JPH0573104A - Method for automatically tuning pid parameter - Google Patents

Abstract

PURPOSE:To present the transfer function of a controlled system and to finally unnecessitate the addition of test signals by a learning function in the case of changing parameters by adaptive operations. CONSTITUTION:While containing a PID loop defining a P (proportion) operation, I (integration) operation and D (differentiation) operation as the basic elements of control operations, this method is equipped with a process parameter calculation part 12 and a PID parameter calculating mechanism to decide the parameters of the respective PID operations based on calculated process parameters, and the process parameter calculation part 12 is composed of a transfer function tuning part to tune the transfer functions of processes by observing the responded waveforms of a control system when stepwise inputs are impressed to the control system, and a gain scheduling part to describe relation between various environmental parameters and process parameters, selects one of the process parameters outputted from both parts, and calculates the respective parameters of the PID operations.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a PID parameter tuning method for a PID controller which is often used for process control of chemical plants, power plants, water and sewage plants and the like.

[0002]

2. Description of the Related Art Conventionally, the following methods have been known as PID parameter auto-tuning methods. (1) Method based on adaptive control theory such as STR or MRACS (2) Method for observing response waveform of process and determining P, I, D parameters by rule base Proceedings, second
Volume 0 Volume 7 p20-p27 "Closed-loop type auto-tuning method for digital process control system", Japanese Patent Publication No. 63-
65964 and the like. These cited examples are methods based on STR, and the pulse transfer function is identified by estimating the parameters of the model using the ARMA model as the target process. In the above cited example, by further converting the pulse transfer function to the transfer function, the PID
The parameter tuning itself is reduced to the idea of a continuous system. In these adaptive control theoretical approaches, adaptive control theory makes it applicable.
The linearity, time invariance, and continuous excitation conditions of the target system are
In an actual application example, it is hard to say that it is always established, and the application range is naturally limited. In the method (2), the step response or the response waveform of the process at the time of adding noise is pattern-recognized, and the parameters of the P, I, and D operating elements are changed according to the rule stored in advance. It goes. As a prior art of this method, for example, in "Measurement Technology", September 1986, p66 to p72, a rule base is constructed using an expert system.
47801 and JP-A-62-241006 use fuzzy logic to describe rules. In these methods, since the parameters of the PID controller are directly calculated, the control system designer cannot be informed of the process because the transfer function of the process is not explicitly shown. The parameter is
It is difficult to know whether it is appropriate or not, and it is difficult to give useful information to the designer when improving the control system. Further, a problem common to (1) and (2) is that some kind of test signal needs to be added when tuning is performed. That is, in (1), in order to operate the adaptive function, it is necessary to add an M-sequence pulse signal, and in (2), a pulse-like or step-like test signal or a change in set value and noise addition are expected. is doing. It doesn't matter during the plant commissioning,
Once the actual operation is started, it is not preferable to add the test signal, and the timing for tuning and the timing for changing the set value or adding noise are not always the same.

[0003]

SUMMARY OF THE INVENTION The object of the present invention is to transfer the controlled object, which is useful information for an engineer, by relaxing the application limit as in the case of using the adaptive control in the above-mentioned prior art. In addition to making it possible to present a function, the learning function ultimately makes it unnecessary to add a test signal when changing the parameters by the adaptive operation.

[0004]

[Means for Solving the Problems] In order to solve the above-mentioned problems, the response of the control system is not taken during the trial run or in the initial stage of the run without taking a mathematical model approach as in the case of using the adaptive control. The transfer function is identified by observing the waveform, and the relationship between the environment variable and the process parameter is learned by the learning function. In actual operation, the time-varying characteristics of the controlled object are matched without adding a special test signal. Fit PID parameters. Also,
Instead of directly calculating the PID parameter from the response waveform, the PID parameter is divided into a part for identifying the transfer function of the controlled object and a part for calculating the PID parameter using the above transfer function. When the transfer function is identified by the response waveform, the transfer function is G
(S) = 1 / (1 + a ₁ s + a ₂ s ² + a ₃ s ³ + a ₄
In the system represented by s ⁴ + ...), the rise time when a step response waveform is taken is determined by the first-order term of s, and the slope at the time of rise is determined by the second-order terms of s, That is, the feature quantity of the step response waveform is the coefficients a ₁ , a ₂ , a ₃ , a ₄ of the denominator sequence of the above transfer function.
The transfer function is identified by recognizing the relationship between various response waveforms and the above-mentioned characteristic amount by utilizing the fact that the above appears.

