A distributed optimization learning control method and system for robot joint modules

By building a model prediction controller, state observer and reinforcement learning model, optimizing the control input of the permanent magnet synchronous motor, the performance degradation of traditional methods under model inaccuracy and external perturbation is solved, and dynamic response and robustness are improved.

CN115890668BActive Publication Date: 2024-11-01SHANGHAI UNIVERSITY OF ELECTRIC POWER
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211445280.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-11-18
Publication Date
2024-11-01
Estimated Expiration
2042-11-18

AI Technical Summary

Technical Problem

Traditional model prediction control methods are sensitive to the parameters of permanent magnet synchronous motors, and are complex in design, making it difficult to maintain control effects in the presence of model inaccuracy, parameter changes and external disturbances, and are not dynamically responsive and robust.

Method used

The robot joint module dispersed optimization learning control method is adopted, and the motion equation of the independent motor control system is established by constructing a model prediction controller, a state observer and a reinforcement learning model, and the photoelectric encoder feedback signal, current signal and magnetic field orientation theory is used to establish the motion equation of the independent motor control system. Construct state space models and state observers, optimize control inputs, and optimize variable prediction time domains through reinforcement learning.

Benefits of technology

It improves position prediction accuracy and system robustness, enhances adaptability, maintains good control effects in complex environments, and reduces dependence on parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115890668B_ABST
    Figure CN115890668B_ABST
Patent Text Reader

Abstract

The present invention relates to a distributed optimization learning control method and system for a robot joint module. The method first realizes the correction of the prediction model deviation by constructing a state observer. Secondly, the optimal control law is obtained by solving the optimization problem of the position following error performance index. Further, self-learning of the prediction time domain is realized by designing a reinforcement learning network. Compared with the prior art, the present invention uses input and output signals to synchronously estimate unmeasurable states and lumped disturbances, and compensates the estimated information of the disturbance and the rotor angular velocity in the output position prediction process, thereby improving the position prediction accuracy and the robustness of the system. At the same time, based on the control idea of ​​optimization learning, it can solve the problems of inaccurate model, decreased control effect when parameters change and external disturbances exist in the system, improve the adaptive ability of the system, and show that the joint module has better robustness to complex environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of robot control and reinforcement learning, and in particular to a distributed optimization learning control method and system for a robot joint module. Background Art

[0002] Since the rise of robotics, in order to meet the growing demand for automated production, the industrial field has taken the lead in opening up a large-scale application of robots and has achieved large-scale development in a very short period of time. The robotic arm is mainly composed of joint modules and connecting rods in series, and its joint module is composed of a servo system controller, a servo motor, a transmission device and a sensor. With the continuous advancement of high-performance motion control technology, the joint module composed of a servo motor and a driver has been widely used in various fields. Among them, permanent magnet synchronous motors have the advantages of high efficiency, light weight, and high power density, so they are widely used in robot servo systems. As the core module of robot actuators, permanent magnet synchronous motors are usually required to have excellent position servo performance to meet the working rhythm and positioning accuracy requirements of assembly line operations. However, permanent magnet synchronous motors are multivariable, strongly coupled and have external interference nonlinear control objects. In this case, traditional linear control methods are difficult to guarantee the operating performance of joint servo systems. With the rapid development of nonlinear control theory, such as adaptive control, sliding mode control, model predictive control, etc., have been successfully applied to the robust control of permanent magnet synchronous motors, but these methods are highly dependent on motor parameters and controller parameters are not easy to adjust. Among them, the idea of ​​model predictive control is to combine the current system's measured values ​​and the prediction model to predict the control output in the future, and to minimize the system's error in the future by solving various constrained optimization problems. Based on this, the model predictive controller is more robust than other control methods, has higher accuracy, and is more widely used. The traditional model predictive control strategy has the following shortcomings:

[0003] 1. Control strategies based on discrete time models, such as the Chinese patent application number CN201710727012.8 that discloses a position prediction control method for a permanent magnet synchronous servo system, and the literature (Xu Nan et al., Improved Predictive Current Control of Permanent Magnet Synchronous Motor [J]. Motor and Control Applications, 2017, 44(07)). The design process of this type of controller is greatly affected by the sampling period. A sampling period that is too small is likely to lead to a larger prediction time domain, which in turn increases the amount of calculation and affects real-time performance. A sampling period that is too large will lead to a decrease in the system's anti-disturbance ability, which increases the difficulty of controller design in practical applications.

