KR102215752B1 - Device and method for notch filter parameter setting based on AI - Google Patents

Abstract

There is provided a notch filter parameter setting apparatus applied to a servo system including a controller, a control object, and a notch filter positioned between the controller and the control object to remove a resonance component of a control signal of the controller. The notch filter parameter setting apparatus receives an input signal input to the control object, an output signal output from the control object, and a parameter of the controller, and the servo system based on the input signal, the output signal, and the parameter of the controller. A data preprocessor that calculates the frequency response of and generates preprocessed data; And a parameter setting unit that sets a notch filter parameter through an artificial neural network based on reinforcement learning using the preprocessed data.

Description

Device and method for notch filter parameter setting based on AI

The present invention relates to a method and apparatus for setting a plurality of notch filter parameters by converting the frequency response of a servo system into a form used for machine learning and learning an artificial neural network through a reinforcement learning method.

[Explanation of nationally supported R&D]

This study was made with the support of the World Class 300 (WC300) project of the Ministry of SMEs and Startups (Development of a robot motion control solution for flexible response to smart machines/collaborative robots, task identification number: S2563339).

The servo system refers to a control system in which the displacement of an object's position, orientation, posture, etc. is used as a control amount (output), and the target value (input) is followed by an arbitrary change. A notch filter, which is a filter that outputs a signal excluding a characteristic frequency band component from an input signal, has been applied to such a servo system. In the servo system, a notch filter is applied to the current command to remove the frequency component that causes resonance to reduce the resonance effect of the system.

1 is a block diagram of a conventional servo system 20 to which a notch filter is applied. Referring to FIG. 1, the servo system 20 includes a controller 210, a control object 220, an encoder 230, and a notch filter 240. The controller 210 generates a current command for controlling the control target 220. The encoder 230 measures the rotation angle of the control object 220 and provides the result to the controller 210 as a feedback. Thereafter, the controller 210 controls the control object 220 using the reference and feedback information.

Here, the notch filter 240 filters a current command output from the controller to remove a resonance component corresponding to the notch frequency. A plurality of notch filters may be used to remove a plurality of resonance components according to the characteristics of the system. If the parameters of the notch filter 240 are properly set, the current passing through the notch filter 240 does not cause a resonance phenomenon in the motor. However, when setting the notch filter 240, several parameters must be set per one notch filter, and when setting a plurality of notch filters, the problem becomes more complicated and it is difficult to set an appropriate notch filter.

P. Schmid, and T. Rehm, "Notch filter tuning for resonant frequency reduction in dual inertia systems," in Conference Record of the 1999 IEEE Industry Applications Conference, pp. 1730-1734, 1999.

The present invention has been devised to solve the above-described problem, and provides a method, apparatus and a servo system including the same for setting a parameter of a notch filter based on AI.

A notch filter parameter setting apparatus according to an embodiment of the present invention is applied to a servo system including a controller, a control object, and a notch filter located between the controller and the control object to remove a resonance component of the control signal of the controller. A notch filter parameter setting apparatus, comprising: receiving an input signal input to the control object, an output signal output from the control object, and a parameter of the controller, and the servo based on the input signal, the output signal, and the parameter of the controller A data preprocessor that calculates the frequency response of the system and generates preprocessed data; And a parameter setting unit that sets a notch filter parameter through an artificial neural network based on reinforcement learning using the preprocessed data.

A method for setting a notch filter parameter according to another embodiment of the present invention is applied to a servo system including a controller, a control object, and a notch filter located between the controller and the control object to remove a resonance component of the control signal of the controller. A notch filter parameter setting method of a notch filter parameter setting apparatus, wherein an input signal input to the control object, an output signal output from the control object, and a parameter of the controller are received, and the input signal, the output signal, and the controller Generating preprocessed data by calculating a frequency response of the servo system based on the parameter; And setting a notch filter parameter through an artificial neural network based on reinforcement learning using the preprocessed data.

A computer program according to another embodiment of the present invention is stored in a medium to execute the method of setting the notch filter parameter by being combined with hardware.