[0005]

In the test run, the time series data of the input layer and the PID of the output layer are used to identify the closed loop transfer function by using the time series (step response) of the measured values when the set value is changed. Closed loop transfer function including controller and plant G (s) = 1 / (1 + a ₁ s + a ₂ s ² + a ₃ s ³
+ A ₄ · s ⁴ + ...) Coefficients a ₁ , a ₂ , a ₃ , of the denominator series
A neural network such as a ₄ ... Is used as the transfer function identifying unit, and the identified transfer function is used to transfer the transfer function Gp (s) = 1 / (c _{1 of} the process to be controlled.
+ C ₂ · s + c ₃ · s ² ) process parameter c ₁ ,
Calculate c ₂ and c ₃ . The PID parameter calculation mechanism determines the PID parameter based on the obtained Gp (s).
Also, the environment variables p1 to pm and the process parameter c ₁ ,
The relationship between c ₂ and c ₃ is the neural network (gain scheduling unit) with the former as the input layer and the latter as the output layer.
The process parameters c ₁ , c ₂ and c ₃ described in _{1 above} and identified at the time of test operation are learned as learning data.
The correlation between the environment variable and the process parameter is obtained by calculating the causality measure so that the validity of the learning of the gain scheduling unit can be judged. When the learning result of the gain scheduling unit is good in terms of the causality scale during actual operation, the neural network of the gain scheduling unit uses the process parameter calculated from the value of the environment variable to set the PID parameter to PID.
It is calculated by the parameter calculation mechanism. The neural network of the transfer function identification unit and the gain schedule unit has an identification mode and a learning mode. The transfer function identification unit, in the learning mode, has a step response waveform represented by the transfer function of G (s) and the coefficients a ₁ , a ₂ ,
By learning the relationship of a ₃ , a ₄ ... And by inputting the time series data when the set value is changed in the identification mode, a ₁ , a ₂ , a ₃ , a ₄ ...
Ask for ... Also, in the learning mode, the gain schedule section uses the above process parameters c ₁ , c ₂ , c
Learn the relationship between ₃ and environment variables p1 to pm. Conversely, in the identification mode, the process parameter is obtained from the environment variable. In this way, by self-learning the relationship between the process parameter to be controlled and the environment variable that affects the process parameter, the PID parameter is adjusted adaptively to changes in the process characteristics, and good control performance is always obtained. Be done. Further, by dividing the part for identifying the transfer function of the controlled object and the part for calculating the PID parameter using this transfer function, it is possible to present the transfer function of the controlled object to the designer and improve the control system. It can be basic data. In addition, the learning function makes it unnecessary to add a test signal when changing the parameters by the adaptive operation.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings. FIG. 1 is an overall configuration diagram of an embodiment of the present invention. Reference numeral 11 denotes a cyclic buffer, which stores the latest n sample periods of data of the measured PV of the control target when the transfer function identification unit is in the identification mode, and learns when the transfer function identification unit is in the learning mode. Step response waveforms of various patterns are stored as data for n sample periods. The input to the cyclic buffer 11 is switched by the changeover switch 13. The changeover switch 13 inputs the measured value PV and the learning data (step response waveform), and outputs the measured value PV in the identification mode and the learning data in the learning mode. 12 is a process parameter calculating unit, the time series data of the cyclic buffer 11, enter the environment parameters P1 to Pm, calculates the process parameters _{c 1, c 2, c 3} . Reference numeral 15 is a PID parameter calculation mechanism using a partial model matching method or the like,
The identified process parameters c ₁ , c ₂ , and c ₃ are input, and PID parameters Kp (proportional gain), Ti (integral time), and Td (differential time) are calculated. At the time of calculation, σ, α1, α2, and α3 are set as set values, which will be described later. Kp, Ti, and Td calculated by the PID parameter calculation mechanism 15 are PID control loop 1
Set to 6.