[0004] 2. Control strategy based on continuous-time model, such as the Chinese patent application number CN202010970766.8, which discloses a cascade control method for permanent magnet synchronous motor based on generalized predictive control, and the literature (Chen Wei et al., Generalized predictive cross-coupling position control of dual-axis linkage system [J]. Control Theory and Applications, 2018, 35(03)). First of all, generalized predictive control based on continuous-time model, as a type of model predictive control, not only has strong robustness, fast dynamic response and convenient processing of system constraints, but also has the advantages of high computational efficiency and easy engineering implementation. However, this method is based on nominal system modeling. Therefore, when there is load interference, model parameter drift, and unmodeled dynamics, the traditional method cannot guarantee that the control parameters are still adapted to the changing working conditions, and its control performance will be affected.

[0005] In summary, the traditional continuous model predictive control method relies on the nominal system model, is very sensitive to parameters, and its design and analysis requires certain manual experience. The control effect will be affected by different system conditions and time-varying disturbance factors, and the parameters are not easy to adjust. Therefore, designing an optimization learning controller with self-adjusting controller parameters has become an urgent problem to be solved. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a distributed optimization learning control method and system for a robot joint module, so as to solve the problems of inaccurate model of the motor control system, decreased control effect during parameter changes and external disturbances, and at the same time improve the dynamic response and robustness of the servo system.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] A robot joint module distributed optimization learning control method comprises the following steps:

[0009] S1, constructing a model predictive controller, wherein the model predictive controller is used to achieve optimal control of the rotor position of the motor of the robot joint module;

[0010] S2. constructing a state observer, wherein the state observer is used to correct the prediction model deviation in the model predictive controller;

[0011] S3, constructing a reinforcement learning model, wherein the reinforcement learning model is used to optimize the variable prediction time domain in the model predictive controller;

[0012] The construction of the state observer includes the following steps:

[0013] S201, collecting feedback signals from a photoelectric encoder installed on the motor rotor of the joint module to obtain the actual position of the rotor;

[0014] S202, collecting motor A and B two-phase current signals, and obtaining q-axis measurement current and d-axis measurement current through Clark transformation and Park transformation;

[0015] S203, based on the actual position of the rotor, the q-axis measured current and the d-axis measured current, according to the magnetic field orientation theory, and with the idea of ​​distributed control, establish the motion equation ∑1 of the independent motor control system in the rotating coordinate system dq;

[0016] S204, determining a given current signal of the q-axis, and constructing a motion equation ∑2 based on the given current signal, the actual position of the rotor and the motion equation ∑1;

[0017] S205, expressing the motion equation ∑2 as a state space model, and constructing a state observer based on the state space model;

[0018] The construction of the model predictive controller includes the following steps:

[0019] S101. Define optimization performance indicators;

[0020] S102, defining constraints of control input;

[0021] S103: Based on the optimization performance index and constraint conditions, define an optimization control input, and obtain the optimization control input by solving the optimization problem.

[0022] The construction of the reinforcement learning model includes the following steps:

[0023] S301, defining position tracking error;

[0024] S302, selecting a state space and an action space;

[0025] S303. Design a reward function.

[0026] Furthermore, the expression of the motion equation ∑1 of the independent motor control system in the rotating coordinate system dq is:

[0027] ∑1:

[0028] Among them, θ m (t) is the actual position of the rotor, L d , L q are the stator inductances of the d and q axes respectively; i d 、i q are the d-axis and q-axis stator currents respectively; n p is the pole pair number; ω m (t) is the actual angular velocity of the rotor; ψ f is the magnetic flux generated by the permanent magnet; J is the moment of inertia of the system; B vis the viscous friction coefficient of the system; T L is the external load torque; ΔT is the torque disturbance caused by the perturbation of model parameters.

