A deep learning-based motor servo system controller parameter self-tuning method

Abstract

This invention discloses a deep learning-based method for self-tuning controller parameters in a motor servo system. The method includes: constructing a hybrid model combining a mechanistic model and feedforward neural network residual compensation; performing multi-objective optimization based on this model to obtain a set of initial offline controller parameter values; and constructing a parameter adjustment model based on a Long Short-Term Memory (LSTM) network during online operation. This model takes the system state sequence as input and outputs the percentage change in parameter offset relative to the offline initial values, achieving online, constrained, and small-range dynamic tuning of the controller parameters. This invention improves the reliability of performance prediction through the hybrid model and ensures engineering safety through a constrained online adjustment mechanism, thereby significantly improving the tracking accuracy, stability, and adaptability of the servo system under complex and variable operating conditions.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and motion control technology, specifically to an offline-online integrated self-tuning method for motor servo system controller parameters based on feedforward neural networks and long short-term memory networks (LSTM). Background Technology

[0002] With the development of electronic manufacturing equipment such as CNC machine tools, industrial robots, aerospace, and dispensing equipment, motor servo systems are facing higher demands for response speed, tracking accuracy, and adaptability to different operating conditions. These systems typically include position loops, speed loops, current loops, as well as control components such as filters and feedforwards. The controller parameters of each component have a significant impact on the system's dynamic response performance, steady-state accuracy, and disturbance rejection performance.

[0003] Existing motor servo system controller parameter tuning methods mainly include two categories: offline tuning methods and online tuning methods.

[0004] Offline tuning methods typically rely on mechanistic modeling, frequency domain identification, parameter identification, or engineering experience to solve for parameters. These methods configure parameters before system operation and offer advantages such as relatively mature implementation and high operational safety. However, in real-world servo systems, factors such as tribolinearity, external disturbances, load variations, and parameter drift often exist, leading to discrepancies between the theoretical model and the actual system. This means that controller parameters obtained through offline tuning often only achieve good results within a limited range of operating conditions and are difficult to adapt to continuously changing conditions during actual operation.

[0005] With the development of data-driven approaches, existing research has attempted to establish a mapping relationship between controller parameters and system performance through surrogate models, neural network models, or other models to improve offline tuning efficiency and parameter optimization capabilities. These methods reduce reliance on precise mechanistic models to some extent; however, complete reliance on black-box models can easily lead to problems such as insufficient model interpretability, unstable generalization ability, and insufficient prediction reliability under unknown operating conditions.

[0006] Online tuning methods typically include gain scheduling, adaptive control, and fuzzy control. These methods can adjust controller parameters based on changes in error, state, or operating conditions, thereby improving system adaptability. However, these methods usually rely on manually constructed tuning rules, operating condition partitioning logic, or online identification results, which limits their portability and scalability across different devices and application scenarios. Data-driven methods such as reinforcement learning and deep learning are also used for online parameter tuning, but if trial-and-error learning is performed directly in a real servo system, it may be affected by factors such as noise, data distribution bias, and insufficient safety constraints, leading to abrupt parameter updates or fluctuations in control performance, posing certain risks in engineering applications. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for self-tuning the parameters of a motor servo system controller based on deep learning.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for self-tuning controller parameters of a motor servo system based on deep learning, comprising the following steps:

[0010] S1. Offline parameter initial value determination stage:

[0011] S11. Establish a mechanistic model of the servo system, wherein the mechanistic model is based on a given controller parameter vector. and working condition feature vector Below, output the theoretical performance index vector. ;

[0012] S12. Based on real system experimental data, calculate the prediction error of the mechanism model, i.e., the residual vector. The controller parameter vector is learned by a feedforward neural network. and working condition feature vector to the residual vector Based on the mapping relationship, a feedforward neural network residual compensation model is constructed to output the residual prediction value. ;

[0013] S13, Output of the mechanism model The output of the feedforward neural network residual compensation model Add them together to construct a hybrid model for estimating the performance of the real system;