FIG. 2 shows a block diagram of the control loop in this embodiment. In this embodiment, the PID
The control unit 16 employs the I-PD method, which uses the idea of feedback compensation, which is a modification of PID. 21
Is a process to be controlled. In this embodiment, the process model is a secondary system, that is, Gp (s) = 1 /
(C ₁ + c ₂ s + c ₃ s ² ). Because I
In the PD control, the PID parameter can be determined up to the second-order term of the process, and conversely, only the second-order term can be compensated, so that no more terms are needed. In process control, practically, there is no problem in most cases, assuming that the control target is the secondary system. In FIG. 2, 22 is a proportional gain, 23 is a proportional term +
Corresponding to the differential term, 24 becomes the integral term. In addition, Kp, T
It is assumed that i and Td have initial values before starting PID parameter tuning.

FIG. 3 shows a detailed configuration of the process parameter calculation unit 12. Reference numeral 31 denotes a transfer function identification unit that, in the identification mode, inputs measured values for n sample periods stored in the cyclic buffer 11 to the input layer of the neural network as input data and outputs the neural network as an output. The denominator series coefficients a ₁ , a ₂ , and a ₃ of the closed-loop transfer function G (s) obtained from the output layer are output. In the learning mode, conversely, the denominator sequence coefficients a ₁ , a ₂ and a ₃ of G (s) are input as learning data 37 from the output side to the output layer of the neural network and input from the cyclic buffer 11. Further, the transfer function G (s) using the above denominator series coefficient = 1 / (1 + a ₁
・ S + a ₂ · s ² + a ₃ · s ³ ) step response waveform is used for learning. It is assumed that this learning is performed before starting the PID parameter tuning. 32 is
A closed-loop transfer function G (s) = 1 / (1 + a ₁ ···) which is a process parameter calculation function and which includes the PID control unit 16 and the process identified by the transfer function identification unit 31.
s + a ₂ · s ² + a ₃ · s ³ ) is a process of calculating the transfer function Gp (s) = 1 / (c ₁ + c ₂ · s + c ₃ · s ² ) of the process to be controlled. The calculation formula of this process is given by the following three formulas. c ₁ = {(Kp · a ₁ ) / Ti} -Kp c ₂ = {(Kp · a ₂ ) / Ti} -Kp · Td c ₃ = (Kp · a ₃ ) / Ti 33 is an identification control function. Yes, the setting value SV of the PID control loop is input, and it is checked whether the setting value has been changed. When the setting value is changed, that is, when it is determined that a step-like input is added to the closed loop system including the process and the PID control loop, the identification control is performed so that the transfer function identification unit 31 starts the identification process. The function operates.
34 is a gain scheduling unit, in the identification mode, the input layer of the environment variables P1 to Pm, are constituted by the process parameters c _1, c _2, c ₃ Neural Network as an output layer. In the learning mode, learning is performed by using the process parameters obtained by the transfer function identifying unit 31 and the process parameter calculating function 32 as learning data and using the values of the respective environment variables at the time when the transfer function identifying unit 31 identifies the transfer function. .. Here, since the environment variables are not always clear at the start of PID parameter tuning, all possible possibilities are input to the input layer. Which environment variable affects the process parameter is determined by the causality scale calculated by the causality scale calculation function 35. In the identification mode, the process parameter is calculated from the current value of each environment variable from the learned relationship between the environment variable and the process parameter. Here, the causality measure of the learning mode is presented to the designer, and the process parameter determined to have low correlation with the environment variable by the designer can be omitted from the input layer. The method of calculating the causality scale will be described later. 36
Is a changeover switch, and the gain scheduling unit 3
4 or the process parameter calculated by the transfer function identification unit 31 or the process parameter calculation function 32 is selected and passed to the PID parameter calculation mechanism 15.

