An adaptive control method for a gun servo system

Abstract

The application relates to the technical field of adaptive control and reinforcement learning, and discloses an adaptive control method for a gun servo system, which comprises the following steps: determining baseline gain parameters according to a low-order nominal model and a reference model; constructing a T-S fuzzy controller, input variables of an antecedent part are divided into multiple rule regions, and activation weights of the rule regions are determined; each rule region of a consequent part is a parameterized structure consistent with an MRAC controller, and the baseline gain parameters are mapped into initial values of parameters of each rule region of the consequent part; during a closed-loop operation process, control error information is input into the antecedent part, expected state variables, actual state variables and the control error information are input into the consequent part, a total control variable is obtained, and each parameter obtained during the closed-loop operation process is used for optimizing and updating parameters of each rule region of the consequent part; and the application can realize adaptive compensation while keeping the model reference structure interpretable.

Description

Technical Field

[0001] This invention relates to the fields of adaptive control and reinforcement learning, and in particular to an adaptive control method for artillery servo systems. Background Technology

[0002] The artillery servo system is a core component of artillery weapon systems, primarily used to drive the turret or barrel to achieve precise tracking and stable control in azimuth and elevation. Its control performance directly affects the artillery's aiming accuracy, firing reaction time, and stability when firing on the move. In actual combat environments, the artillery servo system needs to maintain high-precision, high-dynamic-response tracking capabilities under complex conditions (such as road bumps, temperature changes, and ammunition launch impacts), while simultaneously withstanding interference from multiple nonlinear factors such as dead zones, saturation, friction, measurement noise, and load parameter variations. Therefore, researching an adaptive control method that combines high precision, strong robustness, and engineering deployability is of great significance for improving the combat effectiveness of artillery weapon systems.

[0003] Currently, adaptive control of artillery servo systems mainly employs Model Reference Adaptive Control (MRAC) and its improved methods. MRAC constructs a desired reference model and designs adaptive laws to adjust controller parameters in real time, enabling the output of the controlled object to track the output of the reference model, thereby coping with parameter changes and external disturbances to a certain extent. To further enhance the adaptability to nonlinear factors, some studies have introduced fuzzy logic into MRAC, forming Fuzzy Model Reference Adaptive Control (Fuzzy-MRAC), which uses fuzzy rules to approximate compensation for nonlinear elements. Furthermore, with the development of artificial intelligence technology in recent years, reinforcement learning (RL) methods, especially Deep Deterministic Policy Gradient (DDPG) and its improved algorithm Twin Delayed Deep Deterministic Policy Gradient (TD3), have begun to be explored for application in the self-tuning of control parameters and policy learning of servo systems due to their independence from precise models and autonomous optimization capabilities. In short, adaptive control research in artillery servo systems mainly follows two technical routes: model-based control methods and learning-based control methods. However, both routes have insurmountable drawbacks, and attempts to integrate the related technologies of both have failed to fundamentally solve the problems.

[0004] First, model-based methods, such as Model Reference Adaptive Control (MRAC), have a core limitation: they are highly dependent on the nominal model and matching conditions. When the artillery servo system suffers from severe nonlinear factors such as strong dead zones, saturation, and friction, or when system parameters vary over a wide range, the adaptive law is prone to problems such as parameter drift, decreased convergence speed, and even system instability. A deeper problem is that the control law structure of MRAC is relatively fixed, and its adaptive adjustment capability is limited to the pre-set model structure framework, making it difficult to fully compensate for complex nonlinearities. This has forced researchers to seek more flexible control methods, thus prompting the exploration of reinforcement learning techniques in the field of servo control.

[0005] However, when researchers turn to reinforcement learning-based methods, they face new dilemmas. While pure reinforcement learning methods do not rely on precise models, their policies are typically represented by deep neural networks, exhibiting black-box characteristics. The physical meaning of the control law is unclear, making stability analysis and engineering verification difficult. More importantly, when conducting online exploration on high-value, high-safety-requirement physical systems such as artillery, drastic fluctuations in parameters during random initialization of strategies or training can lead to dangerous control outputs, resulting in high risks for physical training and high deployment barriers. This creates a dilemma: model-based methods have fixed structures but are stable and interpretable, making them difficult to handle complex nonlinearities; learning-based methods are flexible and adaptive but are black-box and uncontrollable, making safe deployment difficult.

[0006] Existing technical solutions attempting to integrate MRAC and reinforcement learning have also failed to effectively resolve the aforementioned contradictions. These solutions often design the reinforcement learning unit as an additional black-box compensator, superimposed on the MRAC control law, resulting in a lack of transparency in the overall control structure. Therefore, simply superimposing a black-box compensator not only fails to solve the fundamental contradiction of MRAC's inability to handle strong nonlinearity, but also introduces new problems such as unprovable stability, unpredictable behavior, and undiagnosable faults. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an adaptive control method for artillery servo systems, which can achieve adaptive compensation for factors such as dead zone, saturation, friction, noise, and parameter mismatch while maintaining the interpretability of the model reference structure.

[0008] This invention provides an adaptive control method for artillery servo systems, comprising the following steps: Construct a low-order nominal model and a reference model based on the artillery servo system, and determine the baseline gain parameters based on the low-order nominal model and the reference model; A Takagi-Sugeno fuzzy controller is constructed, which consists of an antecedent and a consequent. The antecedent uses the system's control error information as the input variable, divides the input variable into multiple regular regions, constructs the membership function of each regular region, obtains the activation degree of each regular region, and thus determines the activation weight of each regular region. The regular regions of the consequent are parameterized structures consistent with the MRAC controller, and the baseline gain parameter is mapped to the initial value of the parameters of each regular region of the consequent. During closed-loop operation, the expected and actual state variables of the artillery servo system are acquired, and the control error information of the system is calculated. The control error information is input into the antecedent part to determine the activation weight of each rule region. The expected state variable, actual state variable, and control error information are input into the consequent part to obtain the local control output of each rule region of the consequent part. The local control output is weighted and synthesized according to the activation weight to obtain the total control quantity, which is output to the actuator of the artillery servo system to drive the artillery servo system to run. The expected state variables, actual state variables, and control error information of the system obtained during closed-loop operation are used to optimize and update the parameters of each rule region in the subsequent part.

[0009] Optionally, the input variables are divided into multiple rule regions, and a membership function for each rule region is constructed, including the following steps: The control error information is normalized to obtain the normalized error. The normalization process uses a preset error calibration value to map the control error information to a preset range. The range of normalization error is divided into multiple regular regions, including the negative error region, the error region near zero, and the positive error region. For each regular region, a membership function is constructed. The membership function includes Gaussian function, triangular function, trapezoidal function or bell function.