[0014] S14. Using the multiple performance indicators output by the hybrid model as the objective function, perform multi-objective optimization to obtain a set of controller parameter combinations as initial values ​​for the offline controller parameters. ;

[0015] S2, Online Parameter Dynamic Adjustment Stage:

[0016] S21. During system operation, construct the online tuning state vector. , At least include system input information System feedback status information Tracking error information The current controller parameters relative to the initial values ​​of the offline controller parameters. Offset percentage And the parameter adjustment action at the previous moment. ;

[0017] S22. Construct a parameter adjustment model based on a Long Short-Term Memory (LSTM) network, wherein the parameter adjustment model uses the state vector... As input, output the parameter adjustment action at the current moment. This refers to the percentage change in controller parameter offset, and the action taken to adjust the parameters. Apply single-step change constraints;

[0018] S23. Adjust the action according to the parameters. Update parameter offset percentage And the percentage offset of the updated parameters. Apply boundary constraints;

[0019] S24. Calculate the actual controller parameters for the next time step. ,in Represents the Hadamard product;

[0020] S25. Calculate the actual controller parameters obtained in step S24. Write the data to the servo controller for control calculations in the next control cycle, and return to step S21 to form a closed-loop online tuning.

[0021] Furthermore, in step S11, the performance indicators include one or more of the following: position overshoot, steady-state error, rise time, settling time, and tracking error.

[0022] Further, in step S21, the system input information This includes one or more of the following: reference position, reference velocity, and reference acceleration; the system feedback status information. This includes one or more of the following: actual position, actual velocity, and actual current; the tracking error information. This includes one or more of the following: position error, velocity error, and rate of change.

[0023] Furthermore, step S2 also includes a model training phase, in which an A-head dynamic prediction model is constructed to simulate system dynamics during training. The A-head model uses a state vector... Given the input, predict the system feedback state at the next time step. During the online deployment phase after model training is complete, only the B-head parameter adjustment model is retained and run.

[0024] Further, in step S24, the boundary constraint is: ,in and These are the lower and upper bounds of the allowable range for the parameter offset percentage, respectively; the single-step change constraint is: ,in This represents the maximum permissible percentage change in offset per step.

[0025] Furthermore, in step S14, the multi-objective optimization employs the non-dominated sorting genetic algorithm NSGA-II.

[0026] Furthermore, the training objective of the A-head dynamic prediction model is to minimize the total loss function. To achieve, the ,in, To predict losses in one step, To predict losses in a rolling manner, Weighted loss for critical phases and These are the weighting coefficients.

[0027] Furthermore, the training objective of the B-head parameter adjustment model is to minimize the total loss function. To achieve, the ,in, For the prior constraint loss term, To track the error loss term, This is the overshoot suppression loss term. For action loss items, For smoothness loss term, For terminal loss items, The weighting coefficients for the prior constraint loss term.

[0028] Furthermore, in step S25, the formation of closed-loop online tuning specifically refers to a cyclic process of state perception, parameter adjustment, control execution, and feedback update.

[0029] Furthermore, the method is used to achieve adaptive control of the servo system under conditions of operating condition switching, friction fluctuations, parameter drift, or external disturbances.

[0030] Compared with the prior art, the advantages of the present invention are as follows:

[0031] 1. Improved model accuracy: By constructing a hybrid model that combines a mechanistic model with feedforward neural network residual compensation, the systematic error between the model and the real system was corrected using a data-driven method while retaining the physical interpretability and stability of the mechanistic model. This significantly improved the accuracy of servo system performance prediction and provided a reliable foundation for subsequent optimization and online adjustment.

[0032] 2. It combines the advantages of offline global optimization and online local fine-tuning: It adopts a two-stage strategy of "offline optimization to determine initial values ​​and online dynamic fine-tuning". In the offline stage, multi-objective optimization is performed based on a high-precision hybrid model to obtain a set of globally optimal initial parameter values ​​with excellent overall performance and meeting safety constraints. In the online stage, the LSTM model is used to make bounded, small-amplitude dynamic adjustments around these initial values, so that the parameters can adapt to changes in operating conditions in real time. This overcomes the shortcomings of poor adaptability of purely offline methods and certain deficiencies in the portability and scalability of purely online methods under different operating conditions.