FIG. 4 schematically shows the structure of the neural network used in the transfer function identifying section 31 and the gain scheduling section 34. 41
Is an output unit, and in the transfer function identification unit 31,
Closed loop transfer function G (s) = 1 / (1 + a ₁ · s + a ₂ ·
The parameters (a ₁ , a ₂ , a ₃ ) of (s ² + a ₃ · s ³ ) are made to correspond to the respective output units, and the gain scheduling unit 3
4, the process transfer function Gp (s) = 1 / (c ₁ + c ₂
・ S + c ₃・ s ² ) parameters are made to correspond to each output unit (that is, the number of output units is 3). 42
Is an input unit, and in the transfer function identification unit, it corresponds to each one of the time series data stored in the cyclic buffer 11 (therefore, the number of input units is n. In FIG. In the gain scheduling unit 34, the environment variables p1 to pm are associated with each other (therefore, the number of input units is m). 43 is a hidden layer unit,
The designer can arbitrarily select the number of units, and the larger the number of units, the more complex the network can be configured. 44
Is a connection between the input layer and the hidden layer, and each unit of the input layer is connected to each unit of the hidden layer. Similarly, 45 is a connection between the hidden layer and the output layer, which is connected from each unit of the hidden layer to each unit of the output layer. This neural network uses a Rumelhart type network, and there is no mutual connection in the same layer of each of the input layer, the hidden layer, and the output layer. Further, in this example, a neural network having a three-layer structure is used, but a hidden layer can be increased and a network having four or more layers can be handled in the same manner. The connections 44 and 45 all have weighting factors, and learning the neural network is nothing but determining the values of these weighting factors.

Next, the algorithm of this neural network will be described. FIG. 5 shows one unit in FIG. 51 is the i-th unit,
52a, 52b, and 52c are unit k to unit i, unit j to unit i, and unit 1 to unit i, respectively.
, 53 indicates the output of unit i. Unit i
The operation of is expressed by the equation: Oi = f (neti) = f (ΣWij · Oj) (1) Where Oi is the output of unit j,
Wij is the weighting factor of the connection from unit i, f (x)
Is a sigmoid function, and f (x) = 1 / (1 + ex
It is represented by p (-x)). In identification mode,
By performing the operation represented by the equation (1) on all units from the input layer to the output layer, the output (a ₁ ,
a ₂ , a ₃ ) is calculated. On the other hand, in the learning mode, the back-propagation algorithm based on the delta rule is used to calculate all weighting factors W in the network.
ij can be calculated. In backpropagation, the variation ΔWij per trial of the weighting factor of each connection is given by the following equation. ΔWij = ηδi · Oj (2) where η is a constant, δi is the deviation from the training data during learning, and Oi is the output of the i-th unit. Note that δi differs depending on whether the i-th unit is the output layer or the hidden layer. If it is a unit of the output layer, δi = (ti−Oi) · f ′ (neti) (3)
(Ti is teacher data of the unit i), and in the case of a hidden layer unit, δi = f ′ (neti) · Σk (δk · Wki) (4) The calculation expressed by the equations (2) to (4) is performed from the output layer to the input layer by 1 for all the teacher data.
If you follow along, it will be one trial. Learning is performed by repeating this trial until Wij converges. For example, the transfer function identification unit 31 uses various cases of the denominator sequence (a ₁ , a ₂ , a ₃ ) of G (s) as the teacher data ti.
The input to the input layer at this time is the transfer function G (s) = 1 /
The step response of the system is represented by (1 + a ₁ s + a ₂ s ² + a ₃ s ³ ).

