Multi-axis mechanical arm predictive control method based on information physical neural network

Abstract

The invention belongs to the technical field of control, and discloses a multi-axis mechanical arm predictive control method based on an information physical neural network. Comprising the following steps that (1) a nonlinear dynamic mechanical arm model is established and configured based on the physical relation between all shafts of the mechanical arm; 2) based on the mechanical arm model, randomly sampling by using a Sobol sequence to generate data, and training the data by using a discrete physical information neural network PINNs; 3) screening data generated by training, and predicting a next state in a noise environment by using an extended Kalman filter; (4) according to the reference trajectory, a nonlinear model is used for predictively controlling the mechanical arm to conduct trajectory tracking, and an optimal control sequence is generated; and (5) according to the current state of the reference track, the optimal control input of the extended Kalman filter is updated, the step (4) and the step (5) are cyclically iterated, and real-time control over the mechanical arm is completed. The position and the speed of the multi-axis mechanical arm can be controlled in real time, smaller errors are obtained, and better control capacity is displayed.

Description

Technical Field

[0001] The present invention relates to the field of control and the application of artificial intelligence technology, and in particular to a multi-axis robotic arm predictive control method based on artificial intelligence information physical neural network. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing technology, multi-axis manipulators are playing an increasingly important role in modern industry. In these application scenarios, the trajectory control of the manipulator is crucial, directly affecting the accuracy, speed and stability of the operation. The traditional proportional-integral-derivative (PID) control method was widely used in industrial applications in the early days due to its simplicity and ease of implementation. However, with the increasing complexity of the operation tasks, the limitations of the PID control scheme in dealing with nonlinear dynamic characteristics, model uncertainties and external disturbances gradually emerged. In order to overcome these limitations, researchers have proposed advanced control methods such as adaptive control and robust control. Adaptive control can automatically adjust the control parameters according to the changes in the system state, thereby improving the adaptability and robustness of the system; robust control enhances the stability and reliability of the control system by considering the uncertainty of the system and external disturbances.

[0003] Although adaptive control and robust control have performed well in improving the adaptability and stability of the system, in order to cope with more complex nonlinear systems and to improve control efficiency, researchers have begun to explore intelligent control methods. However, these intelligent control methods often require a large amount of training data and have high computational costs when dealing with highly complex nonlinear dynamic systems. Therefore, researchers have further explored model predictive control (MPC) that can improve control performance through online optimization. It predicts the future behavior of the system and optimizes the control input by solving the optimization problem online. Due to its advantages in dealing with complex dynamic systems, it has been widely used in the control of complex dynamic systems. Nonlinear model predictive control (NMPC) is an extension of MPC that is specifically used to deal with nonlinear systems. It can handle problems with multivariable coupling and dynamic nonlinear characteristics, which is conducive to improving the accuracy and stability of the control system. Extended Kalman filter (EKF), as an algorithm for estimating the state of the system, is often used in NMPC to provide real-time state estimation, thereby improving control accuracy and system stability. Through prediction and optimization mechanisms, NMPC can achieve high-precision trajectory control and has the ability to handle multivariable coupling, constraints and time-varying characteristics. An important problem of NMPC is that it requires relatively accurate data. However, the models of many systems are relatively complex and have uncertainties. Using numerical analysis methods to process the system models without accurate experimental data often leads to poor control effects of NMPC, and even serious distortion and numerical overflow. Using neural networks to train and generate data is an important means to solve the problem of inaccurate NMPC data. Physically-informed neural networks (PINNs), as an advanced neural network model that combines physical laws and data-driven, have achieved remarkable results in scientific computing and engineering in recent years. Unlike other neural networks, PINNs embed physical laws such as dynamic equations and boundary conditions into the loss function of the neural network, so that the model not only relies on data during training, but is also guided by physical constraints, thereby improving the generalization ability and physical consistency of the model. In addition, due to the addition of physical information, PINNs can converge faster during training, reducing training time and computational costs. Summary of the invention

[0004] In view of the problems of insufficient real-time performance and high energy consumption in the existing technology, the present invention proposes a multi-axis robotic arm predictive control method based on information physical neural network. This method combines the advantages of discrete physical information neural network, extended Kalman filter and nonlinear model predictive control algorithm, and can ensure that the robotic arm can better track the reference trajectory.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] A multi-axis robot arm predictive control method based on cyber-physical neural network comprises the following steps:

[0007] 1) Establish a nonlinear dynamic robot model and configure it based on the physical relationship between the axes of the robot;

[0008] 2) Based on the robot model, the Sobol sequence random sampling is used to generate data, and the discrete physical information neural network PINNs is used to train the data;

[0009] 3) Filter the data generated by training and use the extended Kalman filter to predict the next state in a noisy environment;

[0010] 4) Based on the reference trajectory, the nonlinear model is used to predict and control the robot arm to track the trajectory and generate the optimal control sequence;

[0011] 5) According to the current state of the reference trajectory, the optimal control input of the extended Kalman filter is updated, and steps 4) and 5) are iterated cyclically to complete the real-time control of the robotic arm.

[0012] The step 1) comprises the following steps:

[0013] 1.1) Construct the dynamic equation of the robot arm:

[0014]

[0015] Where M(q) is the inertia matrix, is the Coriolis force matrix between the joints, g(q) is the gravity vector, B is the input matrix of the control system, u is the joint torque, Γ is the matrix of nonlinear friction influence, ω is the nonlinear disturbance vector, q is the joint position vector, is the joint velocity vector, is the joint acceleration vector;

[0016] 1.2) Define the order of each variable and matrix:

[0017] Define M(q) and are all (n×n)-order matrices, g(q), ω, q, as well as are all (n×1)-dimensional vectors, B is a (n×m)-order matrix, u is a (m×1)-dimensional vector, and Γ is a (n×n)-order matrix, where n is the number of joints or degrees of freedom of the robot arm, and m is the number of actuators or inputs of the system;

[0018] 1.3) Define the state variables and write them as state space expressions:

[0019] Defining state variables but q=[q1 ,q 2 ,…,q n ] T , and M(q) is a positive semidefinite matrix. Assuming that the system output is y, the state space expression is:

[0020]

[0021] Where I is the identity matrix.

[0022] The step 2) comprises the following steps:

[0023] 2.1) Use the Sobol sequence to randomly generate the initial position and speed data of each axis of the robot;

[0024] 2.2) Determine the calculation method of each matrix in the robotic arm model:

[0025] Assume that the connecting rod is a uniform rod, m i is the mass of the ith connecting rod, l i is the length of the ith connecting rod, q i is the angle of the i-th joint, q j is the angle of the jth joint, is the angular velocity of the ith joint, is the angular velocity of the jth joint, g is the gravitational acceleration constant, then:

[0026]

[0027] 2.3) Calculate the predicted value of joint position using the trapezoidal method and And update the original state: Assume the sampling period is Δt, the total number of steps is n, and the initial state is but:

[0028]

[0029] Update the original status:

[0030]

[0031] 2.4) Define the loss function L of PINNs:

[0032]

[0033] Among them, L data is the data loss, L phys is the physical loss, α, β are weight factors,

[0034]

[0035] Among them, Ndata is the number of data points, N phys is the number of physical points sampled;

[0036] 2.5) Select Adam optimizer for training.

[0037] The step 3) comprises the following steps:

[0038] 3.1) Construct the measurement equation:

[0039] Use the state equation to predict the state at the next moment: k =h(x k )+v k , where z k is the observed value, h is the measurement function, v k is the measurement noise;

[0040] 3.2) Use Extended Kalman Filter for prediction:

[0041] Use the measurement equation to predict the measurement value: Process noise ν k ~N(0,Q), where Q is the process noise covariance matrix and f is the system state update function; calculate the Jacobian matrix: Then the updated covariance matrix P k+1|k =FP k F T +Q, where P k is the state covariance matrix of the previous moment;

[0042] 3.3) Update the extended Kalman filter:

[0043] Use the measurement equation to predict the measured value: Measurement noise v k ~N(0, R), where R is the measurement noise covariance matrix; calculate its Jacobian matrix: The Kalman gain K = P k+1|k H T (HP k+1|k +R) -1 , update the state estimate according to the Kalman gain and the state covariance matrix P k+1 ,in P k+1 =(I-KH)P k+1|k .