[0033] 3. Significantly improved system tracking accuracy and dynamic performance: During the online tuning phase, the LSTM parameter tuning model can learn and output optimized parameter tuning strategies based on real-time state sequences (including reference, feedback, error, and historical tuning actions). This strategy aims to directly optimize tracking errors, suppress overshoot, and enhance convergence, enabling the system to maintain high-precision trajectory tracking and good dynamic response even when facing complex operating conditions such as friction changes and load disturbances.

[0034] 4. Ensures high safety and robustness in engineering applications: By setting the online adjustment target as a "parameter offset percentage" rather than an absolute parameter value, and limiting the offset and its single-step changes within preset boundaries, it ensures that parameter updates are always continuously and smoothly adjusted within a safe range near the offline initial value. This design effectively prevents system instability or drastic performance fluctuations caused by parameter mutations, significantly reducing the risks of directly applying data-driven methods in real control systems.

[0035] 5. Enhanced interpretability and generalization ability: This method does not rely entirely on a "black box" model. Its core framework integrates prior physical mechanisms (mechanistic models), providing a better interpretability foundation for the entire parameter tuning process. Furthermore, both the offline hybrid model and the online LSTM model can learn from historical operating data, enabling the method to adapt to different equipment characteristics and diverse operating conditions, exhibiting strong generalization ability and transfer potential.

[0036] 6. Improved automation and efficiency in the tuning process: The entire process (from hybrid model construction and offline optimization to online LSTM model training and deployment) is highly automated, reducing reliance on the experience and knowledge of control engineers and extensive manual parameter tuning. Especially in the offline optimization stage, it can systematically seek optimization in the parameter space, which is more efficient than traditional trial-and-error methods, providing an effective way for rapid deployment and performance improvement. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a method for self-tuning the parameters of a motor servo system controller based on deep learning, according to an embodiment of the present invention.

[0039] Figure 2 This is a structural diagram of the hybrid model according to an embodiment of the present invention;

[0040] Figure 3 This is a structural diagram of the online parameter tuning servo system according to an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, this invention relates to a deep learning-based method for self-tuning parameters of a motor servo system controller, applicable to servo control systems affected by operating condition switching, friction fluctuations, parameter drift, and external disturbances. This method constructs a hybrid model combining a mechanistic model and a feedforward neural network residual compensation model. Figure 2 The system then performs multi-objective optimization based on the hybrid model to obtain initial values ​​for the offline controller parameters. Next, a parameter tuning model based on a long short-term memory network is constructed. This model outputs the percentage change in parameter offset relative to the initial offline controller parameters based on the system's operating state sequence, thereby achieving online, small-range dynamic tuning of the controller parameters. Specifically:

[0043] Step 1: Establish a hybrid model and perform multi-objective optimization to obtain initial values ​​for the controller parameters.

[0044] 1) Establish a system mechanism model

[0045] A mechanistic model of the motor servo system is established to output the theoretical performance indicators of the system under given controller parameter combinations and operating conditions. Let the controller parameter vector be... The working condition feature vector is The theoretical performance index vector output by the mechanistic model is denoted as:

[0046]

[0047] in, This represents a vector of theoretical performance indicators output by the mechanistic model. These performance indicators can be selected based on actual application requirements, including one or more of the following: position overshoot, steady-state error, rise time, settling time, and tracking error.

[0048] 2) Construct residuals and establish a feedforward neural network compensation model

[0049] Due to the presence of tribolinearity, external disturbances, parameter perturbations, and unmodeled dynamics in real servo systems, there is often a discrepancy between the output of the mechanistic model and the experimental results of the real system. Therefore, a feedforward neural network is constructed to learn the error components that the mechanistic model fails to accurately describe and to compensate for the output of the mechanistic model. Let the performance index vector measured in the real system experiment under the same controller parameter combination and operating conditions be:

[0050]

[0051] The residual vector between the mechanistic model and the real system is then defined as:

[0052]

[0053] in, This represents the prediction error of the mechanistic model on the performance of the real system.