[0029] Further, determine the q-axis given current signal Will As the system input, the actual rotor position θ m (t) is used as the system output to construct the motion equation ∑2, which is expressed as follows:

[0030] ∑2:

[0031] in: is θ m The second derivative of (t), b(t) = K t / J is the controller gain, K t is the torque constant; u=i q * is the control input; f(t) is the lumped disturbance including current loop tracking error, load torque, viscous friction and torque disturbance caused by model parameter perturbation; K t The expression is:

[0032]

[0033] Further, the actual rotor position θ is selected m (t) and the actual angular velocity of the rotor ω m (t) are state variables x1 and x2 respectively, and the motion equation ∑2 is expressed as a state space model ∑3, which is expressed as follows:

[0034] ∑3:

[0035] The state observer ∑4 is constructed based on the state space model ∑3, and its expression is:

[0036] ∑4:

[0037] Where: b0 is the nominal value of b(t), z1 is θ m (t) is the observed value, z2 is ω m (t), z3 is the observed value of f(t); l i (i=1, 2, 3) is the state observer gain, which is used to adjust the observation performance of the system.

[0038] Furthermore, the optimization performance index is defined, and its expression is:

[0039]

[0040] Where: T pis a variable prediction time domain; is the expected reference position of the system after time τ; θ m (t+τ) is the position prediction output of the system after time τ.

[0041] Furthermore, the constraints of the control input are defined as:

[0042] u min <u<u max

[0043] Where: u max for i q Maximum limit value, u min for i q Minimum limit value.

[0044] Furthermore, based on the optimization performance index and constraints, the optimization control input u is defined * ,

[0045] u * =argminJ(u)

[0046] The optimal control input u is obtained by solving the following optimization problem * :

[0047]

[0048] Where: θ r (t),ω r (t), a r (t) are the rotor reference position, rotor reference angular velocity and rotor reference angular acceleration respectively.

[0049] Furthermore, the reinforcement learning model is constructed to predict the variable prediction time domain T in the model predictive controller. p Optimize, variable prediction time domain T p The selection process is equivalent to a Markov decision process, which includes the following steps:

[0050] Define the position tracking error e(t) as:

[0051] e(t)=θ m (t)-θ r (t)

[0052] Choose the state space as:

[0053]

[0054] Choose the action space as T p , design the reward function r(t) to satisfy:

[0055]

[0056] Where: β i (i=1,2,3) is the reward coefficient, ∈ i (i=1, 2) is the error threshold;

[0057] The variable prediction time domain T is realized by the reinforcement learning model p Self-learning to obtain the optimal prediction time domain.

[0058] Furthermore, when the actual position of the rotor is close to the rotor reference position, positive rewards β1 and β2 are given, and β1>β2; when the actual position of the rotor is far from the rotor reference position, a negative reward -β3 is given.

[0059] A robot joint module distributed optimization learning control system, used to implement a robot joint module distributed optimization learning control method as described above, including a state observation module, a model prediction control module and a parameter optimization module;

[0060] The model prediction control module is used to achieve optimal control of the motor rotor position of the robot joint module;

[0061] The state observation module is used to correct the prediction model deviation in the model prediction control module;

[0062] The parameter optimization module is used to realize self-learning of the variable prediction time domain to obtain the optimal prediction time domain.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention utilizes input and output signals to synchronously estimate unmeasurable states and lumped disturbances, and compensates the disturbances and estimated information of the rotor angular velocity into the output position prediction process, thereby improving the position prediction accuracy and the robustness of the system.

[0065] 2. The present invention is based on the control concept of optimization learning, which can solve the problems of inaccurate model, decreased control effect when parameters change and external disturbances occur in the system, improves the adaptive ability of the system, and shows that the joint module has better robustness to complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is the overall system block diagram of the present invention;

[0067] Figure 2 This is a comparison chart of the position tracking effect of the optimized learning controller and the P-PI controller under constant disturbance;

[0068] Figure 3 This is a comparison chart of the position tracking effect of the optimized learning controller and the P-PI controller under sinusoidal disturbance;

[0069] Figure 4 Schematic diagram of q-axis current for optimizing learning controller under constant disturbance;

[0070] Figure 5 It is the schematic diagram of the q-axis current of the P-PI controller under constant disturbance;

[0071] Figure 6 Schematic diagram of q-axis current for optimizing learning controller under sinusoidal disturbance;

[0072] Figure 7 Schematic diagram of the q-axis current of the P-PI controller under sinusoidal disturbance. DETAILED DESCRIPTION

[0073] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0074] The present invention firstly realizes the correction of the prediction model deviation by constructing a state observer. Secondly, the optimal control law is obtained by solving the optimization problem of the position following error performance index. Furthermore, the self-learning of the prediction time domain is realized by designing a reinforcement learning network.