[0044] The step 4) comprises the following steps:

[0045] 4.1) Set the trajectory equation of the reference trajectory

[0046] 4.2) Construct the cost function J:

[0047]

[0048] Among them, V is the state error weight matrix, W is the control input weight matrix, U is the terminal state error weight matrix, and L is the prediction time domain length;

[0049] 4.3) Define the constraints of the system:

[0050] The dynamic constraints of a system are defined as the dynamic equations of the system;

[0051] The input constraints of the system are: min ≤u k ≤u max ;

[0052] The state constraints of the system are: min ≤x k ≤x max ;

[0053] 4.4) Under the three constraints, find the solution that satisfies J min The optimal control input sequence u * .

[0054] The step 5) comprises the following steps:

[0055] 5.1) Find the solution that satisfies J min The optimal control input sequence u * = {u 0 ,u 1 …,u L-1}, take the first value u in the sequence * (0) as feedback variable;

[0056] 5.2) According to the current state of the reference trajectory, select a group of PINNs predictions that are closest to the current reference trajectory state And extract the corresponding Will with u * The first value of the sequence is compared, and the control input with the smaller difference from the current trajectory state is selected as the optimal control input of the extended Kalman filter at the next moment.

[0057] 5.3) Using the extended Kalman filter calculated at the previous moment As the state input of the extended Kalman filter at the next moment, steps 4) and 5) are iterated in a loop to complete the control of the robotic arm.

[0058] A multi-axis robotic arm predictive control system based on cyber-physical neural network, comprising:

[0059] A robot model building module, which is used to build and configure a nonlinear dynamic robot model based on the physical relationship between the axes of the robot;

[0060] Discrete physical information neural network module, which is used to generate data based on the robot model using Sobol sequence random sampling and train the data using discrete physical information neural network PINNs;

[0061] The extended Kalman filter module is used to filter the data generated by the training and use the extended Kalman filter to predict the next state in a noisy environment;

[0062] The nonlinear model predictive control module is used to track the trajectory of the robotic arm using the nonlinear model predictive control according to the reference trajectory and generate the optimal control sequence. According to the current state of the reference trajectory, the optimal control input of the extended Kalman filter is updated to complete the real-time control of the robotic arm.

[0063] A multi-axis robot arm predictive control device based on information physical neural network includes a memory and a processor; the memory is used to store computer programs; the processor is used to implement the multi-axis robot arm predictive control method based on information physical neural network when executing the computer program.

[0064] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-axis robot arm predictive control method based on an information physical neural network is implemented.

[0065] The present invention has the following beneficial effects and advantages:

[0066] 1. In order to solve the problems of long response time and poor trajectory tracking effect of the robot arm in the intelligent manufacturing scenario, a unified model is built for the multi-axis robot arm. This model considers the interaction force between the axes and takes the position and speed of each axis as the state variables, so that the method proposed in this invention has good practical significance.

[0067] 2. Aiming at the trajectory tracking problem of a multi-axis manipulator, the present invention proposes a predictive control method for a multi-axis manipulator based on an information physical neural network, which achieves precise control in the absence of prior data and in the presence of noise. The algorithm proposed in the present invention uses discrete PINNs to train the data obtained by random sampling without data. Discrete PINNs can maintain a faster training speed and better training results when the amount of data is large. After using the extended Kalman filter for state estimation, the nonlinear model predictive control algorithm is used to predict the state at the next moment, select the optimal control law and input feedback. Through analysis, the control method is effective and feasible. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is the principle diagram of the method of the present invention;

[0069] Figure 2 It is a model structure diagram of the discrete PINNs part in the present invention;

[0070] Figure 3 This is a model structure diagram of the extended Kalman filter part in the present invention;

[0071] Figure 4 This is a model structure diagram of the nonlinear model predictive control part in the present invention. DETAILED DESCRIPTION

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0073] The present invention is mainly divided into three parts: a discrete PINNs part, an extended Kalman filter part and a nonlinear model predictive control part.

[0074] Figure 1 Shown is a schematic diagram of the method of the present invention, and the specific implementation process is described as follows.

[0075] Step 1: Robot dynamics model and configuration;

[0076] The dynamic equation of the mechanical arm designed in the present invention is:

[0077]

[0078] M(q) is a (n×n)-order matrix; is a (n×n)-order matrix; g(q) is a (n×1)-dimensional vector; B is a (n×m)-order matrix; u is a (m×1)-dimensional vector; Γ is a (n×n)-order matrix; ω is a (n×1)-dimensional vector; q is a (n×1)-dimensional vector; is a (n×1)-dimensional vector; is a (n×1)-dimensional vector, where n is the number of joints or degrees of freedom of the robot and m is the number of actuators or inputs of the system.

