KR-20260067428-A - Motion Stage Control Gain Tuning System and Method and Recording Medium thereof

KR20260067428AKR 20260067428 AKR20260067428 AKR 20260067428AKR-20260067428-A

Abstract

The present invention relates to a motion stage control gain tuning system and method and a recording medium thereof, and more specifically, to a system and method capable of efficiently and precisely deriving an optimal gain value by generating a machine learning-based model and comparing it with a target response characteristic based on an optimization algorithm, and a recording medium thereof. A motion stage control gain tuning system according to an embodiment of the present invention comprises a processor that acquires response characteristics of a motion stage according to PID gain values, generates and trains a machine learning-based model, and derives an optimal gain value through comparison with a target response characteristic based on an optimization algorithm.

Inventors

박재현
구정인
이학준
곽다솜

Assignees

한국생산기술연구원

Dates

Publication Date: 20260513
Application Date: 20241104

Claims (16)

A processor that acquires the response characteristics of a motion stage according to PID gain values, creates and trains a machine learning-based model, and derives an optimal gain value through comparison with a target response characteristic based on an optimization algorithm; A motion stage control gain tuning system including
In paragraph 1, The above processor is, A data unit that collects response characteristic data based on changes in PID gain values through motion stage driving, equalizes the learning influence of each feature through data scaling in preprocessing, and removes outliers from the acquired data using an outlier removal method; A model unit that selects a model to be used for the above machine learning and prediction, performs training using preprocessed acquired data as input, separates the training data into training data and test data, performs model training with the training data, and evaluates the training results with the test data; A scaling unit that sets a target value to be derived through machine learning of the above model and performs scaling, and scales the search range for PID gain optimization to match the range used for model training. A motion stage control gain tuning system characterized by having a gain optimization unit that derives an optimal gain value that minimizes an objective function based on the above optimization algorithm, and converts it to fit the original data range through a de-scaling operation.
In paragraph 2, The above-described scaling unit is characterized by redesigning the model through hyper-parameter optimization and then replacing the model to perform scaling operations when performance is improved compared to the existing model.
In paragraph 2, A motion stage control gain tuning system characterized by driving the motion stage by setting the PID control gain to the above-mentioned optimal gain value and further comprising a verification unit to check whether it is identical to the target value.
In paragraph 2, A motion stage control gain tuning system characterized by the scaler used in the above data section being standardized or normalized.
In paragraph 2, A motion stage control gain tuning system characterized by using MAPE (Mean Absolute Percentage Error) as the evaluation metric in the above model section to evaluate whether the mean absolute percentage error falls within the critical performance range.
(A) A step of collecting response characteristic data according to changes in PID gain values through motion stage driving in the data section; (B) A preprocessing step in the above data section, wherein the learning influence of each feature is equalized through data scaling and outliers in the acquired data are removed using an outlier removal method; (C) A step in which a model is selected for machine learning and prediction in the model section, and training is performed using preprocessed acquired data as input; (D) A step in which the above model unit separates the learning data into training data and test data, performs model training with the training data, and evaluates the learning results with the test data; (E) A step of redesigning the model through hyperparameter optimization in the scaling section and replacing the model only if its performance is improved compared to the existing model; (F) A step of setting a target value to be derived through model learning in the above scaling unit, and adjusting the data range to the y-value resulting from model learning through scaling of the target value; (G) A step of scaling the search range for PID gain optimization in the above scaling unit to match the range used for model training; and (H) A step of optimizing the gain based on an optimization algorithm in the gain optimization unit; A motion stage control gain tuning method including
In Paragraph 7, The scaler used in step (B) above is standardization or normalization, and A motion stage control gain tuning method characterized by having one or more of the following outlier removal methods: IQR (Inter Quantile Range), Z-score, and Modified Z-score.
In Paragraph 7, A motion stage control gain tuning method characterized by separating training data and test data in an 8:2 ratio in step (D) above.
In Paragraph 7, A motion stage control gain tuning method characterized by using MAPE as the evaluation indicator in step (D) above to evaluate whether the average absolute percentage error falls within the threshold performance.
In Paragraph 7, A motion stage control gain tuning method characterized by setting an initial hyperparameter search range in step (E) above, performing a hyperparameter optimization algorithm, training a model with the derived hyperparameters, and then replacing the model when the target model evaluation performance is improved compared to the existing model.
In Paragraph 11, A motion stage control gain tuning method characterized in that the above hyperparameter optimization algorithm is Bayesian optimization, which is a method for finding an optimal solution that maximizes or minimizes the objective function based on Bayesian probability.
In Paragraph 7, A motion stage control gain tuning method characterized in that, in step (H) above, the optimization algorithm is one or more of Bayesian Optimization, Powell minimize, and Basin Hopping.
In Paragraph 7, A motion stage control gain tuning method characterized by descaling the optimal gain range to match the original data range in order to obtain an actually applicable gain value in the above (H) step.
In Paragraph 7, (I) A step of verifying by applying the optimal gain value of the above (H) step to the motion stage in the verification unit; A motion stage control gain tuning method characterized by further performing
A computer-readable recording medium for a program to implement a motion stage control gain tuning method according to any one of claims 7 through 15.