[0054] Using the controller parameter vector and operating condition feature vector as input, and the residual vector as supervision label, a feedforward neural network residual compensation model is trained. The predicted residual value output by the feedforward neural network is then:

[0055]

[0056] in, This represents the prediction of the residual by the feedforward neural network.

[0057] 3) Construct a hybrid model

[0058] Adding the output of the mechanistic model to the prediction residual of the feedforward neural network yields the estimation result of the hybrid model for the performance index of the real system, i.e., the output of the hybrid model is:

[0059]

[0060] Right now:

[0061]

[0062] in, This represents the estimated performance index output by the hybrid model.

[0063] In this way, the mechanistic model is used to characterize the dominant dynamic laws of the system, and the feedforward neural network is used to compensate for the systematic deviation between the mechanistic model and the real system, thus forming a hybrid model that combines basic physical constraints and error correction capabilities.

[0064] 4) Multi-objective optimization based on a hybrid model

[0065] After obtaining the hybrid model, a multi-objective optimization problem is established using the multiple performance indices output by the hybrid model as the objective function. The feasible region of the controller parameters is then searched to obtain a Pareto candidate solution set. The multi-objective optimization algorithm can be NSGA-II or other multi-objective optimization algorithms. Further, considering practical engineering constraints, system stability requirements, parameter boundary constraints, and comprehensive performance evaluation criteria, a set of controller parameter combinations that meet operational safety and comprehensive performance requirements is selected from the Pareto candidate solution set as the initial values ​​for the offline controller parameters. .

[0066] Step 2: Establish LSTM learning and parameter tuning rules and perform online tuning.

[0067] 1) Constructing the online tuning state vector

[0068] Let the first The system state vector at each time step is:

[0069]

[0070] in, This indicates system input information. This indicates that the system is providing feedback status information. Indicates tracking error information. This represents the percentage offset of the current controller parameters from their initial offline controller parameters. This indicates the parameter adjustment action at the previous moment, i.e., the percentage change in the controller parameter offset at the previous moment. The system input information may include one or more of the following: reference position, reference velocity, and reference acceleration; the feedback status information may include one or more of the following: actual position, actual velocity, and actual current; the tracking error information may include one or more of the following: position error, velocity error, and their rate of change.

[0071] This invention uses a parameter offset percentage instead of directly outputting the absolute parameter value of the controller during the online phase. This restricts the online tuning process to the vicinity of the initial value of the offline controller parameters, which helps to reduce the risk of parameter mutation and improve the safety of system operation.

[0072] 2) Establish an A-head dynamic prediction model

[0073] To describe the impact of current state and parameter shift changes on the system response at the next time step during the training phase, a dynamic prediction model is constructed. This model predicts the system feedback state at the next time step based on the current state vector and parameter change information, and its output is expressed as:

[0074]

[0075] in, This represents the predicted system feedback state value at the next moment.

[0076] The A-head model does not directly participate in online parameter updates during the actual deployment phase. Instead, it serves as a learnable dynamic environment for the B-head model during the training phase, avoiding trial-and-error online parameter tuning on real server systems.

[0077] 3) Establish a B-head parameter adjustment model

[0078] Construct a B-head parameter adjustment model, wherein the B-head model uses a state vector As input, the output shows the adjustment action of each controller parameter at the current moment, i.e., the percentage change in parameter offset:

[0079]

[0080] in, This indicates the parameter adjustment action at the current moment.

[0081] Update the current parameter offset percentage based on the B-head output action:

[0082]

[0083] Further, the actual controller parameters at the next time step are obtained:

[0084]

[0085] in, This represents the Hadamard product.

[0086] To ensure engineering safety during online tuning, parameter updates must satisfy parameter boundary constraints and single-step change constraints:

[0087]

[0088]

[0089] in, and These represent the lower and upper bounds of the allowable range for the parameter offset percentage, respectively. This represents the maximum permissible percentage change in offset per step. Through the above constraints, online parameter adjustments are prevented from being drastic and abrupt, but rather are made through bounded, continuous, small corrections near the initial values ​​of the offline controller parameters.