The apparatus and method for setting a notch filter parameter according to an embodiment of the present invention calculates a frequency response (Board plot data) using a control target input signal, an output signal, and a parameter of a controller, and uses a DQN machine learning method. The parameters of the notch filter may be automatically set based on the frequency response (Board plot data).

1 is a block diagram of a conventional servo system to which a notch filter is applied.
2 is a block diagram of a servo system according to an embodiment of the present invention.
3 is a block diagram of a parameter setting device according to an embodiment of the present invention.
4 is a block diagram of a parameter setting unit according to an embodiment of the present invention.
5 is a block diagram of a network storage unit according to an embodiment of the present invention.
6 is a flowchart of a method for setting a notch filter parameter according to an embodiment of the present invention.
7 is a flowchart of a step of setting a notch filter parameter according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION The detailed description of the present invention to be described later refers to the accompanying drawings, which illustrate specific embodiments in which the present invention may be practiced. These embodiments are described in detail sufficient to enable a person skilled in the art to practice the present invention. The terms used in the present specification have selected general terms that are currently widely used as possible while considering functions, but this may vary according to the intention or custom of a technician working in the art, or the emergence of new technologies. In addition, in certain cases, there are terms arbitrarily determined by the applicant, and in this case, the meaning will be described in the description of the corresponding specification. Accordingly, terms used in the present specification should be interpreted based on the actual meaning of the term and the entire contents of the present specification, not a simple name of the term.

2 is a block diagram of a servo system according to an embodiment of the present invention. 3 is a block diagram of a parameter setting device according to an embodiment of the present invention. 4 is a block diagram of a parameter setting unit according to an embodiment of the present invention. 5 is a block diagram of a network storage unit according to an embodiment of the present invention.

2 to 5, the servo system 10 according to an embodiment of the present invention includes a parameter setting device 100, a controller 110, a control target 120, an encoder 130, and a notch filter 140. ). The controller 110 generates a current command for controlling the control object 120. The encoder 130 measures the rotation angle of the control object 120 and provides the result to the controller 110 as a feedback signal. Thereafter, the controller 110 controls the control object 120 using the reference and feedback information. The control target 120 is a control target to be controlled by the servo system 10, and corresponds to a control target and a plant. The notch filter 140 filters the current command S1 output from the controller to remove a resonance component corresponding to the notch frequency. That is, a current command from which the resonance component has been removed through the notch filter 140 is provided to the control target 120. The parameter setting device 100 sets the parameters of the notch filter 140 based on AI and provides it to the notch filter 140. The parameter setting apparatus 100 according to an embodiment of the present invention may set and provide a parameter of the notch filter 140 through a reinforcement learning method using an artificial neural network among AI-based methods. The reinforcement learning method using an artificial neural network can select an action that maximizes reward from among the selectable actions while learning the system by itself even in a complex system, so it can more efficiently set the notch filter parameter. However, the AI-based notch filter parameter setting method of the present invention is not limited to the above-described method.

The notch filter 140 according to an embodiment of the present invention may be designed with a transfer function expressed by Equation 1 below.

[Equation 1]

는 라플라스 연산자(Laplace operator),

은 노치 필터 주파수(notch frequency),

는 노치 필터 깊이(notch depth parameter),

는 노치 필터 너비(notch width parameter)이다.)(here,

Is the Laplace operator,

Is the notch filter frequency,

Is the notch depth parameter,

Is the notch width parameter.)

Here, the notch filter frequency, the notch depth parameter, and the notch width parameter are control parameters of the notch filter 140 and correspond to targets for setting the notch filter parameters. That is, the number of control parameters of the notch filter 140 may be three, but is not limited thereto. The number of control parameters of the notch filter 140 may vary according to design conditions.

The parameter setting apparatus 100 uses a notch filter frequency, a notch depth parameter, and a notch width parameter adjusted through a reinforcement learning method using an artificial neural network among AI-based methods. It can be provided as 140. Here, a plurality of notch filters 140 may be provided in the servo system 10, and the parameter setting device 100 may set and provide parameters for each of the plurality of notch filters 140. Hereinafter, the configuration and function of the parameter setting device 100 will be described in more detail.

The parameter setting device 100 includes a data preprocessing unit 101 and a parameter setting unit 102.