[0075] like Figure 1 As shown, a robot joint module distributed optimization learning control system provided by this embodiment includes a state observation module, a model prediction control module and a parameter optimization module;

[0076] Among them, the model predictive control module includes a prediction model unit, an optimization solution unit, and a performance index and constraint condition unit; the model predictive control module is used to realize the optimization control of the motor rotor position of the robot joint module;

[0077] The state observation module is connected to the model prediction control module and is used to correct the prediction model deviation in the model prediction control module;

[0078] The parameter optimization module is used to implement reinforcement learning on the variable prediction horizon to obtain the optimal prediction horizon, and pass the optimal prediction horizon to the model predictive control module.

[0079] Based on the robot joint module distributed optimization learning control system, a robot joint module distributed optimization learning control method can be implemented, and the method includes the following steps:

[0080] S1. Construct a model predictive controller, which is used to optimize the position of the motor rotor of the robot joint module;

[0081] S2. construct a state observer, which is used to correct the prediction model deviation in the model predictive controller;

[0082] S3, building a reinforcement learning model, the reinforcement learning model is used to optimize the variable prediction horizon in the model predictive controller;

[0083] The construction of the state observer includes the following steps:

[0084] Step 1) By collecting the feedback signal of the photoelectric encoder installed on the rotor of the joint module motor, the actual rotor position θ is obtained. m (t);

[0085] The motor A and B two-phase current signals are collected through the hardware ADC, and the q-axis measurement current i is obtained through Clark Transformation and Park Transformation. q and the d-axis measuring current i d ;

[0086] Step 2): Based on the actual rotor position, q-axis measured current and d-axis measured current, according to the field orientation theory and the distributed control concept, the motion equation ∑1 of the independent motor control system in the rotating coordinate system dq is established:

[0087] ∑1:

[0088] Among them, θ m (t) is the actual position of the rotor, L d , L q are the stator inductances of the d and q axes respectively; i d 、i q are the d-axis and q-axis stator currents respectively; n p is the pole pair number; ω m (t) is the actual angular velocity of the rotor; ψ f is the magnetic flux generated by the permanent magnet; J is the moment of inertia of the system; B v is the viscous friction coefficient of the system; T L is the external load torque; ΔT is the torque disturbance caused by the perturbation of model parameters.

[0089] Further, determine the q-axis given current signal As the system input, the actual rotor position θ m (t) is used as the system output to construct the motion equation ∑2:

[0090] ∑2:

[0091] in: is θ m The second derivative of (t), b(t) = K t / J is the controller gain, K t is the torque constant; u=i q * is the control input; f(t) is the lumped disturbance including current loop tracking error, load torque, viscous friction and torque disturbance caused by model parameter perturbation; K t The expression is:

[0092]

[0093] Step 3): Select the actual rotor position θ m (t) and the actual angular velocity of the rotor ω m (t) are state variables x1 and x2 respectively, and the motion equation ∑2 is expressed as a state space model ∑3, which is expressed as follows:

[0094] ∑3:

[0095] The state observer ∑4 is constructed based on the state space model ∑3, and its expression is:

[0096] ∑4:

[0097] Where: b0 is the nominal value of b(t), z1 is θ m (t) is the observed value, z2 is ω m (t), z3 is the observed value of f(t); l i (i=1, 2, 3) is the state observer gain, which is used to adjust the observation performance of the system.

[0098] The construction of the model predictive controller includes the following steps:

[0099] Step 4): Define the optimization performance index, whose expression is:

[0100]

[0101] Where: T p is a variable prediction time domain; is the expected reference position of the system after time τ; θ m (t+τ) is the position prediction output of the system after time τ.

[0102] Furthermore, the constraints of the control input are defined as:

[0103] u min <u<u max

[0104] Where: umax for i q Maximum limit value, u min for i q Minimum limit value.

[0105] Based on the optimization performance index and constraints, define the optimal control input u * :

[0106] u * =argminJ(u)

[0107] The optimal control input is obtained by solving the optimization problem of the following position following error performance index, that is, the optimal control law u * :

[0108]

[0109] Where: θ r (t),ω r (t), a r (t) are the rotor reference position, rotor reference angular velocity and rotor reference angular acceleration respectively.