[0010] Optionally, the activation weight of each rule region is determined in the following way: The membership function value corresponding to each rule region is used as the original activation degree of that rule region. The original activation values ​​of each rule region are normalized to obtain the normalized activation value, which is the ratio of the original activation value of each rule region to the sum of the original activation values ​​of all rule regions. The activation weight of each rule region is determined based on the normalized activation degree, where the sum of the normalized activation degrees of each rule region is 1.

[0011] Optionally, the local control outputs of each rule region in the consequent are determined in the following way: Construct local control laws for each rule region. The local control laws are expressed as the product of the parameter vectors of each rule region and the regression vector in the consequent part. In this context, each element in the parameter vector of each rule region in the consequent part corresponds to the weighting coefficients of the desired state quantity, the actual state quantity, and the control error information, respectively. The initial values ​​of the parameter vectors of each rule region in the consequent part are obtained from the baseline gain parameter. The regression vector includes the desired state quantity, the actual state quantity, and the control error information.

[0012] Optionally, the baseline gain parameter is mapped to the initial values ​​of the parameters of each regular region in the consequent part, including the following steps: The baseline gain parameter is directly assigned to the parameter vector of each regular region in the consequent part of each regular region, so that each regular region has the same baseline control law in the initial state. Alternatively, based on the error partitioning characteristics corresponding to each rule region, the baseline gain parameter can be scaled, biased, or otherwise equivalently mapped to obtain differentiated initial values ​​for different rule regions.

[0013] Optionally, the total control quantity is obtained by weighting and synthesizing the local control outputs according to the activation weights, including the following steps: Multiply the local control output of each rule region by its corresponding normalized activation degree to obtain the weighted control component of each rule region; The total control output of the Takagi-Sugeno fuzzy controller is obtained by summing the weighted control components of each rule region.

[0014] Optionally, the Twin Delayed Deep Deterministic Policy Gradient algorithm can be used to optimize and update the parameters of each rule region in the consequent part.

[0015] Optionally, the parameters of each rule region in the consequent part are optimized and updated, including the following steps: Two independent evaluation networks are constructed to estimate the value of state-action pairs, and a target value is constructed based on the target action generated by the target policy network. The target value is taken as the smaller value between the outputs of the two evaluation networks to reduce the overestimation of value. By minimizing the evaluation network loss function and updating the evaluation network parameters, the evaluation network learns the long-term control effect under different states and control actions. The parameters of each rule region are optimized based on the output of the evaluation network for the consequent part of each rule region.

[0016] Optionally, during the optimization and update process of the parameters of each rule region in the consequent part, one of the following or engineering constraints may also be included: The parameters of each rule region in the consequent part of the selected key rule region remain fixed and are not updated, while the parameters of each rule region in the consequent part of the other rule regions participate in learning and optimization. The key rule regions include rule regions where the absolute value of the error is less than the preset error threshold, boundary rule regions, or rule regions that have achieved the expected control effect. A downsampling strategy is adopted to update the parameters once every D control cycles, and the parameter iteration frequency is set to 1 / D of the control frequency; Apply upper and lower bound constraints or norm constraints to the parameters of each rule region in the consequent part of each rule region.

[0017] Optionally, the actual state quantities include one of current, angular velocity, or angular position, and the control error information includes one or more of the following: system tracking error, the rate of change of system tracking error, and the integral of system tracking error.

[0018] The technical solution provided by the embodiments of the present invention has the following advantages compared with the prior art: The present invention provides an adaptive control method for artillery servo systems, which adopts a two-stage plus structured parameterization approach: the first stage determines the baseline gain parameters based on a low-order nominal model and a reference model; the second stage designs each regular region of the follow-up part as a parameterized structure consistent with the MRAC controller, and learns the regular-level local gain instead of the output of an additional black-box compensator; therefore, it is more transparent and easier to verify and engineer. The baseline gain parameters are mapped to the initial values ​​of the regular region parameters of the consequent part of the T-S fuzzy controller, so that the initial strategy is in a stable and usable region rather than being randomly initialized. The antecedent part uses error partitioning to form the regular activation degree. The local control outputs of each regular region of the consequent part are weighted and synthesized to obtain the total control quantity. The parameters in each local control output correspond to the expected state quantity, the actual state quantity, and the control error information, thus giving it a clear engineering physical meaning, which is convenient for subsequent analysis, constraint, and update. During the later operation of the consequent part, only the regular region parameters of the consequent part are updated (some rules / parameters are frozen or regularized when necessary). This allows for adaptive compensation for factors such as dead zone, saturation, friction, noise, and parameter mismatch while maintaining the interpretability of the model reference structure, without the need for precise modeling. Attached Figure Description

[0019] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. 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.

[0020] Figure 1A flowchart of an adaptive control method for an artillery servo system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the membership function of the antecedent part of the Takagi-Sugeno fuzzy controller provided in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0022] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] First, embodiments of the present invention provide an adaptive control method for artillery servo systems, specifically as follows: Figure 1 As shown, it includes the following steps: A low-order nominal model based on the artillery servo system and a reference model designed according to the desired closed-loop performance are constructed. Baseline gain parameters are determined based on the low-order nominal model and the reference model. The low-order nominal model is a simplified mathematical description of the main dynamic characteristics of the artillery servo system, used to retain the main control features of the system without excessively increasing modeling complexity. The reference model describes the ideal closed-loop response that the control system aims to achieve. The baseline gain parameters are a set of initial controller parameters used to initialize the parameters of each regular region in the subsequent Takagi-Sugeno fuzzy controller. The method of obtaining the baseline gain parameters is not limited to a specific controller design; in one embodiment, it can be obtained through proportional-integral controller tuning, while in another embodiment, it can be determined by combining model reference control concepts. Their common purpose is to provide a parameter starting point with clear physical meaning and usable initial closed loop for the subsequent Takagi-Sugeno fuzzy controller.

[0024] The purpose of this step is not to allow the baseline control law to operate independently as the final online control law in the long term, but rather to provide a set of parameters with clear physical meaning, usable initial closed-loop parameters, and conducive to subsequent optimization for the fuzzy reinforcement learning controller. This avoids starting the subsequent learning process with random parameters, improving the stability and safety of the controller's initial operation.