The data preprocessor 101 receives a current command from which the resonance component is removed, a feedback signal provided from the encoder 130, and a parameter of the controller 110, and calculates the frequency response of the servo system 100. Here, the current command from which the resonance component is removed is an input signal input to the control object 120, and the feedback signal provided from the encoder 130 corresponds to an output signal output from the control object 120. The data preprocessor 101 receives an input signal of the control object 120, an output signal of the control object 120, and parameters of the controller 110, and a bode plot corresponding to the frequency response of the servo system 100 plot) data can be generated. That is, the parameter setting device 100 may set a parameter of the notch filter 140 by using the Bode plot data of the servo system 100. In addition, the data preprocessor 101 may convert the frequency response into a form used for machine learning. The data preprocessor 101 may further perform an operation of changing the frequency response (Board plot data) to a data size that can be processed by the parameter setting unit 102. For example, when the data size is large, the data preprocessor 101 may use a method of sampling a frequency component and changing it to a data size, or a method of bundling a series of previously used data and processing it as one data. Bode plot data whose data size has been adjusted, that is, pre-processed data, is provided to the parameter setting unit 102.

The parameter setting unit 102 generates a notch filter parameter by using the preprocessed data. The parameter setting unit 102 uses an artificial neural network based on reinforcement learning. For example, DQN can be used. DQN is an algorithm that performs reinforcement learning in a wider state space by adding artificial neural networks to Q-learning technology. DQN estimates a function (Q function) that determines the Q value, rather than storing individual Q values. DQN inputs the current state to the Q function and extracts the Q value as a result value. Here, the current state is the preprocessed data, and the Q value may be a value used to update the notch filter parameter. The parameter setting unit 102 learns a function for determining the Q value determined using the value network for each periodic training period. The parameter setting unit 102 may perform reinforcement learning during the training period using the behavior and compensation values stored in the storage device and update the notch filter parameter.

Specifically, the parameter setting unit 102 includes a network storage unit 1020, a parameter update unit 1022, a compensation calculation unit 1024, a machine learning unit 1026, and a completion condition determination unit 1028.

,

및

를 사용할 수 있다. The network storage unit 1020 receives the preprocessed data and outputs a Q-value used to update the notch filter parameter to the parameter update unit 1022. Output data output from the network storage unit 1020 may be used for an action to be performed by the machine learning unit 1026 to be described later. As a parameter to be output, in Equation 1 per notch filter

,

And

You can use

The network storage unit 1020 may be configured as an artificial neural network. DQN can approximate a function (Q function) that determines a Q value using an artificial neural network of three or more layers. That is, the network storage unit 1020 may approximate a mathematical function having a relationship between an input (preprocessed data) and an output (Q value). 4 is a block diagram illustrating an example of a network storage unit 1020. As shown in FIG. 4, the network storage unit 1020 includes an input layer 1020A, a hidden layer 1020B, and an output layer 1020C. The hidden layer 1020B may be composed of a convolutional layer and a fully connected layer, and may approximate a function (Q function) for determining a Q value using an artificial neural network. The number of the input layer 1020A, the output layer 1020C, and the hidden layer 1020B may be configured differently according to a task to be solved or a function to be approximated. The size of the input layer 1020A is the same as the size of the preprocessed data, and the size of the output layer 1020C is the same as the number of actions that adjust the notch filter parameter.

The parameter update unit 1022 may update the notch filter parameter based on the Q value input from the network storage unit 1020.

)을 계산할 수 있다. 보상값(

)은 기계 학습부(1026)에서 수행할 강화 학습을 위한 데이터일 수 있다. 보상값(

)은 노치 필터 파라미터를 상기 서보 시스템의 주파수 응답에 적용하여 얻을 수 있는 이득 여유, 위상 여유일 수 있으나, 이에 한정되는 것은 아니다.The compensation calculation unit 1024 is based on the notch filter parameter input from the parameter update unit 1022, the compensation value (

) Can be calculated. Compensation value(

) May be data for reinforcement learning to be performed by the machine learning unit 1026. Compensation value(

) May be a gain margin or a phase margin that can be obtained by applying the notch filter parameter to the frequency response of the servo system, but is not limited thereto.