[0090] 4) A-head loss function design

[0091] The training objective of the A-head model is to obtain a surrogate model capable of describing the short-term dynamic evolution of the system. The total loss function of the A-head model can be designed as follows:

[0092]

[0093] in, This is a one-step prediction loss used to constrain the prediction error of the system feedback state at the next time step; This is the rolling prediction loss, used to constrain the cumulative error during continuous multi-step recursion. Weighted loss for critical stages is used to highlight the fitting accuracy of critical intervals such as sudden changes in operating conditions, start-up and shutdown phases, or sudden changes in acceleration. and These are the corresponding weighting coefficients.

[0094] Let the number of system sampling points be... The one-step prediction loss can be expressed as:

[0095]

[0096] Let the rolling prediction step size be... Then the rolling prediction loss can be expressed as:

[0097]

[0098] The weighted loss for the critical stage can be expressed as:

[0099]

[0100] in, The weights for critical stages are set to larger values ​​in the operating condition switching interval, start-stop interval, or other critical intervals that have a significant impact on control performance.

[0101] 5) Design of B-head loss function

[0102] The training objective of the B-head model is to learn a parameter tuning strategy that improves trajectory tracking performance, suppresses overshoot, maintains motion smoothness, and enhances convergence during the stopping phase. The total loss function of the B-head model is designed as follows, including prior behavioral constraints. Tracking error term Overshoot penalty item in critical stage Action amplitude penalty item Smooth motion penalty item The penalty term for convergence during the stopping phase .

[0103]

[0104] 1. Prior behavioral constraints

[0105]

[0106] in, This indicates the parameter adjustment action in the sample; This is used to constrain the consistency between the B-head output parameter adjustment actions and the parameter adjustment actions in the samples, so that the network can inherit the existing parameter tuning experience during the training process and improve training stability.

[0107] 2. Tracking error term

[0108]

[0109] in, Weighting is applied to larger trajectory tracking error ranges. Higher weights are assigned to larger trajectory error ranges to improve tracking performance during these critical phases.

[0110] 3. Overshoot penalty

[0111]

[0112] in, This is the weight for the overshoot-sensitive region. This term is used to apply additional constraints near the end of acceleration / deceleration or other overshoot-sensitive phases, guiding the network to output more robust parameter adjustments.

[0113] 4. Action amplitude penalty item

[0114]

[0115] in, This is the motion amplitude penalty coefficient. This item is used to suppress excessively large single-step parameter offset percentage changes, reducing aggressive behavior during online adjustment.

[0116] 5. Smoothness of movement penalty item

[0117]

[0118] in, This is the motion smoothing penalty coefficient. This parameter effectively reduces high-frequency jitter during parameter adjustment, making the output motion more continuous.

[0119] 6. Convergence penalty term during the stopping phase

[0120]

[0121] in, The system sampling length during the stopping phase. This is the tracking error penalty coefficient for the stopping phase. This term is used to enhance the final value convergence performance during the stopping phase, suppressing steady-state errors and slow convergence issues.

[0122] 6) Online deployment

[0123] After the model training is completed, only the B-head parameter adjustment model is retained in the actual deployment phase to run as an online controller parameter self-tuner.

[0124] like Figure 3 As shown, the system collects reference trajectory, feedback status, error information, current parameter offset percentage, and parameter adjustment actions from the previous moment in real time during each control cycle, forming an input sequence and feeding it into the B-head model. The B-head model outputs the parameter correction amount at the current moment, i.e., the change in parameter offset percentage. Subsequently, the parameter correction amount undergoes single-step change limiting processing, and the updated parameter offset percentage is subject to boundary constraint processing. It is then mapped to the current actual controller parameters and written to the servo controller to participate in the control calculation of the next control cycle. The system continues to collect new real feedback status and update the input sequence, thus forming a closed-loop online tuning process of "status perception—parameter adjustment—control execution—feedback update".