[0025] A Takagi-Sugeno fuzzy controller is constructed, which consists of an antecedent and a consequent. The antecedent uses the system's control error information as the input variable, divides the input variable into multiple regular regions, constructs the membership function of each regular region, obtains the activation degree of each regular region, and thus determines the activation weight of each regular region. The regular regions of the consequent are parameterized structures consistent with the MRAC controller, and the baseline gain parameter is mapped to the initial value of the parameters of each regular region of the consequent. During closed-loop operation, the expected and actual state variables of the artillery servo system are acquired, and the control error information of the system is calculated. The control error information is input into the antecedent part to determine the activation weight of each rule region. The expected state variable, actual state variable, and control error information are input into the consequent part to obtain the local control output of each rule region of the consequent part. The local control output is weighted and synthesized according to the activation weight to obtain the total control quantity, which is output to the actuator of the artillery servo system to drive the artillery servo system to run. The expected state variables, actual state variables, and control error information of the system obtained during closed-loop operation are used to optimize and update the parameters of each rule region in the subsequent part.

[0026] In this embodiment, the construction of the low-order nominal model includes the following steps: For artillery servo systems, a low-order nominal model can be established first for the speed control object.

[0027] In a typical embodiment, for the pitch velocity loop or azimuth velocity loop of the artillery servo system, a first-order inertial element can be used as a low-order nominal model, and its expression can be written as: in, Let be the derivative of the state variable with respect to time. This is a state variable of the low-order nominal model, used to characterize the current internal dynamic state of the low-order nominal model; when the low-order nominal model is a first-order model, this state variable can directly correspond to the output of the low-order nominal model. For control input, it represents the drive command applied to the actuator of the artillery servo system, which can correspond to motor current command, torque command, voltage command or its equivalent drive quantity in different implementation scenarios. The output of the controlled object represents the actual output of the artillery servo system. In this embodiment, In a velocity loop control scenario, this typically corresponds to the actual angular velocity of the turret or barrel. In a current loop control scenario, it typically corresponds to the current of the gun motor or the armature current of the drive motor. In a position loop control scenario, it typically corresponds to the actual angular position of the turret azimuth axis or the barrel pitch axis. and These are the system parameters of the low-order nominal model, where... Used to characterize the dynamic properties of the system itself. These parameters are used to characterize the strength of the effect of the control input on the system output. They can be obtained through experimental identification, mechanistic modeling, fitting historical empirical data, or engineering tuning.

[0028] It should be noted that the low-order nominal model does not require high-precision, full-condition accurate modeling of the artillery servo system, but only needs to reflect the basic dynamic response characteristics of the system within the main operating range.

[0029] This is because artillery servo systems typically experience complex nonlinear factors during actual operation, such as dead zones, saturation, friction, backlash, external impact disturbances, sensor noise, and load variations. Establishing a completely accurate mathematical model in the initial stage is not only difficult to implement but also hinders the application of subsequent data-driven optimization methods. Therefore, this invention first constructs a low-order nominal model to provide an interpretable and computable initial control foundation.

[0030] Building a reference model involves the following steps: After obtaining the low-order nominal model, a reference model is further constructed based on the desired control performance.

[0031] The reference model is used to give the ideal closed-loop response process that the artillery servo system should achieve, that is, to specify how the system should respond, rather than to describe how the system actually responds at present.

[0032] The reference model can be represented as: in, For the derivative with respect to time, This is the state vector of the reference model, used to characterize the dynamic evolution process within the reference model. This state vector can be a one-dimensional scalar or a multi-dimensional vector; when the reference model is designed as a second-order dynamic model, It typically contains two state components, used to describe the internal states during the process of the desired output change. It should be noted that... They do not necessarily correspond directly to physical quantities that can be directly measured in artillery systems, but are used as intermediate variables to mathematically characterize the desired dynamic process. In this embodiment, the externally given command represents the target input to the artillery servo system. In other scenarios... It can also be indicated that in the current loop scenario, the external given command can be the desired current, and the reference model output is the desired current response; in the velocity loop scenario, the external given command can be the desired angular velocity, and the reference model output is the desired angular velocity response; it can also be indicated that in the position loop scenario, the external given command can be the desired angular position, and the reference model output is the desired angular position response. This is the state matrix of the reference model, used to describe the changes in the internal state of the reference model over time. The input matrix of the reference model is used to describe the given external instructions. The influence relationship on the state changes of the reference model; This is the output matrix of the reference model, used to represent the internal states of the reference model. Mapped to reference model output . The output is the reference model output, used to represent the desired closed-loop output, i.e., the actual system output. The ideal response target to be tracked.

[0033] In other words, in this embodiment of the invention, the control objective of the artillery servo system is not simply to make the actual output... Track input commands Instead, it makes the actual output Track the output of the reference model as much as possible The reason for adopting this approach is that the reference model can incorporate the expected response speed, overshoot characteristics, damping level, and steady-state accuracy of the system into its dynamic structure in advance, thereby giving the subsequent controller design a clearer target orientation.

[0034] Designing reference model parameters based on performance indicators includes the following steps: Furthermore, reference model parameters can be designed based on the low-order nominal model and the expected closed-loop performance indicators. In a typical embodiment, a proportional-integral controller can be designed first based on the low-order nominal model to obtain the desired closed-loop response that meets engineering requirements. Its control law can be expressed as:

[0035] in, The control input used when constructing the desired closed-loop response based on a low-order nominal model; The proportional gain is used to quickly adjust the control action based on the current error. The integral gain is used to continuously adjust the control action based on the accumulated historical error, in order to reduce the steady-state error. The system tracking error represents the expected instruction. Compared with actual output The difference between them.

[0036] In this process, by adjusting and The value of can enable the low-order nominal model closed-loop system to obtain the desired rise rate, settling time, damping characteristics, and steady-state error level, and further determine the reference model parameters accordingly. , and In one embodiment, the closed-loop transfer relationship formed by the low-order nominal model and the proportional-integral controller can be rearranged to obtain the corresponding equivalent closed-loop dynamic expression, and the state matrix, input matrix, and output matrix of the reference model can be determined based on the equivalent closed-loop dynamic expression.

[0037] In the case of a first-order low-order nominal model combined with a proportional-integral controller, the closed-loop system can usually exhibit equivalent second-order dynamic characteristics. Therefore, the reference model can be designed as an equivalent second-order model. The significance of designing the reference model as an equivalent second-order dynamic model is that the second-order model is convenient for directly describing the desired dynamic performance using commonly used engineering indicators such as natural frequency, damping ratio, overshoot, and settling time. Therefore, it is more suitable for expressing the engineering control objectives of artillery servo systems.