The machine learning unit 1026 may update the artificial neural network by applying a reinforcement learning method in consideration of a compensation value. The machine learning unit 1026 may perform Q-learning represented by Equation 2 below.

<Equation 2>

는 Q 함수,

는 상태(state),

는 행동(action),

은 보상값,

는 할인율(discount factor)이다. 할인율은 0에서 1 사이의 값을 가지며, 0에 가까우면 현재, 1에 가까우면 미래에 대한 보상의 중요성을 강조하는 수치이다.)(here,

Is the Q function,

Is the state,

Is the action,

Is the compensation value,

Is the discount factor. The discount rate has a value between 0 and 1, and when it is close to 0, it is a number that emphasizes the importance of rewards for the present, and 1 for the future.)

가 최대값을 가질 수 있도록 노치 필터 파라미터를 업데이트 하는 것이다. 파라미터 업데이트부(1022)에서

가 최대가 되는 행동(

)를 선택할 수 있으며, 이러한 행동(

)은 노치 필터 파라미터인 노치 필터 주파수(notch frequency), 노치 필터 깊이(notch depth parameter), 노치 필터 너비(notch width parameter) 중 적어도 하나를 조절하는 행동일 수 있다. 즉, 강화 학습 방법은 상기 Q 함수가 최대가 되는 행동을 선택하는 것으로, 상기 행동은 상기 노치 필터 파라미터를 조절하는 행동을 포함할 수 있다. 여기서, 노치 필터가 복수로 구성되는 경우, 복수의 노치 필터 중 적어도 하나의 파라미터를 조절하는 것을 포함할 수 있다. 파라미터 업데이트부(1022)에서 현재 상태(

)에 대한

가 최대가 되도록 노치 필터의 파라미터가 설정될 수 있다.Q-learning is basically a reinforcement learning algorithm composed of environment, agent, state, action, and reward. First, the agent can move to a new state by taking action. The agent is rewarded from the environment for the actions taken. In the end, the agent’s ultimate goal is to receive the maximum reward,

Is to update the notch filter parameter so that it has the maximum value. In the parameter update unit 1022

Is the maximum behavior (

), and this behavior (

) May be an action of adjusting at least one of a notch filter parameter such as a notch frequency, a notch depth parameter, and a notch width parameter. That is, the reinforcement learning method selects an action in which the Q function is maximized, and the action may include an action of adjusting the notch filter parameter. Here, when a plurality of notch filters are configured, it may include adjusting at least one parameter of the plurality of notch filters. In the parameter update unit 1022, the current state (

)for

The parameter of the notch filter may be set so that is maximum.

) 또는 네트워크 저장부(1020)의 업데이트 횟수를 고려할 수 있다. 또한, 완료 조건 판단부(1028)는 완료 조건을 판단하기 위하여 기계 학습부(1026)의 학습 시간, 보상값(

) 또는 네트워크 저장부(1020)의 업데이트 횟수를 저장하는 메모리를 더 포함할 수 있다. The completion condition determination unit 1028 may determine a condition for completing the parameter setting, and if a preset condition is satisfied, complete the parameter setting and provide the currently set notch filter parameter to the notch filter 140. have. In order to determine the completion condition, the completion condition determination unit 1028 includes a learning time of the learning unit 1024 and the compensation value (

) Or the number of updates of the network storage unit 1020 may be considered. In addition, the completion condition determination unit 1028 includes a learning time and a compensation value of the machine learning unit 1026 in order to determine the completion condition.

) Or a memory for storing the number of updates of the network storage unit 1020 may be further included.

)을 다시 계산할 수 있다. 기계 학습부(1026)는 제공받은 보상값(

), Q 함수, 설정된 파라미터를 학습하고, 인공신경망을 업데이트할 수 있다. 기계 학습부(1026)는, 인공신경망을 업데이트하기 위하여 경사 하강법(gradient descent) 혹은 경사 하강법의 변형 방법을 적용할 수도 있다.If the completion condition determination unit 1028 determines that the parameter setting has not been completed, the above-described process may be repeated. That is, the Q function for deriving the currently set parameter may be derived again, and the parameter update unit 1022 may update the notch filter parameter by using the derived Q function. The compensation calculation unit 1024 is based on the currently input notch filter parameter.