[0079] Define the state variables and write them as state space expressions:

[0080] Defining state variables but q=[q 1 ,q 2 ,…,q n ] T , and M(q) is a positive semidefinite matrix. Assuming that the system output is y, the state space expression can be written as:

[0081]

[0082] Figure 2 The model structure diagram of the discrete PINNs part in the present invention is shown, and the specific implementation process is described as follows.

[0083] Step 2: Configuration of discrete PINNs;

[0084] Step 2.1: Use the Sobol sequence to randomly generate the initial position and velocity data of each axis of the robot arm.

[0085] Step 2.2: Determine how to calculate the matrices in the robot arm equation:

[0086] Assume that the connecting rod is a uniform rod, m i is the mass of the ith connecting rod, l i is the length of the ith connecting rod, q i is the angle of the i-th joint, q j is the angle of the jth joint, is the angular velocity of the ith joint, is the angular velocity of the jth joint, g is the gravitational acceleration constant (about 9.81 m / s 2 ).

[0087] Inertia Matrix Gravity vector Coriolis Force Matrix

[0088] Step 2.3: Discretize the dynamic equations. In order to effectively handle the dynamic changes of the system in actual robot arm control and ensure the stability of the control process, the present invention chooses to use the trapezoidal method. The trapezoidal method is a numerical integration method with second-order accuracy. Compared with the improved Euler method, it can provide a more accurate numerical solution. Calculate the predicted value using the trapezoidal method and And update the original state:

[0089] Assume the sampling period (step length) is Δt, the total number of steps is n, and the initial state is

[0090]

[0091] Update the original status:

[0092]

[0093] Step 2.4: Generate data using PINN training.

[0094] Design a multilayer perceptron (MLP) network with several hidden layers, several neurons in each layer, and define the loss function:

[0095] Data loss: Physical damage: Total loss function:

[0096]

[0097] Where α and β are weight factors used to balance the impact of data loss and physical loss.

[0098] Step 2.5: Select the Adam optimizer for training and print the loss value regularly to monitor the training process.

[0099] After initializing the neural network model, define the optimizer. The present invention selects the Adam optimizer for training and fine tuning. For each training cycle, the state variables need to be predicted first. Then calculate Determine the values ​​of all variables. Then, calculate L data and L phys , through the neural network back propagation, use the chain rule to calculate the gradient of the loss function relative to each network parameter and update the model parameters, according to the calculated gradient, use the optimizer to update the model parameters, regularly print the loss value to monitor the training process, output

[0100] Figure 3 Shown is a model structure diagram of the extended Kalman filter part in the present invention, and the specific implementation process is described as follows.

[0101] Step 3: Configuration of the extended Kalman filter.

[0102] Step 3.1: Introduce the measurement equation;

[0103] Use the state equation to predict the state at the next moment: k =h(x k )+v k . Where z kis the observed value, h is the measurement function, v k is the measurement noise, which is zero-mean Gaussian noise.

[0104] Step 3.2: Prediction process of extended Kalman filter:

[0105] Use the measurement equation to predict the measured value: Process noise ν k ~N(0,Q). Where ν k is the process noise in the prediction process, Q is the process noise covariance matrix, and f is the system state update function. Calculate the Jacobian matrix Then the covariance matrix P of the next step is k+1|k =FP k F T +Q, where P k is the state covariance matrix at the previous moment.

[0106] Step 3.3: Update process of extended Kalman filter:

[0107] Use the measurement equation to predict the measured value: Measurement noise v k ~N(0, R). Where R is the measurement noise covariance matrix. Calculate its Jacobian matrix We can get the Kalman gain K = P k+1|k H T (HP k+1|k +R) -1 . Update the state estimate according to the Kalman gain and the state covariance matrix P k+1 ,in P k+1 =(I-KH)P k+1|k .

[0108] Figure 4 Shown is the model structure diagram of the nonlinear model predictive control part in the present invention, and the specific implementation process is described as follows.

[0109] Step 4: Nonlinear model predictive controller design.

[0110] Step 4.1: Set the trajectory equation for the reference trajectory

[0111] Step 4.2: Define the cost function:

[0112]

[0113] Among them, V is the state error weight matrix, W is the control input weight matrix, U is the terminal state error weight matrix, and L is the prediction time domain length.