Description

Motion Stage Control Gain Tuning System and Method and Recording Medium thereof The present invention relates to a motion stage control gain tuning system and method and a recording medium thereof, and more specifically, to a system and method capable of efficiently and precisely deriving an optimal gain value by generating a machine learning-based model and comparing it with a target response characteristic based on an optimization algorithm, and a recording medium thereof. Generally, gain tuning of a PID controller is performed through iterative trial-and-error, and time and cost are incurred as the PID gain values must be readjusted due to system changes or experimental environment changes, such as component replacement or weight changes. Since it is difficult to obtain consistent results and subjective judgment may be involved when the user manually adjusts PID control gain tuning, research and patents have been disclosed that enable automatic control gain tuning in any system or environment according to rules. However, there was a problem in that the calculations were very complex and it was not suitable for ultra-precision stage control systems that require periodic tuning according to environmental changes. For example, Patent No. 10-1849464 (Method for Automatic PID Gain Tuning) discloses a method for automatically adjusting gains using parameters to determine the slope of the sliding mode and the control period of the system in a PID controller for a nonlinear system, but there was a problem that it was difficult to use in actual field conditions where various disturbances could occur. In addition, Patent No. 10-1481645 (Method for Automatic PID Gain Tuning of a Multi-Joint Robot) discloses a method for automatically setting the PID gains of a multi-joint robot, wherein initial values for proportional gain, derivative gain, and integral gain are set using the moment of inertia and natural frequency of each arm, and PID control is performed by sequentially increasing the derivative gain, proportional gain, and integral gain until the performance is satisfactory; however, there was a problem in that PID control had to be executed continuously to obtain the gain values. Meanwhile, machine learning is a technology that enables computers to learn, detect, classify, and predict information. Since it eliminates human intervention in complex calculations, it can rapidly produce objective and precise results, leading to ongoing research applying this technology across various industries. Although research on machine learning-based control gain tuning methods is actively underway in PID control across various fields, high performance of machine learning models is required for nano-level precision in ultra-precision stages. Figure 1 is a diagram showing the calculation structure of a typical PID controller. FIG. 2 is a block diagram of a motion stage control gain tuning system according to an embodiment of the present invention. FIG. 3 is a flowchart illustrating a motion stage control gain tuning method according to an embodiment of the present invention. Figure 4 is a detailed flowchart of the hyperparameter optimization steps shown in Figure 3. The configuration and operation of embodiments of the present invention will be described below with reference to the attached drawings. It should be noted that identical components in the drawings are represented by the same reference numbers and symbols whenever possible, even if they are shown on different drawings. In the following description of the present invention, if it is determined that a detailed description of related known functions or configurations may unnecessarily obscure the essence of the present invention, such detailed description will be omitted. Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. First, to aid in understanding the present invention, the PID controller of the ultra-precision motion stage will be explained with reference to FIG. 1. Figure 1 is a diagram showing the calculation structure of a typical PID controller. The ultra-precision motion stage implements high-speed and ultra-precision motion through feedback control using actuators and position sensors. At this time, feedback control mainly utilizes a PID controller, and the PID controller has a structure that measures the output of the target to be controlled as shown in Fig. 1 and calculates the error by comparing it with the desired setpoint. The above PID controller determines control performance through proportional, integral, and differential gain tuning, and is configured as shown in Equation 1 below, with each term being proportional to the error value, the integral of the error value, and the differential of the error value. However, as mentioned above, the control gain tuning of a PID controller is performed thro