) Can be recalculated. The machine learning unit 1026 receives the provided compensation value (

), Q function, and set parameters, and update the artificial neural network. The machine learning unit 1026 may apply a gradient descent method or a transformation method of the gradient descent method to update the artificial neural network.

The parameter setting device 100 according to an embodiment of the present invention calculates a frequency response (Board plot data) using a control target input signal, an output signal, and a parameter of a controller, and uses a DQN machine learning method to perform a frequency response. The parameters of the notch filter can be automatically set based on the (Board plot data).

Hereinafter, a method of setting a notch filter parameter according to another embodiment of the present invention will be described.

6 is a flowchart of a method of setting a notch filter parameter according to another embodiment of the present invention. 7 is a flowchart of a step of setting a notch filter parameter according to another embodiment of the present invention.

The notch filter parameter setting method according to the present embodiment may be a parameter setting method applied to the notch filter parameter setting apparatus 100 described in FIGS. 2 to 5 described above. That is, it corresponds to a parameter setting method of a parameter setting device applied to a servo system including a controller, a control target, and a notch filter located between the controller and the control target to remove the resonance component of the control signal of the controller. For description of the embodiment, the apparatus for setting parameters of the notch filter of FIGS. 2 to 5 may be referred to.

6 and 7, a method for setting a notch filter parameter according to another embodiment of the present invention includes the step of generating preprocessed data by calculating the frequency response of a servo system (S100) and through an artificial neural network based on reinforcement learning. And setting the notch filter parameter (S110).

First, the frequency response of the servo system is calculated to generate preprocessed data (S100).

The step S100 of generating preprocessed data by calculating the frequency response of the servo system may be performed in the data preprocessor 101 of the parameter setting device. The data preprocessor 101 receives an input signal of the control object 120, an output signal of the control object 120, and parameters of the controller 110, and a bode plot corresponding to the frequency response of the servo system 100 plot) data can be generated. The data preprocessor 101 may convert the frequency response into a form used for machine learning. The data preprocessor 101 may further perform an operation of changing the frequency response (Board plot data) to a data size that can be processed by the parameter setting unit 102. For example, when the data size is large, the data preprocessor 101 may use a method of sampling a frequency component and changing it to a data size, or a method of bundling a series of previously used data and processing it as one data.

A notch filter parameter is set through an artificial neural network based on reinforcement learning (S110).

The step (S110) of setting the notch filter parameter through an artificial neural network based on reinforcement learning may be performed by the parameter setting unit 102. The parameter setting unit 102 may set a notch filter parameter using a DQN, an artificial neural network based on reinforcement learning. The parameter setting unit 102 learns a function (Q function) for determining a determined Q value using an artificial neural network for each periodic training period. The parameter setting unit 102 may perform reinforcement learning during the training period using the behavior and compensation values stored in the storage device and update the notch filter parameter.

Specifically, in the step of setting the notch filter parameter through an artificial neural network based on reinforcement learning (S110), receiving preprocessed data, outputting a Q value, and updating the notch filter parameter using the Q value (S112) ; Calculating a compensation value based on the notch filter parameter (S114); Updating the artificial neural network through a reinforcement learning method in consideration of the compensation value (S116); And determining whether the setting of the notch filter parameter is completed (S118).

The step of receiving the preprocessed data, outputting a Q value, and updating the notch filter parameter using the Q value (S112) is performed in the network storage unit 1020 and the parameter update unit 1022. The network storage unit 1020 may be configured as an artificial neural network. DQN can approximate a function (Q function) that determines the value of Q by using a value network of three or more layers. That is, the network storage unit 1020 may estimate a Q function, which is a mathematical function having a relationship between an input (preprocessed data) and an output (Q value). The parameter update unit 1022 may update the notch filter parameter based on the Q value input from the network storage unit 1020.