[0114] Step 4.3: System constraints definition.

[0115] The dynamic constraints of the system are defined as the dynamic equations of the system. The input constraints of the system are: min ≤u k ≤u max The state constraints of the system are: min ≤x k ≤x max .

[0116] Step 4.4: Determine control objectives:

[0117] The problem is transformed into finding the solution J under three constraints: min The optimal control input sequence u * .

[0118] Step 5: Find the optimal control law and repeat the cycle.

[0119] Step 5.1: Find the condition that satisfies J min The optimal control input sequence u * = {u 0 ,u 1 …,u L-1}, take the first value u in the sequence * (0) as the feedback variable.

[0120] Step 5.2: Based on the current state of the reference trajectory, select a group of PINNs predictions that are closest to the current reference trajectory state And extract the corresponding Combine it with u * The first value of the sequence is compared, and the control input with the smaller difference from the current trajectory state is selected as the optimal control input of the EKF at the next moment.

[0121] Step 5.3: Use the EKF calculation at the previous moment As the state input of the EKF at the next moment, the optimal control law at each moment can be obtained by iterating steps 4) and 5) in a loop.

[0122] It should be understood by those skilled in the art that the embodiments of the present application can be provided as methods or robotic arm products. Therefore, the present application can adopt the form of embodiments combining software and hardware aspects. Moreover, the present application can deploy the method on a multi-axis robotic arm that supports programming in the form of Python code.

[0123] The present application is described with reference to the method flow chart according to the embodiment of the present application. It should be understood that each process in the flow chart can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a computer, controller, embedded processor or other programmable device to configure the algorithm so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flow chart. Figure 1 A device that specifies functions in a process or multiple processes.

[0124] After these computer program instructions are deployed on other devices, the program can be downloaded to the multi-axis robot arm through the communication protocol supported by the robot arm, so that it can execute the algorithm proposed by the present invention after power-on, thereby providing the instructions executed on the computer or other programmable device for the multi-axis robot arm to realize the process Figure 1 The steps of a specified function in a process or multiple processes.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A predictive control method for multi-axis robotic arms based on cyber-physical neural networks, It is characterized in that The following steps are involved: 1) Establish a nonlinear dynamic robot model and configure it based on the physical relationship between the axes of the robot; 2) Based on the robot model, the Sobol sequence random sampling is used to generate data, and the discrete physical information neural network PINNs is used to train the data; 3) Filter the data generated by training and use the extended Kalman filter to predict the next state in a noisy environment; 4) Based on the reference trajectory, the nonlinear model is used to predict and control the robot arm to track the trajectory and generate the optimal control sequence; 5) According to the current state of the reference trajectory, the optimal control input of the extended Kalman filter is updated, and steps 4) and 5) are iterated cyclically to complete the real-time control of the robotic arm.

2. According to claim 1, a multi-axis robot arm predictive control method based on cyber-physical neural network, It is characterized in that The step 1) comprises the following steps: 1.1) Construct the dynamic equation of the robot arm: Where M(q) is the inertia matrix, is the Coriolis force matrix between the joints, g(q) is the gravity vector, B is the input matrix of the control system, u is the joint torque, Γ is the matrix of nonlinear friction influence, ω is the nonlinear disturbance vector, q is the joint position vector, is the joint velocity vector, is the joint acceleration vector; 1.2) Define the order of each variable and matrix: Define M(q) and are all (n×n)-order matrices, g(q), ω, q, as well as are all (n×1)-dimensional vectors, B is a (n×m)-order matrix, u is a (m×1)-dimensional vector, and Γ is a (n×n)-order matrix, where n is the number of joints or degrees of freedom of the robot arm, and m is the number of actuators or inputs of the system; 1.3) Define the state variables and write them as state space expressions: Defining state variables but q=[q 1 ,q 2 ,…,q n ] T , and M(q) is a positive semidefinite matrix. Assuming that the system output is y, the state space expression is: Where I is the identity matrix.