[0110] The variable prediction time domain T in step 4) above p The selection process is equivalent to a Markov decision process, which includes the following steps:

[0111] Define the position tracking error e(t) as:

[0112] e(t)=θ m (t)-θ r (t)

[0113] Choose the state space as:

[0114]

[0115] When the constraint on the control energy is considered constant, choose T p As an action, adjust and explore it, and improve and optimize the action according to the size of the reward obtained after the action, that is, T p As a dynamic parameter, it is updated in real time according to environmental changes;

[0116] The current motor control system is used as the environment, which is responsible for giving the real-time status after the parameters change. The reward function defines that the reward gradually increases as the error decreases, and a positive reward is given when the error decreases to a certain extent. The reward function is defined as:

[0117]

[0118] Where: β i (i=1,2,3) is the reward coefficient, ∈ i(i=1, 2) is the error threshold;

[0119] The variable prediction time domain T is realized through reinforcement learning model p Self-learning is performed to obtain the optimal prediction time domain. When the actual rotor position is close to the rotor reference position, positive rewards β1 and β2 are given, and β1>β2; when the actual rotor position is far from the rotor reference position, a negative reward -β3 is given, thereby maximizing the reward function to improve dynamic performance.

[0120] In specific implementation, the electrical parameter specifications of the permanent magnet synchronous motor selected in the robot joint module of this embodiment are shown in Table 1:

[0121] Table 1 Electrical parameters of permanent magnet synchronous motor

[0122]

[0123] As a specific implementation case, this embodiment selects the DDPG reinforcement learning algorithm for T p Self-learning. In other embodiments, other types of reinforcement learning algorithms can be selected according to different needs to achieve the same function. Among them, the Critic network in the DDPG reinforcement learning algorithm has two hidden layers, each with 64 units; there are 5 units in the input layer, including 4 states and 1 action. The Actor network has one hidden layer, a fully connected layer consisting of 128 neurons, and the output layer activation function is selected as a hyperbolic tangent function. In addition, the gain selection in the state observer ∑4 satisfies l1=3ω0, l2=3ω0 2 , l3=ω0 3 , where ω0 is the observer bandwidth factor.

[0124] The effectiveness of the present invention is illustrated by simulation experiments. In combination with the permanent magnet synchronous motor joint module selected in this example, as shown in Table 2, the parameters of the optimization learning controller and the P-PI dual closed-loop controller of the present invention are respectively set as follows:

[0125] Table 2 Controller parameter selection

[0126]

[0127] Where: K θ,p and K ω,p They are the proportional gain coefficients of the position loop and the speed loop, K ω,i is the speed loop integral gain coefficient.

[0128] The following two working conditions are designed:

[0129] 1) Given θ r =500° step signal, the disturbance torque is set to T at 0.2s L=0.1Nm.

[0130] 2) Given θ r =500° step signal, the disturbance torque is set to T at 0.2s L =0.1sin(2πt)Nm.

[0131] Figures 2 to 5 The position tracking curves and q-axis current response curves of the optimized learning controller proposed in the present invention and the traditional P-PI dual closed-loop controller under constant load disturbance and sinusoidal load disturbance are given respectively. It can be clearly seen from the experimental results that the optimized learning control design method provided by the present invention can well suppress different disturbances in the system. In particular, it can be seen from the position output curve of the sinusoidal load disturbance that the optimized learning control method provided by the present invention has been significantly improved compared with the traditional dual closed-loop controller, and the prediction period T of the predictive control is p ,According to the system model combined with the reinforcement learning mechanism, dynamic update,the controller parameters are self-learning, which improves the dynamic response and robustness of the servo system.