)은 강화 학습을 위한 데이터일 수 있다. 보상값(

)은 노치 필터 파라미터를 상기 서보 시스템의 주파수 응답에 적용하여 얻을 수 있는 이득 여유, 위상 여유일 수 있으나, 이에 한정되는 것은 아니다.The step of calculating the compensation value based on the notch filter parameter (S114) is performed by the compensation calculation unit 1024. Compensation value(

) May be data for reinforcement learning. Compensation value(

) May be a gain margin or a phase margin that can be obtained by applying the notch filter parameter to the frequency response of the servo system, but is not limited thereto.

)에서 Q 값이 최대가 되는 행동을 선택하는 것으로, 상기 행동은 상기 노치 필터 파라미터를 조절하는 행동을 포함할 수 있다. 여기서, 노치 필터가 복수로 구성되는 경우, 복수의 노치 필터 중 적어도 하나의 파라미터를 조절하는 것을 포함할 수 있다. The step of updating the artificial neural network (S116) through the reinforcement learning method considering the compensation value is performed by the machine learning unit 1026. The machine learning unit 1026 may update the artificial neural network by applying a reinforcement learning method in consideration of a compensation value. The reinforcement learning method is the characteristic state (

) To select an action having a maximum Q value, and the action may include an action of adjusting the notch filter parameter. Here, when a plurality of notch filters are configured, it may include adjusting at least one parameter of the plurality of notch filters.

) 또는 네트워크 저장부(1020)의 업데이트 횟수를 고려할 수 있다. 또한, 완료 조건 판단부(1028)는 완료 조건을 판단하기 위하여 기계 학습부(1026)의 학습 시간, 보상값(

) 또는 네트워크 저장부(1020)의 업데이트 횟수를 저장하는 메모리를 더 포함할 수 있다. 완료 조건 판단부(1028)에서 파라미터 설정이 완료되지 않은 것으로 판단되는 경우, 상술한 과정이 반복될 수 있다. 즉, 현재 설정된 파라미터를 도출하기 위한 Q 함수를 다시 도출할 수 있으며, 파라미터 업데이트부(1022)는 다시 도출된 Q 함수를 이용하여 노치 필터 파라미터를 업데이트 할 수 있다. 보상 계산부(1024)는 현재 입력받은 노치 필터 파라미터를 바탕으로 보상값(

)을 다시 계산할 수 있다. 기계 학습부(1026)는 제공받은 보상값(

), Q 함수, 설정된 파라미터를 학습하고, 인공신경망을 업데이트할 수 있다. 기계 학습부(1026)는, 인공신경망을 업데이트하기 위하여 경사 하강법(gradient descent) 혹은 경사 하강법의 변형 방법을 적용할 수도 있다.The step of determining whether the setting of the notch filter parameter is completed (S118) is performed by the completion condition determining unit 1028. The completion condition determination unit 1028 may determine a condition for completing the parameter setting, and if a preset condition is satisfied, complete the parameter setting and provide the currently set notch filter parameter to the notch filter 140. have. In order to determine the completion condition, the completion condition determination unit 1028 includes a learning time of the learning unit 1024 and the compensation value (

) Or the number of updates of the network storage unit 1020 may be considered. In addition, the completion condition determination unit 1028 includes a learning time and a compensation value of the machine learning unit 1026 in order to determine the completion condition.

) Or a memory for storing the number of updates of the network storage unit 1020 may be further included. If the completion condition determination unit 1028 determines that the parameter setting has not been completed, the above-described process may be repeated. That is, the Q function for deriving the currently set parameter may be derived again, and the parameter update unit 1022 may update the notch filter parameter by using the derived Q function. The compensation calculation unit 1024 is based on the currently input notch filter parameter.

) Can be recalculated. The machine learning unit 1026 receives the provided compensation value (

), Q function, and set parameters, and update the artificial neural network. The machine learning unit 1026 may apply a gradient descent method or a transformation method of the gradient descent method to update the artificial neural network.