[0125] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0126] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for self-tuning controller parameters of a motor servo system based on deep learning, characterized in that, Includes the following steps: S1. Offline parameter initial value determination stage: S11. Establish a mechanistic model of the servo system, wherein the mechanistic model is based on a given controller parameter vector. and working condition feature vector Below, output the theoretical performance index vector. ; S12. Based on real system experimental data, calculate the prediction error of the mechanism model, i.e., the residual vector. The controller parameter vector is learned by a feedforward neural network. and working condition feature vector to the residual vector Based on the mapping relationship, a feedforward neural network residual compensation model is constructed to output the residual prediction value. ; S13, Output of the mechanism model The output of the feedforward neural network residual compensation model Add them together to construct a hybrid model for estimating the performance of the real system; S14. Using the multiple performance indicators output by the hybrid model as the objective function, perform multi-objective optimization to obtain a set of controller parameter combinations as initial values ​​for the offline controller parameters. ; S2, Online Parameter Dynamic Adjustment Stage: S21. During system operation, construct the online tuning state vector. , At least include system input information System feedback status information Tracking error information The current controller parameters relative to the initial values ​​of the offline controller parameters. Offset percentage And the parameter adjustment action at the previous moment. ; S22. Construct a parameter adjustment model based on a Long Short-Term Memory (LSTM) network, wherein the parameter adjustment model uses the state vector... As input, output the parameter adjustment action at the current moment. This refers to the percentage change in controller parameter offset, and the action taken to adjust the parameters. Apply single-step change constraints; S23. Adjust the action according to the parameters. Update parameter offset percentage And the percentage offset of the updated parameters. Apply boundary constraints; S24. Calculate the actual controller parameters for the next time step. ,in Represents the Hadamard product; S25. Calculate the actual controller parameters obtained in step S24. Write the data to the servo controller for control calculations in the next control cycle, and return to step S21 to form a closed-loop online tuning.

2. The self-tuning method according to claim 1, characterized in that, In step S11, the performance indicators include one or more of the following: position overshoot, steady-state error, rise time, settling time, and tracking error.

3. The self-tuning method according to claim 1, characterized in that, In step S21, the system inputs information. This includes one or more of the following: reference position, reference velocity, and reference acceleration; the system feedback status information. This includes one or more of the following: actual position, actual velocity, and actual current; the tracking error information. This includes one or more of the following: position error, velocity error, and rate of change.

4. The self-tuning method according to claim 1, characterized in that, Step S2 also includes a model training phase, in which an A-head dynamic prediction model is constructed to simulate system dynamics during training. The A-head model uses a state vector... Given the input, predict the system feedback state at the next time step. During the online deployment phase after model training is complete, only the B-head parameter adjustment model is retained and run.

5. The self-tuning method according to claim 1 or 4, characterized in that, In step S24, the boundary constraints are: ,in and These are the lower and upper bounds of the allowable range for the parameter offset percentage, respectively; the single-step change constraint is: ,in This represents the maximum permissible percentage change in offset per step.

6. The self-tuning method according to claim 1, characterized in that, In step S14, the multi-objective optimization adopts the non-dominated sorting genetic algorithm NSGA-II.

7. The self-tuning method according to claim 4, characterized in that, During the model training phase, the training objective of the A-head dynamic prediction model is to minimize the total loss function. To achieve, the ,in, To predict losses in one step, To predict losses in a rolling manner, Weighted loss for critical phases and These are the weighting coefficients.

8. The self-tuning method according to claim 4, characterized in that, During the model training phase, the B-head parameters adjust the model's training objective by minimizing the total loss function. To achieve, the ,in, For the prior constraint loss term, To track the error loss term, This is the overshoot suppression loss term. For action loss items, For smoothness loss term, For terminal loss items, The weighting coefficients for the prior constraint loss term.

9. The self-tuning method according to claim 1, characterized in that, In step S25, the formation of closed-loop online tuning specifically refers to a cyclic process of state perception, parameter adjustment, control execution, and feedback update.

10. The self-tuning method according to claim 1, characterized in that, The method is used to achieve adaptive control of the servo system under conditions of operating condition switching, friction fluctuation, parameter drift or external disturbance.

Images