[0038] It should be noted that the reference model is not required to be completely identical to the actual physical structure of the artillery servo system. Its essential function is to define a reasonable and achievable desired response trajectory. Therefore, even if the artillery servo system actually has higher-order dynamics, coupling terms, or nonlinear factors, a combination of the low-order nominal model and the reference model can still provide a clear, stable, and engineering-meaning design benchmark for subsequent controller initialization.

[0039] To obtain the baseline gain parameters, the following steps are included: After obtaining the low-order nominal model and reference model, a baseline control law for initialization can be constructed. To ensure that the parameters of the subsequent fuzzy controller have clear physical meaning and facilitate engineering analysis, this invention selects a regression vector with control interpretability to construct the baseline controller.

[0040] The regression vector can be represented as: in, The desired angular velocity command represents the input information for controlling the target. This represents the actual output of the artillery servo system, indicating the current response status of the system. The system tracking error can be defined as follows: In the current loop, it is the difference between the desired current and the actual current; in the velocity loop, it is the difference between the desired angular velocity and the actual angular velocity; and in the position loop, it is the difference between the desired angular position and the actual angular position.

[0041] This is the error integral term, used to characterize the cumulative result of the error over a period of time. Its function is to reduce steady-state error and improve continuous tracking accuracy. Therefore, in the regression vector... The three components reflect the target input information, output feedback information, and error history accumulation information, respectively, and can jointly characterize the main control information required to form the current control decision.

[0042] Based on the above regression vector, the baseline control law can be written as: in, Output to the baseline controller; The baseline gain parameter vector, whose components correspond to the weighting coefficients of each component of the regression vector. In other words, The parameters in the table are used to characterize: the system's response strength to the target command, the degree of correction to the current output feedback, and the ability to compensate for the error integral term.

[0043] Baseline gain parameter vector The baseline gain parameter can be obtained through MRAC control design, controller tuning results under the nominal model, or parameter conversion combined with the reference model objective. In one embodiment, when the baseline control law adopts the same form as the proportional-integral control structure, the baseline gain parameter can be composed of proportional gain, output feedback gain, and integral gain; furthermore, the initial values ​​of the parameters of each regular region in the consequent part can be obtained by directly assigning values ​​based on the baseline gain parameter or by equivalent conversion.

[0044] In this invention, this set of parameters is not intended to be used continuously as the final online control law over a long period, but rather serves as the source of initial values ​​for the parameters of each rule region in the subsequent Takagi-Sugeno fuzzy controller. In other words, this invention does not simply adopt the traditional MRAC as the final solution, but rather leverages its interpretable structure and available initial control performance to provide a stable starting point for subsequent data-driven optimization.

[0045] After the above steps, the final output of this step is a set of baseline gain parameters. .

[0046] The baseline gain parameter has the following functions: 1. As the initialization source of the parameters of each rule region in the subsequent part of the Takagi-Sugeno fuzzy controller, it enables each rule to have usable control capabilities in the initial state.

[0047] 2. This allows the subsequent learning process to begin near stable parameters with clear physical meaning, rather than starting with random parameters, thereby reducing the risk of closed-loop divergence caused by the initial exploration.

[0048] 3. It provides an engineering interpretation basis for subsequent rule-level parameter optimization, so that the learned parameter changes can correspond to the actual control behavior of the artillery servo system, which facilitates subsequent analysis, constraints and deployment.

[0049] In summary, by constructing a low-order nominal model, designing a reference model, and extracting baseline gain parameters, a foundation was laid for the structured initialization of the Takagi-Sugeno fuzzy controller. This step retains the advantages of model-based methods in terms of initial usability and interpretability, while also reserving adjustment space for subsequent optimization learning based on closed-loop data. This lays the foundation for an adaptive control method in artillery servo systems that balances safety, adaptability, and engineering feasibility.

[0050] The present invention provides an adaptive control method for artillery servo systems, which adopts a two-stage plus structured parameterization approach: the first stage determines the baseline gain parameters based on a low-order nominal model and a reference model; the second stage designs each regular region of the follow-up part as a parameterized structure consistent with the MRAC controller, and learns the regular-level local gain instead of the output of an additional black-box compensator; therefore, it is more transparent and easier to verify and engineer. The baseline gain parameters are mapped to the initial values ​​of the post-rule parameters of each rule region in the consequent part of the T-S fuzzy controller, ensuring that the initial strategy is in a stable and usable region rather than being randomly initialized. The antecedent part uses error partitioning to form the rule activation degree. The local control outputs of each rule region in the consequent part are weighted and synthesized to obtain the total control quantity. The parameters in each local control output correspond to the expected state quantity, the actual state quantity, and the control error information, thus giving it clear engineering and physical meaning, which facilitates subsequent analysis, constraint, and updating. During the later stages of operation, the consequent part only updates the parameters of each rule region (freezing or regularizing some rules / parameters when necessary). This allows for adaptive compensation for factors such as dead zone, saturation, friction, noise, and parameter mismatch while maintaining the interpretability of the model reference structure, without the need for precise modeling.

[0051] After obtaining the baseline gain parameters mentioned above, a Takagi-Sugeno fuzzy controller is further constructed, and the baseline gain parameters are mapped to the initial values ​​of the parameters of each rule region in the consequent part. This allows the Takagi-Sugeno fuzzy controller to inherit the basic control capabilities of the baseline control law in the initial stage and improve the availability and stability of the initial closed-loop operation.

[0052] The Takagi-Sugeno fuzzy controller consists of an antecedent and a consequent. The antecedent describes the degree to which the current control condition belongs to each rule region, while the consequent provides the local control law corresponding to each rule. The final control output is obtained by weighting and summing the local control results of each rule according to the degree of rule activation.

[0053] Optionally, the input variables are divided into multiple rule regions, and a membership function for each rule region is constructed, including the following steps: The control error information is normalized to obtain the normalized error. The normalization process uses a preset error calibration value to map the control error information to a preset range. The range of normalization error is divided into multiple regular regions, including the negative error region, the error region near zero, and the positive error region. For each regular region, a membership function is constructed. The membership function includes Gaussian function, triangular function, trapezoidal function or bell function.