The operation by the parameter setting method of the notch filter according to the embodiments described above may be implemented at least partially as a computer program and recorded on a computer-readable recording medium. A computer-readable recording medium in which a program for implementing an operation according to the parameter setting method of a notch filter according to the embodiments is recorded and includes all types of recording devices in which data that can be read by a computer is stored. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tapes, floppy disks, and optical data storage devices. In addition, the computer-readable recording medium may be distributed over a computer system connected through a network, and computer-readable codes may be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing this embodiment may be easily understood by those skilled in the art to which this embodiment belongs.

Although the above has been described with reference to the embodiments, the present invention should not be construed as being limited by these embodiments or drawings, and those skilled in the art will have the spirit and scope of the present invention described in the following claims. It will be understood that various modifications and changes can be made to the present invention within the range not departing from.

10: servo system
100: parameter setting device
101: data preprocessor
102: parameter setting unit
1020: network storage
1022: parameter update unit
1024: compensation calculation unit
1026: Machine Learning Department
1028: Completion condition determination unit
110: controller
120: control target
130: encoder
140: notch filter

Claims (11)

A notch filter parameter setting apparatus applied to a servo system including a controller, a control object, and a notch filter positioned between the controller and the control object to remove a resonance component of a control signal of the controller,
Receives an input signal input to the control object, an output signal output from the control object, and a parameter of the controller, and calculates a frequency response of the servo system based on the input signal, the output signal, and parameters of the controller. A data preprocessor for generating preprocessed data; And
A parameter setting unit that sets a notch filter parameter through an artificial neural network based on reinforcement learning using the preprocessed data,
The parameter setting unit,
A network storage unit configured with an artificial neural network for receiving the preprocessed data, outputting a Q value for updating the notch filter parameter, and estimating a Q function for determining the Q value;
A parameter update unit for updating a notch filter parameter based on a Q value provided from the network storage unit;
A compensation calculation unit that calculates a compensation value based on the notch filter parameter provided by the parameter update unit;
A machine learning unit that updates the artificial neural network through a reinforcement learning method in consideration of the compensation value; And
A completion condition determination unit determining a condition for completing the setting of the parameter, completing the setting of the parameter when a preset condition is satisfied, and providing a parameter of a currently set notch filter to the notch filter,
The notch filter includes a notch filter frequency, a notch depth parameter, and a notch width parameter as the notch filter parameters,
The reinforcement learning method is to select an action in which the Q function is maximized, and the action comprises adjusting at least one of the notch filter frequency, the notch filter depth, and the notch filter width. Parameter setting device. delete delete delete The method of claim 1,
The notch filter is configured in plural in the servo system,
Wherein the notch filter parameter setting device sets parameters of each of the plurality of notch filters. A notch filter parameter setting method of a notch filter parameter setting device applied to a servo system including a controller, a control object, and a notch filter positioned between the controller and the control object to remove a resonance component of a control signal of the controller,
Receives an input signal input to the control object, an output signal output from the control object, and a parameter of the controller, and calculates a frequency response of the servo system based on the input signal, the output signal, and parameters of the controller. Generating preprocessed data; And
And setting a notch filter parameter through an artificial neural network based on reinforcement learning using the preprocessed data,
The step of setting the notch filter parameter,
An artificial neural network receiving the preprocessed data and outputting a Q value, the artificial neural network estimating a Q function that determines the Q value;
Updating the notch filter parameter using the Q value;
Calculating a compensation value based on the notch filter parameter;
Updating the artificial neural network through a reinforcement learning method in consideration of the compensation value; And
Determining whether the setting of the notch filter parameter is complete,
The determining whether the setting of the notch filter parameter is complete may include determining a condition for completing the setting of the parameter, and if a preset condition is satisfied, setting the parameter is completed and the parameter of the currently set notch filter is transferred to the notch filter. And
The notch filter includes a notch filter frequency, a notch depth parameter, and a notch width parameter as the notch filter parameters,
The reinforcement learning method selects an action in which the Q function is maximized, and the action comprises adjusting at least one of the notch filter frequency, the notch filter depth, and the notch filter width. delete delete delete The method of claim 6,
The notch filter is configured in plural in the servo system,
The notch filter parameter setting apparatus sets a parameter of each of the plurality of notch filters. A computer program stored in a medium to be combined with hardware to execute the notch filter parameter setting method according to any one of claims 6 and 10.

Images