[0132] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A distributed optimization learning control method for a robot joint module, characterized in that: The following steps are involved: Constructing a model predictive controller, wherein the model predictive controller is used to achieve optimal control of the rotor position of the motor of the robot joint module; Constructing a state observer, wherein the state observer is used to correct the prediction model deviation in the model predictive controller; Constructing a reinforcement learning model for optimizing a variable prediction horizon in a model predictive controller; The construction of the state observer includes the following steps: Collect the feedback signal of the photoelectric encoder installed on the motor rotor of the joint module to obtain the actual position of the rotor; Collect the motor A and B two-phase current signals, and obtain the q-axis measurement current and d-axis measurement current through Clark transformation and Park transformation; Based on the actual position of the rotor, the q-axis measured current and the d-axis measured current, according to the magnetic field orientation theory and the distributed control idea, the motion equation ∑1 of the independent motor control system in the rotating coordinate system dq is established; Determine a given current signal of the q-axis, and construct a motion equation ∑2 based on the given current signal, the actual position of the rotor and the motion equation ∑1; Representing the motion equation ∑2 as a state space model, and constructing a state observer based on the state space model; The construction of the model predictive controller includes the following steps: Define optimization performance indicators; Define constraints that control inputs; Based on the optimization performance index and constraint conditions, an optimization control input is defined, and the optimization control input is obtained by solving the optimization problem; The construction of the reinforcement learning model includes the following steps: Define position tracking error; Select state space and action space; Design reward functions; The expression of the motion equation ∑1 of the independent motor control system in the rotating coordinate system dq is: ∑1: Among them, θ m (t) is the actual position of the rotor, L d , L q are the stator inductances of the d and q axes respectively; i d 、i q are the d-axis and q-axis stator currents respectively; n p is the pole pair number; ω m (t) is the actual angular velocity of the rotor; ψ f is the magnetic flux generated by the permanent magnet; J is the moment of inertia of the system; B v is the viscous friction coefficient of the system; T L is the external load torque; Δf is the torque disturbance caused by the perturbation of model parameters; Determine the q-axis given current signal Will As the system input, the actual rotor position θ m (t) is used as the system output to construct the motion equation ∑2, which is expressed as follows: ∑2: in: is θ m The second derivative of (t), b(t) = K t / J is the controller gain, K t is the torque constant; u=i q * is the control input; f(t) is the lumped disturbance including current loop tracking error, load torque, viscous friction and torque disturbance caused by model parameter perturbation; K is expressed as: Select the actual rotor position θ m (t) and the actual angular velocity of the rotor ω m (t) are state variables x1 and x2 respectively, and the motion equation ∑2 is expressed as a state space model ∑3, which is expressed as follows: ∑3: The state observer ∑4 is constructed based on the state space model ∑3, and its expression is: ∑4: Where: b0 is the nominal value of b(t), z1 is θ m (t) is the observed value, z2 is ω m (t), z3 is the observed value of f(t); l i (i=1, 2, 3) is the state observer gain, which is used to adjust the observation performance of the system.

2. A robot joint module distributed optimization learning control method according to claim 1, characterized in that: Define the optimization performance index, its expression is: Where: T p is a variable prediction time domain; is the expected reference position of the system after time τ; θ m (t+τ) is the position prediction output of the system after time τ.

3. A robot joint module distributed optimization learning control method according to claim 2, characterized in that: The constraints for the control input are defined as: in min <u<u max Where: u max for i q Maximum limit value, u min for i q Minimum limit value.

4. A robot joint module distributed optimization learning control method according to claim 3, characterized in that: Based on the optimization performance index and constraints, define the optimization control input u * , he * =argminJ(u) The optimal control input u is obtained by solving the following optimization problem * : Where: θ r (t),ω r (t), a r (t) are the rotor reference position, rotor reference angular velocity and rotor reference angular acceleration respectively.

5. A robot joint module distributed optimization learning control method according to claim 3, characterized in that: The reinforcement learning model is constructed to predict the variable prediction time domain T in the model predictive controller. p Optimize the variable prediction time domain T p The selection process is equivalent to a Markov decision process, which includes the following steps: Define the position tracking error e(t) as: e(t)=θ m (t)-θ r (t) Choose the state space as: Choose the action space as T p , design the reward function r(t) to satisfy: Where: β i (i=1,2,3) is the reward coefficient, ∈ i (i=1, 2) is the error threshold; The variable prediction time domain T is realized by the reinforcement learning model p Self-learning to obtain the optimal prediction time domain.

6. A robot joint module distributed optimization learning control method according to claim 5, characterized in that: When the actual position of the rotor is close to the rotor reference position, positive rewards β1 and β2 are given, and β1>β2; when the actual position of the rotor is far from the rotor reference position, a negative reward -β3 is given.

7. A robot joint module distributed optimization learning control system, characterized in that: A method for implementing a distributed optimization learning control method for a robot joint module as described in any one of claims 1 to 6, comprising a state observation module, a model prediction control module and a parameter optimization module; The model prediction control module is used to achieve optimal control of the motor rotor position of the robot joint module; The state observation module is used to correct the prediction model deviation in the model prediction control module; The parameter optimization module is used to realize self-learning of the variable prediction time domain to obtain the optimal prediction time domain.