[0054] Specifically, in this embodiment, the selection of the antecedent variable includes the following steps: To characterize the current tracking state of the artillery servo system, it is necessary to first select antecedent variables. Antecedent variables can be selected from system tracking error, error rate of change, system output, error integral, or a combination thereof. In this embodiment, taking the speed loop control of the artillery servo system as an example, the system control input can also be a motor current command, torque command, voltage command, or their equivalent driving quantity, and the system output is selected as the actual angular velocity of the turret azimuth axis or the barrel elevation axis. That is:

[0055] in, Discrete time The desired angular velocity command, Discrete time Actual angular velocity output, This represents the tracking error at the current moment.

[0056] In this embodiment, tracking error is selected as the antecedent variable because: tracking error can directly reflect the current control effect and deviation direction; at the same time, using only a single antecedent variable can reduce the antecedent dimension, reduce the number of fuzzy rules, and facilitate engineering implementation and real-time calculation.

[0057] It should be noted that the present invention does not limit the antecedent variable to only error; in other embodiments, error change rate, output quantity or integral quantity can also be introduced as additional antecedent variables according to the characteristics of the controlled object, so as to further improve the accuracy of rule division.

[0058] The construction of error partitioning and membership functions includes the following steps: After determining the antecedent variable, the range of values ​​for the antecedent variable is partitioned to form multiple regular regions.

[0059] Error partitioning refers to dividing tracking errors into zones. The range of values ​​is divided into several sub-intervals, with different sub-intervals corresponding to different control conditions. For example, the error range can be divided into the negative error region, the error region near zero, and the positive error region; when higher division accuracy is required, it can be further subdivided into multiple regions such as negative large, negative small, near zero, positive small, and positive large.

[0060] In a typical embodiment, the error can first be normalized to obtain the normalized error. : in, A preset error calibration value is used to normalize the error to a preset range, such as [-1, 1]. Subsequently, a membership function can be constructed for each error partition. In one embodiment, the first... The membership function corresponding to each rule can be expressed using a Gaussian function.

[0061] in, Indicates normalization error For the first The membership degree of a rule; Indicates the first The center location of each error partition; Indicates the first The width parameter of each error partition. The membership value is usually between 0 and 1. The larger the value, the more the current error belongs to the working condition area described by the corresponding rule.

[0062] In practical applications, membership functions can also take other forms such as triangular functions, trapezoidal functions, or bell-shaped functions, as long as they can characterize the degree to which the error belongs to different regular regions. This invention uses a Gaussian membership function, such as... Figure 2 As shown.

[0063] Optionally, the activation weight of each rule region is determined in the following way: The membership function value corresponding to each rule region is used as the original activation degree of that rule region. The original activation values ​​of each rule region are normalized to obtain the normalized activation value, which is the ratio of the original activation value of each rule region to the sum of the original activation values ​​of all rule regions. The activation weight of each rule region is determined based on the normalized activation degree, where the sum of the normalized activation degrees of each rule region is 1.

[0064] In this embodiment, determining the rule activation degree includes the following steps: After obtaining the membership degree of each rule, the activation degree of each rule can be further determined.

[0065] In this embodiment, since only one antecedent variable, error, is used, therefore the first... The original activation of the rule It can be directly given by the corresponding membership function value, that is: To facilitate the weighted fusion of the outputs of each rule, the activation values ​​of each rule can be further normalized to obtain normalized activation values: in, To represent the total number of fuzzy rules, Indicates the first The normalized activation of each rule at the current moment. After normalization, the sum of the activations of each rule is 1, which makes it easier to perform a weighted summation of the local control outputs of each rule to obtain the final control quantity.

[0066] In this embodiment of the invention, constructing a local control law for a rule consequent includes the following steps: After determining the antecedent rules, the consequent local control laws corresponding to each rule are further constructed. To ensure that the subsequent learning parameters have clear physical meaning and remain consistent with the baseline control law in step 1, in a typical embodiment, the first... The consequent local control law of a rule can be written as:

[0067] in, For the first The local control output corresponding to each rule; For the first The parameter vectors of each rule region in the consequent part of a rule; The regression vector used for the local control law.

[0068] In one embodiment, the regression vector can be represented as: in, For discrete error integral terms, To control the sampling period, each component in the regression vector has the following meaning:

[0069] This represents the desired state quantity, used to reflect the current control objective; Represents the actual state quantity, used to reflect the current system response state; It represents error accumulation information, which is used to reduce steady-state error and enhance continuous tracking capability.

[0070] Therefore, the parameters in the consequent local control law correspond to the weighting coefficients of the target input term, output feedback term, and error integral term, respectively, and have clear engineering and physical meanings, facilitating subsequent analysis, constraints, and updates. It should be noted that in other embodiments, the consequent regression vector can also be extended or replaced according to specific control requirements, as long as it maintains the structure and interpretability of the local control law.

[0071] Optionally, the local control outputs of each rule region in the consequent are determined in the following way: Construct local control laws for each rule region. The local control laws are expressed as the product of the parameter vectors of each rule region and the regression vector in the consequent part. In this context, each element in the parameter vector of each rule region in the consequent part corresponds to the weighting coefficients of the desired state quantity, the actual state quantity, and the control error information, respectively. The initial values ​​of the parameter vectors of each rule region in the consequent part are obtained from the baseline gain parameter. The regression vector includes the desired state quantity, the actual state quantity, and the control error information.

[0072] Optionally, the baseline gain parameter is mapped to the initial values ​​of the parameters of each regular region in the consequent part, including the following steps: The baseline gain parameter is directly assigned to the parameter vector of each regular region in the consequent part of each regular region, so that each regular region has the same baseline control law in the initial state. Alternatively, based on the error partitioning characteristics corresponding to each rule region, the baseline gain parameter can be scaled, biased, or otherwise equivalently mapped to obtain differentiated initial values ​​for different rule regions.

[0073] Optionally, the total control quantity is obtained by weighting and synthesizing the local control outputs according to the activation weights, including the following steps: Multiply the local control output of each rule region by its corresponding normalized activation degree to obtain the weighted control component of each rule region; The total control output of the Takagi-Sugeno fuzzy controller is obtained by summing the weighted control components of each rule region.

[0074] In this embodiment of the invention, the initial mapping from baseline gain parameters to parameters of each regular region includes the following steps: After constructing the structure of each rule region in the consequent part, the baseline gain parameters obtained above need to be used to initialize the parameters of each rule region in the consequent part. In a typical embodiment, the baseline gain parameter vector obtained above can be directly assigned to the parameter vector of each rule region in the consequent part of each rule, that is:

[0075] in, The baseline gain parameter vector obtained from the above steps, For the first The parameter vectors of each rule region in the consequent part of the rule at the initial time.