3. According to claim 1, a multi-axis robot arm predictive control method based on cyber-physical neural network, It is characterized in that The step 2) comprises the following steps: 2.1) Use the Sobol sequence to randomly generate the initial position and speed data of each axis of the robot; 2.2) Determine the calculation method of each matrix in the robotic arm model: Assume that the connecting rod is a uniform rod, m i is the mass of the ith connecting rod, l i is the length of the ith connecting rod, q i is the angle of the i-th joint, q j is the angle of the jth joint, is the angular velocity of the ith joint, is the angular velocity of the jth joint, g is the gravitational acceleration constant, then: 2.3) Calculate the predicted value of joint position using the trapezoidal method and And update the original state: Assume the sampling period is Δt, the total number of steps is n, and the initial state is but: Update the original status: 2.4) Define the loss function L of PINNs: Among them, L data is the data loss, L phys is the physical loss, α, β are weight factors, Among them, N data is the number of data points, N phys is the number of physical points sampled; 2.5) Select Adam optimizer for training.

4. The predictive control method for a multi-axis robot arm based on a cyber-physical neural network according to claim 1, It is characterized in that The step 3) comprises the following steps: 3.1) Construct the measurement equation: Use the state equation to predict the state at the next moment: k =h(x k )+υ k , where z k is the observed value, h is the measurement function, υ k is the measurement noise; 3.2) Use Extended Kalman Filter for prediction: Use the measurement equation to predict the measurement value: Process noise v k ~N(0,Q), where Q is the process noise covariance matrix and f is the system state update function; calculate the Jacobian matrix: Then the updated covariance matrix P k+1|k =FP k F T +Q, where P k is the state covariance matrix of the previous moment; 3.3) Update the extended Kalman filter: Use the measurement equation to predict the measurement value: Measurement noise k ~N(0, R), where R is the measurement noise covariance matrix; calculate its Jacobian matrix: Get Kalman gain K = P k+1| k H T (HP k+1|k +R) -1 , update the state estimate according to the Kalman gain and the state covariance matrix P k+1 ,in P k+1 =(I-KH)P k+1|k .

5. The predictive control method for a multi-axis robot arm based on a cyber-physical neural network according to claim 1, It is characterized in that The step 4) comprises the following steps: 4.1) Set the trajectory equation of the reference trajectory 4.2) Construct the cost function J: Among them, V is the state error weight matrix, W is the control input weight matrix, U is the terminal state error weight matrix, and L is the prediction time domain length; 4.3) Define the constraints of the system: The dynamic constraints of a system are defined as the dynamic equations of the system; The input constraints of the system are: min ≤u k ≤u max ; The state constraints of the system are: min ≤x k ≤x max ; 4.4) Under the three constraints, find the solution that satisfies J min The optimal control input sequence u * .

6. The predictive control method for a multi-axis robot arm based on cyber-physical neural network according to claim 1, It is characterized in that The step 5) comprises the following steps: 5.1) Find the solution that satisfies J min The optimal control input sequence u * = {u 0 ,u 1 …,u L-1 }, take the first value u in the sequence * (0) as feedback variable; 5.2) According to the current state of the reference trajectory, select a group of PINNs predictions that are closest to the current reference trajectory state And extract the corresponding Will Compare with the first value of the u* sequence and select the control input with the smaller difference from the current trajectory state as the optimal control input of the extended Kalman filter at the next moment 5.3) Using the extended Kalman filter calculated at the previous moment As the state input of the extended Kalman filter at the next moment, steps 4) and 5) are iterated in a loop to complete the control of the robotic arm.

7. A predictive control system for a multi-axis robotic arm based on cyber-physical neural network, It is characterized in that include: A robot model building module, which is used to build and configure a nonlinear dynamic robot model based on the physical relationship between the axes of the robot; Discrete physical information neural network module, which is used to generate data based on the robot model using Sobol sequence random sampling and train the data using discrete physical information neural network PINNs; The extended Kalman filter module is used to filter the data generated by the training and use the extended Kalman filter to predict the next state in a noisy environment; The nonlinear model predictive control module is used to track the trajectory of the robotic arm using the nonlinear model predictive control according to the reference trajectory and generate the optimal control sequence. According to the current state of the reference trajectory, the optimal control input of the extended Kalman filter is updated to complete the real-time control of the robotic arm.

8. A multi-axis robotic arm predictive control device based on cyber-physical neural network, It is characterized in that It comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement a multi-axis robot arm predictive control method based on an information physical neural network as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, It is characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the multi-axis robot arm predictive control method based on information physical neural network as described in any one of claims 1 to 6 is implemented.

Images