Next, the calculation method of the causality scale will be described. The causality measure is defined by the following equation when the hidden layer is one layer. CO = Σj (W1kj · W2ji) (5) where CO is the input unit i with respect to the output unit k
, W1kj is a weight coefficient between the output unit k and the hidden layer unit j, and W2ji is a weight coefficient between the hidden layer unit j and the input unit i. The magnitude of the absolute value of CO indicates the magnitude of the correlation, and the sign indicates whether the correlation is positive or negative. The causality measure calculating function 35 calculates the equation (5) for all input / output units.

Finally, the PID parameter calculation mechanism 15
The processing will be described. Process parameter calculation unit 12
The process transfer function Gp (s) was identified at
Therefore, the subsequent calculation of PID parameters is performed by various known
The delivery method can be applied. In this application example,
PID parameters are calculated by the Dell matching method
It That is, the transfer function showing the desired response is Gm
(S) = 1 / (1 + σ · α₁・ S + σ²・ 0₂・ S²+ Σ³
・ Α₃・ S³), The closed loop transfer function G (s) is
Calculate PID parameters to match m (s)
It Here, σ is the rise time. σ, α₁, Α₂, Α
₃Is a PID parameter for defining a desired control system.
This is the set value for the data calculation mechanism. The PID parameters are
It is calculated by a formula. (Note that in this embodiment,
Me α₁= 1. ) Kp = {c₃/ (Α₃・ Σ²)}-C₁  Ti = (1-α₃・ C₁・ Σ²/ C₃) Σ Td = (c₃・ Α₂-C₂・ Α₃・ Σ) σ / (c₃-Α₃・ C₁・ Σ²) In addition, this PID parameter auto-tuning method
Is realized by the above algorithm.
For control computer and digital instrumentation system
Easily on top of a loop or multi-loop controller etc.
Can build

FIG. 6 shows an example in which the present invention is applied to DO (dissolved oxygen content) control in a sewage treatment process (activated sludge method). In the figure, reference numeral 61 is an aeration tank in which organic substances in wastewater in the tank are treated by bacteria. This process is performed by supplying oxygen to the bacteria, and therefore is accelerated by supplying air from the blower 62 to the aeration tank 61 through the pipe 63. The DO control is to control the dissolved oxygen amount in the aeration tank 61 to a constant value by the control valve 64. In the present embodiment, the DO loop 65 in which the measured value of the DO meter 67 is PV is used as the primary adjustment system, and the measured value of the air flow meter 68 is used as the secondary adjustment system in PV.
Is configured as a cascade system of two PID loops 66. In the PID loops 65 and 66, the P
Apply the ID auto-tuning method. Since the sewage treatment process includes microbial treatment, it is known that its dynamic characteristics fluctuate depending on factors such as liquid temperature, microorganisms, and the state of influent sewage. Therefore, the fluctuation of the dynamic characteristics can be considered as a change of the gain and the time constant in the construction of the control loop. Absorbance 601 of
The gain scheduling unit 34 learns the relationship between these measured values, the gain, and the time constant. In this learning, the transfer function identification unit 31 uses the process parameters obtained by the process parameter calculation function 32 as learning data.
Liquid temperature at the time when the transfer function was identified by MLSS69,
All the values of the absorbance 601 of the inflowing wastewater are input to the input layer for learning. Then, which environment variable influences the process parameter is determined by the causality scale calculated by the causality scale calculation function 35. In the identification mode, the process parameter is calculated from the current value of each environment variable from the learned relationship between the environment variable and the process parameter. Since the secondary adjustment system can be regarded as having no fluctuation in the dynamic characteristics, the gain scheduling unit 34 is not used,
The PID parameter is determined only by the transfer function identification unit 31. In this embodiment, since the transfer function of the plant can be determined from the MLSS, the liquid temperature, and the absorbance after the end of learning at the beginning of the plant operation, the PID parameter calculation mechanism 15
However, without adding a special test signal, the PID parameter can be changed by following the fluctuation of the dynamic characteristics, and a good control performance can always be obtained.