[0076] When the above method is used, each rule can have the same baseline control law in the initial state, thereby ensuring that the Takagi-Sugeno fuzzy controller has a basically usable closed-loop control capability before it starts learning or in the early stages of learning.

[0077] In the subsequent learning process based on closed-loop data, the parameter vectors of each rule region in the consequent of each rule can be gradually differentiated under different operating conditions, thereby forming a local control law adapted to different error regions. In other embodiments, the baseline gain parameter can also be scaled, biased, or otherwise equivalently mapped according to the error partition characteristics corresponding to different rules to obtain differentiated initial values ​​for different rules.

[0078] In this embodiment of the invention, the total control output of the Takagi-Sugeno fuzzy controller includes the following steps: After obtaining the normalized activation values ​​and local control outputs for each rule, the total control output of the Takagi-Sugeno fuzzy controller can be expressed as: Substituting further into the local control law, we get: in This is the total control output at the current moment.

[0079] Therefore, the final control input of the Takagi-Sugeno fuzzy controller is essentially the result of a weighted fusion of the local linear control laws according to the activation degree of the current operating condition. This structure retains both the fuzzy controller's ability to switch local control laws according to the operating condition and the physical interpretability of local parameters, thus providing a structural basis for subsequent data-driven optimization of the parameters of each rule region only for the consequent.

[0080] In this invention, the membership structure and its parameters of the antecedent can be preset and can remain fixed or undergo only limited fine-tuning during the learning process; the subsequent reinforcement learning process mainly updates the parameters of each rule region of the consequent to achieve performance optimization under the constraint of the structure.

[0081] Optionally, the Twin Delayed Deep Deterministic Policy Gradient algorithm can be used to optimize and update the parameters of each rule region in the consequent part.

[0082] Optionally, the parameters of each rule region in the consequent part are optimized and updated, including the following steps: Two independent evaluation networks are constructed to estimate the value of state-action pairs, and a target value is constructed based on the target action generated by the target policy network. The target value is taken as the smaller value between the outputs of the two evaluation networks to reduce the overestimation of value. By minimizing the evaluation network loss function and updating the evaluation network parameters, the evaluation network learns the long-term control effect under different states and control actions. The parameters of each rule region are optimized based on the output of the evaluation network for the consequent part of each rule region.

[0083] In this embodiment of the invention, after the Takagi-Sugeno fuzzy controller is initialized, interactive data is collected during the closed-loop operation of the artillery servo system. The Twin Delayed Deep Deterministic PolicyGradient algorithm is used to optimize the parameters of each rule region in the subsequent part. This allows the controller to gradually correct local control parameters based on actual operating data while maintaining structural interpretability, thereby improving its adaptability to nonlinear factors and changes in operating conditions.

[0084] It should be noted that the optimization target in this step is not all the weights of the black-box neural network policy, but rather the parameters of each rule region in the consequent part of the Takagi-Sugeno fuzzy controller mentioned above; the membership structure and its parameters of the antecedent can remain fixed, or only undergo limited fine-tuning. The Twin Delayed Deep Deterministic Policy Gradient algorithm optimizes and updates the parameters of each rule region in the consequent part, including the following steps:

[0085] During the operation of the artillery servo system, at each discrete moment Collect closed-loop interaction data and construct sample quintuples: in, Indicates time The system status; Indicates time Control actions; Indicates time Instant rewards; Indicates time The state at the next moment; This indicates the end-of-round flag, set to 1 when the control process corresponding to the current sample ends, and 0 otherwise.

[0086] In a typical embodiment, state This may include system tracking error, error integral, actual output, and other information that can characterize the current control condition; control actions. Control commands output by the Takagi-Sugen fuzzy controller; instant rewards It is used to evaluate the current control effect and can be constructed based on tracking error, control smoothness, control energy, or a combination thereof.

[0087] By continuously collecting the above samples, a closed-loop interactive dataset reflecting the actual operation of the artillery servo system can be obtained.

[0088] To improve sample utilization efficiency and reduce the correlation between consecutively sampled data, the aforementioned sample quintuples can be sequentially stored in the experience replay pool. During parameter updates, a small batch of samples is randomly drawn from the experience replay pool as training data for the current round of parameter optimization. The purpose of employing the experience replay mechanism is twofold: firstly, to improve the reuse rate of collected data, and secondly, to reduce the impact of strong correlation between samples at adjacent time points on the stability of parameter updates.

[0089] Construct two independent evaluation networks and Used to estimate state-action pairs The value of this. For samples obtained from the experience replay pool, a target value can be constructed:

[0090] in, Discount factor; Represents the target evaluation network; This indicates that the target policy network determines the state based on the next time step. The generated target action. Using two evaluation networks and taking the smaller value from the target values ​​can mitigate the overestimation problem that may be caused by a single evaluation network, thereby improving the stability of parameter learning.

[0091] Furthermore, the two evaluation networks are updated by minimizing the following loss function: in, This represents the current number of samples in the mini-batch.

[0092] Through the above process, the evaluation network can gradually learn the long-term control effects under different states and control actions.

[0093] After the evaluation network is updated, the parameters of the Takagi-Sugeno fuzzy controller are further updated using the deterministic policy gradient method.

[0094] In this invention, the strategy function It is not a typical black-box neural network, but rather the Takagi-Sugeno fuzzy controller constructed in the steps described above, whose parameters... It mainly consists of the parameters of each rule region in the consequent part.

[0095] By maximizing the evaluation network The policy parameters are updated based on the evaluation value of the current policy's output action, i.e.: Since the control output of the Takagi-Sugeno fuzzy controller is obtained by weighted summation of the local consequents of each rule according to their activation degree, the above update is essentially a correction of the local control coefficients of each rule under different operating conditions based on closed-loop data feedback.

[0096] It should be noted that in this invention, the learning process mainly involves iteratively updating the parameters of each rule region in the consequent part; the antecedent partition structure can remain fixed, or only be finely adjusted in small steps, in order to ensure the stability and interpretability of the overall controller structure.

[0097] To further improve learning stability, in a typical embodiment, the policy network is updated less frequently than the evaluation network; that is, the policy parameters are updated only after the evaluation network has been updated several times.

[0098] In addition, a target policy network and a target evaluation network can be set up, and the target network parameters can be updated using a soft update method.

[0099] For example, the target network parameters can be updated in the following way: in, Indicates the current network parameters. Indicates the target network parameters. Represents the soft update coefficient, and By employing delayed updates and a soft update mechanism for the target network, the risks of oscillations and divergences during parameter updates can be reduced, and the smoothness of the closed-loop learning process can be improved.

