is a figure which shows the change of ISE (square integration error) with respect to the learning frequency about different constant A. FIG. It is a figure which shows the change of the output value of embodiment and the comparative example which fixed the learning coefficient. It is a figure corresponding to FIG. 11 at the time of making a target value and a steady gain differ. It is a figure which shows the change of the output value of embodiment and the comparative example which made the learning period correspond to the control period. FIG. 14 is a diagram corresponding to FIG. 13 when the dead time is changed. It is a block diagram of the control system which concerns on other embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Temperature controller 2 Control object 3 PID controller 4 Adjustment means 5 Reference model 6 Database 7 PID tuning part

Claims (6)

Control means for controlling an object to be controlled in accordance with the control parameter, and a vector including a control system characteristic amount is accumulated as a state vector, and the control means controls the control means based on database data in which a control system characteristic parameter is accumulated. Adjusting means for adjusting the parameters,
The adjusting means is provided with a new state vector, and based on the state vector of the database, a characteristic parameter generating unit that generates a characteristic parameter corresponding to the new state vector, and the characteristic parameter generating unit A characteristic parameter correction unit that corrects the characteristic parameter in a direction in which the control error or the model error is reduced by using a correction coefficient, and the characteristic parameter corrected by the characteristic parameter correction unit is stored in the database There,
The characteristic parameter correction unit calculates the correction coefficient according to a predetermined calculation formula ,
The characteristic parameter of the control system is a control parameter of the control means,
The calculation formula includes a term of the control parameter, and includes at least one of a term of a target value and a term of a control amount,
Further, the calculation formula is such that the correction coefficient is proportional to the control parameter, and the square of the target value, the square of the maximum value of the target value, the square of the control amount, and the maximum of the control amount. The control device is characterized by being inversely proportional to either the square of the value or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount. . The control device according to claim 1 , wherein the characteristic parameter correction unit corrects the characteristic parameter at a period equal to or longer than the dead time of the control target. A device for adjusting the control parameter of a control device that controls a control object according to a control parameter,
A vector including a control system feature amount is accumulated as a state vector, and an adjustment unit that adjusts a control parameter of the control device based on data in a database in which a control system characteristic parameter is accumulated,
The adjusting means is provided with a new state vector, and based on the state vector of the database, a characteristic parameter generating unit that generates a characteristic parameter corresponding to the new state vector, and the characteristic parameter generating unit A characteristic parameter correction unit that corrects the characteristic parameter in a direction that reduces a control error or a model error by using a correction coefficient, and stores the characteristic parameter corrected by the characteristic parameter correction unit in the database. And
The characteristic parameter correction unit calculates the correction coefficient according to a predetermined calculation formula ,
The characteristic parameter of the control system is a control parameter of the control device,
The calculation formula includes a term of the control parameter, and includes at least one of a term of a target value and a term of a control amount,
Further, the calculation formula is such that the correction coefficient is proportional to the control parameter, and the square of the target value, the square of the maximum value of the target value, the square of the control amount, and the maximum of the control amount. A control parameter characterized by being inversely proportional to either the square of the value or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount Adjustment device. The control parameter adjustment device according to claim 3 , wherein the characteristic parameter correction unit corrects the characteristic parameter at a period equal to or longer than the dead time of the control target. A method of adjusting the control parameter of a control device that controls a controlled object according to a control parameter,
A vector containing a control system feature amount is stored in the database as a state vector, and a control system characteristic parameter is stored in the database, and a new state vector is given, so that A step of generating a characteristic parameter corresponding to the new state vector, a correction step of correcting the generated characteristic parameter in a direction in which a control error or a model error is reduced using a correction coefficient, and a corrected characteristic parameter, Accumulating in the database, and adjusting the control parameter based on a characteristic parameter;
In the correction step, the correction coefficient is calculated according to a predetermined calculation formula ,
The characteristic parameter of the control system is a control parameter of the control device,
The calculation formula includes a term for the control parameter, a term for the control parameter, a term for the target value, and a term for the controlled variable.
Further, the calculation formula is such that the correction coefficient is proportional to the control parameter, and the square of the target value, the square of the maximum value of the target value, the square of the control amount, and the maximum of the control amount. A control parameter characterized by being inversely proportional to either the square of the value or the product of the target value or the maximum value of the target value and the control amount or the maximum value of the control amount Adjustment method. The control parameter adjustment method according to claim 5 , wherein a correction period of the characteristic parameter in the correction step is equal to or longer than a dead time of the control target.

Images