[0014]

According to the present invention, since the transfer function is identified by the pattern recognition approach of the step response waveform during the plant commissioning using the neural network, it is difficult to apply the conventional adaptive control approach. It is possible to adapt the PID parameters even to time deformation and non-linear systems, and to learn the relationship between environmental variables and process parameters during test operation. By the same idea as the gain scheduling method, good control performance can be realized by adaptive fitting of PID parameters without adding a test signal that is not preferable for operation.
In addition, since the transfer function of the controlled object can be presented to the designer by dividing it into a part for identifying the transfer function of the controlled object and a part for calculating the PID parameter using this transfer function, it is a basis for improving the control system. It can be data. In addition, since the gain scheduling part has a causality measure calculation function, the causality measure of the learning mode can be presented to the designer, and the process parameter determined by the designer to have a low correlation with the environment variable is omitted from the input layer. be able to.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of a PID parameter / order tuning method of the present invention.

FIG. 2 is a block diagram of a control system

FIG. 3 is a configuration of a process parameter calculation unit of the present invention

[Fig. 4] Schematic diagram of neural network

FIG. 5: Individual unit of neural network

[Fig. 6] D of the sewage treatment process (activated sludge method) according to the present invention
Example applied to O (dissolved oxygen content) control

[Explanation of symbols]

 11 cyclic buffer 12 process parameter calculation unit 13 changeover switch 15 PID parameter calculation mechanism 16 PID control loop

Claims (5)

[Claims] 1. A process control method including a PID control loop having a proportional action (P action), an integral action (I action), and a derivative action (D action) as basic elements of the control action.
And a process parameter calculation unit, and a PID parameter calculation mechanism that determines the parameters of each of P operation, I operation, and D operation based on the calculated process parameter.
The process parameter calculation unit includes a transfer function identification unit that identifies a process transfer function by observing a response waveform of the control system when a stepwise input is applied to the control system, and various environment. The gain scheduling unit describes the relationship between variables and process parameters, selects one of the process parameters output by both, and based on the output of either one, the parameters of P operation, I operation, and D operation. A PID parameter auto-tuning method characterized by: 2. The transfer function identification unit according to claim 1,
When a step-like set value change is applied to the control system,
It has a function of automatically starting a tuning operation for identifying a closed loop transfer function including a PID controller and a controlled object, and when identifying the closed loop transfer function, a time series of the latest n sample periods of process measurement values. Is an input layer, and the coefficient of the closed loop transfer function is an output layer, and the PID parameter auto-tuning method is configured by a neural network having at least one hidden layer. 3. The closed transfer function G (s) = 1 / (1 + a ₁ · s) according to claim 1 or 2, wherein the transfer function identification unit includes an identification target including a PID controller and a control target.
+ A ₂ · s ² + a ₃ · s ³ ), the latest n sample periods of the process measurement value are used as the input layer, and the coefficients a ₁ , a ₂ , a of the denominator sequence of the closed-loop transfer function G (s) are used. ₃ is an neural network having at least one hidden layer as an output layer, and the strength of the coupling between the units forming the neural network is the coefficient a ₁ of the denominator sequence of the closed loop transfer function G (s), A PID parameter auto-tuning method characterized in that it is determined by learning step response waveforms of various combination patterns of a ₂ and a ₃ . 4. The gain scheduling unit according to claim 1, wherein an environment variable affecting a process parameter is used as an input layer, and a process transfer function of a control target is Gp (s) =
A neural network having at least one hidden layer, the output parameters of which are process parameters c ₁ , c ₂ , and c ₃ on the assumption that they are represented by 1 / (c ₁ + c ₂ · s + c ₃ · s ² ). The PID parameter auto-tuning method, wherein the training data of the neural network is the process parameter identified by the transfer function identification unit. 5. The gain scheduling section according to claim 1, wherein the gain scheduling section has a function of calculating a causality measure for evaluating a correlation between an environment variable and a process parameter.
ID parameter auto tuning method.