[0100] After the above update process, this step finally yields a set of rule region parameters for the consequent part after closed-loop interactive data optimization.

[0101] Compared with the initialization parameters, the updated parameters of each rule region in the successor part can better adapt to non-ideal factors such as dead zones, saturation, friction, noise interference and parameter changes that occur in the actual operation of the artillery servo system.

[0102] Because this invention maintains the overall structure of the Takagi-Sugeno fuzzy controller unchanged during the optimization process, and primarily updates the parameters of each rule region in the consequent part, the resulting controller retains characteristics such as clear rule partitioning, interpretable local parameters, and ease of constraint and engineering tuning. It employs TD3's dual Critic + delayed update mechanism, but the target is the parameters of each rule region in the structured fuzzy consequent part rather than the black-box network strategy; thus achieving stable and controllable offline / online updates.

[0103] Furthermore, in practical applications, mechanisms such as rule freezing, integral channel correction, downsampling updates, and parameter constraints can be combined to impose security restrictions and engineering constraints on the above learning process. The optimization and updating of parameters for each rule region in the consequent part also includes one of the following or engineering constraints: The parameters of each rule region in the consequent part of the selected key rule region remain fixed and are not updated, while the parameters of each rule region in the consequent part of the other rule regions participate in learning and optimization. The key rule regions include rule regions where the absolute value of the error is less than the preset error threshold, boundary rule regions, or rule regions that have achieved the expected control effect. A downsampling strategy is adopted to update the parameters once every D control cycles, and the parameter iteration frequency is set to 1 / D of the control frequency; Apply upper and lower bound constraints or norm constraints to the parameters of each rule region in the consequent part of each rule region.

[0104] Specific engineering constraints / safety mechanisms (optional during iteration due to the interpretability of the controller) include the following steps: (1) Prior fixed: Key rules can be selected (such as small error zone or boundary rules, or certain situations can be determined). (The expected control effect of the artillery has been achieved) Maintain No updates; the remaining rules are learned to reduce the risks of online optimization of physical objects and enhance the applicability of the project.

[0105] (2) Integral channel σ-mod correction: The update of the integral channel parameters can be represented as the superposition of a correction term related to the deviation of the current parameter from the reference value on the basis of the original gradient update, so that the integral channel parameters converge to a preset reasonable range during the learning process, and suppress parameter drift caused by noise / integral bias.

[0106] (3) Iteration update frequency: The downsampling update strategy is adopted, and the update is performed once every D control cycles (i.e., the iteration frequency is 1 / D of the control frequency). The fast closed-loop control process and the relatively slow parameter learning process are appropriately separated on the time scale to avoid frequent changes in parameters within each control sampling cycle, thereby reducing the control output fluctuation caused by rapid parameter jitter and improving the stability of the physical system operation.

[0107] (4) Parameter projection / gain limiting: for Applying upper and lower bounds or norm constraints ensures that the control quantity and closed-loop bandwidth remain within safe ranges. This prevents parameters from abnormally increasing, reversing their sign, or exceeding acceptable engineering ranges during the learning process, thus avoiding excessive control output, excessively high closed-loop bandwidth, or increased risk of system oscillation.

[0108] Optionally, the actual state quantities include one of current, angular velocity, or angular position, and the control error information includes one or more of the following: system tracking error, the rate of change of system tracking error, and the integral of system tracking error.

[0109] Compared with the prior art, the present invention also has the following advantages: The method provided in this embodiment of the invention does not rely on long-term online updates of analytical adaptive laws. Instead, it uses baseline parameters as interpretable initialization and then optimizes rule consequents using closed-loop data from TD3. Furthermore, it introduces combined mechanisms such as rule freezing, σ-mod, and downsampling updates to adapt to real-world learning, eliminating the need for a precise model of the controlled object.

[0110] The method provided in this invention restricts the strategy structure to error partitioning + local linear consequents, and uses baseline parameters embedded in the consequents as a stable starting point. This makes the method interpretable and easy to constrain, reducing exploration risks. The interpretability of fuzzy rules is improved.

[0111] The learning units of this method are aligned with the MRAC regression structure (r, y, I_u), and the learned data is rule-level local gain rather than the output of an additional black-box compensator; therefore, it is more transparent and easier to verify and engineer. The interpretability of fuzzy rules is improved.

[0112] The practical effects that artillery can achieve: Explainable, easy for engineering analysis and safety constraint implantation: The method of this invention adopts a control law structure of "error partitioning - rule weighting - local linear consequents", and the control output is obtained by linear combination of each rule consequent. Compared with the black box function approximation strategy structure, it is easier to conduct mechanism analysis and parameter tuning of the stability, saturation constraints and error convergence process of the artillery servo system. It can also reduce the exploration risk of online learning of physical artillery by combining rule freezing, gain constraints and other means.

[0113] Reducing reliance on precise models and adapting to the challenging modeling characteristics of artillery servo systems: This invention only requires low-order nominal model and reference model parameters for initialization to obtain MRAC baseline parameters and map them to T-S fuzzy rule consequents; subsequently, nonlinear effects (such as dead zone, saturation, friction, parameter changes / mismatches, etc.) are compensated and corrected through closed-loop interactive data-driven parameter iteration. Therefore, in artillery servo scenarios where precise models are difficult to obtain or operating conditions change frequently, it can improve adaptability to nonlinearity and uncertainty while maintaining the reference model tracking framework.

[0114] The training process is more controllable and the risk of divergence is reduced: Since the controller parameters can provide a stable starting point and limit the learning to the parameters of each rule region that mainly affect the consequent (the antecedent partition can be fixed or finely adjusted), combined with mechanisms such as dual evaluation network, delayed update and update frequency configuration, the value estimation bias and drastic parameter drift can be suppressed, making the online optimization process smoother and more controllable, and improving the deployability and safety under the closed-loop condition of artillery servo physical objects.

[0115] In summary, this method constructs a controller composed of interpretable rules without relying on a precise system model, and achieves compensation for nonlinearity and uncertainty through data-driven iteration. Therefore, it is easier to apply safety constraints and achieve safer and more controllable online optimization and performance improvement in the practical application of artillery servo systems.

[0116] Typical Implementation Example (Artillery Servo Speed ​​Loop) Object: Taking the pitch speed loop of a gun servo system as an example. System input is... The system outputs the shaft angular velocity as the desired angular velocity command. (The feedback signal can be obtained from encoder differential or gyroscope). Reference Model Tuning can be performed based on a low-order nominal model of the shaft system (such as first-order inertia + gain), and the second-order dynamics corresponding to the PI-type reference model can be selected as the desired closed-loop response. The reference model output is... The controller input is the error. The output is the motor current / torque command. .

[0117] Model nonlinearity assumption: 1. Dead zone saturation 2. Measurement of noise .

[0118] State / Feature Design: MDP State Desirability ,in for Integral term; Critic feature Actor characteristics Rewards can be set .

[0119] Training and update strategy: Run the closed loop and collect samples at a fixed sampling period Ts; perform parameter update once every D steps based on the experience pool data; Critic can be trained with each update; Actor further adopts a strategy to delay updates (e.g., update Actor once every 2 Critic updates).

[0120] Deployment: First verify the stability of the initial values ​​of this method in simulation / semi-physical verification, and then gradually expand the learning.

[0121] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. An adaptive control method for artillery servo systems, characterized in that, Includes the following steps: Construct a low-order nominal model and a reference model based on the artillery servo system, and determine the baseline gain parameters based on the low-order nominal model and the reference model. A Takagi-Sugeno fuzzy controller is constructed, which includes an antecedent part and a consequent part. The antecedent part uses the control error information of the system as the input variable, divides the input variable into multiple regular regions, constructs the membership function of each regular region, obtains the activation degree of each regular region, and thus determines the activation weight of each regular region. Each rule region of the aftermath is a parameterized structure consistent with the MRAC controller, and the baseline gain parameter is mapped to the initial value of each rule region parameter of the aftermath. During closed-loop operation, the expected and actual state variables of the artillery servo system are acquired, and the control error information of the system is calculated. The control error information is input into the preceding part to determine the activation weight of each rule region; The desired state quantity, actual state quantity, and control error information are input into the subsequent part to obtain the local control output of each rule region of the subsequent part; the local control output is weighted and synthesized according to the activation weight to obtain the total control quantity, which is output to the actuator of the artillery servo system to drive the artillery servo system to run. The parameters of each rule region in the subsequent part are optimized and updated by using the expected state variables, actual state variables, and control error information of the system obtained during closed-loop operation.

2. The adaptive control method for an artillery servo system as described in claim 1, characterized in that, The process of dividing the input variables into multiple rule regions and constructing the membership function for each rule region includes the following steps: The control error information is normalized to obtain a normalized error. The normalization process uses a preset error calibration value to map the control error information to a preset range. The normalization error range is divided into multiple regular regions, including a negative error region, a zero-near error region, and a positive error region. For each regular region, a membership function is constructed, which includes a Gaussian function, a triangular function, a trapezoidal function, or a bell-shaped function.

3. An adaptive control method for an artillery servo system as described in claim 1 or 2, characterized in that, The activation weights of each rule region are determined in the following way: The membership function value corresponding to each rule region is used as the original activation degree of that rule region. The original activation values ​​of each rule region are normalized to obtain normalized activation values, which are the ratios of the original activation values ​​of each rule region to the sum of the original activation values ​​of all rule regions. The activation weight of each rule region is determined based on the normalized activation degree, wherein the sum of the normalized activation degrees of each rule region is 1.

4. The adaptive control method for an artillery servo system as described in claim 1, characterized in that, The local control outputs of each rule region in the consequent are determined in the following manner: Construct local control laws for each rule region, whereby the local control laws are expressed as the product of the parameter vectors of each rule region and the regression vector in the consequent part; In this context, each element in the parameter vector of each rule region of the consequent part corresponds to the weighting coefficient of the expected state quantity, the actual state quantity, and the control error information, respectively. The initial value of each parameter vector of the rule region of the consequent part is obtained from the baseline gain parameter. The regression vector includes the expected state quantity, the actual state quantity, and the control error information.

5. The adaptive control method for an artillery servo system as described in claim 4, characterized in that, The baseline gain parameter is mapped to the initial values ​​of the parameters of each regular region in the consequent part, including the following steps: The baseline gain parameter is directly assigned to the parameter vector of each regular region in the consequent part of each regular region, so that each regular region has the same baseline control law in the initial state. Alternatively, based on the error partitioning characteristics corresponding to each rule region, the baseline gain parameter can be scaled, biased, or otherwise equivalently mapped to obtain differentiated initial values ​​for different rule regions.

6. The adaptive control method for artillery servo systems as described in claim 3, characterized in that, The process of weighting and synthesizing the local control outputs according to the activation weights to obtain the total control quantity includes the following steps: Multiply the local control output of each rule region by its corresponding normalized activation degree to obtain the weighted control component of each rule region; The total control output of the Takagi-Sugeno fuzzy controller is obtained by summing the weighted control components of each rule region.

7. An adaptive control method for an artillery servo system as described in any one of claims 1, 2, and 4, characterized in that, The Twin Delayed Deep Deterministic Policy Gradient algorithm is used to optimize and update the parameters of each rule region in the consequent part.

8. The adaptive control method for an artillery servo system as described in claim 7, characterized in that, The parameters of each rule region in the consequent section are optimized and updated, including the following steps: Two independent evaluation networks are constructed to estimate the value of state-action pairs, and a target value is constructed based on the target action generated by the target policy network. The target value is taken as the smaller value between the outputs of the two evaluation networks to reduce the overestimation of value. By minimizing the evaluation network loss function and updating the evaluation network parameters, the evaluation network learns the long-term control effect under different states and control actions. The parameters of each rule region are optimized based on the output of the evaluation network for the consequent part of each rule region.

9. The adaptive control method for an artillery servo system as described in claim 7, characterized in that, The process of optimizing and updating the parameters of each rule region in the subsequent part also includes one of the following or engineering constraints: The parameters of each rule region in the consequent part of the selected key rule region remain fixed and are not updated, while the parameters of each rule region in the consequent part of the other rule regions participate in learning and optimization. The key rule region includes rule regions with an absolute error value less than a preset error threshold, boundary rule regions, or rule regions that have achieved the expected control effect. A downsampling strategy is adopted to update the parameters once every D control cycles, and the parameter iteration frequency is set to 1 / D of the control frequency; Apply upper and lower bound constraints or norm constraints to the parameters of each rule region in the consequent part of each rule region.

10. The adaptive control method for an artillery servo system as described in claim 1, characterized in that, The actual state quantities include one of current, angular velocity, or angular position, and the control error information includes one or more of system tracking error, the error rate of change of system tracking error, and the error integral of system tracking error.

Images