Predictive control method for robot flexible joints based on dual neural network model

Abstract

The present invention belongs to the field of intelligent control technology and discloses a predictive control method for robot flexible joints based on a dual neural network model, comprising collecting dynamic parameters of each flexible joint in the robot under various working conditions and constructing a motion state database; constructing a real-time prediction model based on the motion state database; collecting real-time state parameters and inputting them into the real-time prediction model to predict real-time characteristic parameters; constructing a future prediction model based on the motion state database; inputting the real-time characteristic parameters into the future prediction model to predict future motion parameters; dynamically optimizing the future motion parameters to obtain optimized motion parameters and controlling the flexible joints; the present invention can give full play to the advantages of deep learning in nonlinear dynamic system modeling and optimization control, effectively improve the control accuracy, response speed and adaptability of flexible joints, and enhance the real-time performance and prediction accuracy of the control system.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control, and in particular relates to a robot flexible joint predictive control method based on a dual neural network model. Background Art

[0002] With the rapid development of robotics technology, flexible joint robots have been widely used in industrial automation, medical rehabilitation, bionic robots and other fields; compared with traditional rigid joint robots, flexible joint robots have better compliance, lightweight and safety, and can achieve efficient and precise motion control in complex environments; however, the dynamic characteristics of flexible joints are complex, with problems such as nonlinearity, time-varying, strong coupling and elastic deformation, which bring great challenges to the control system; traditional robot joint control methods, such as PID control, adaptive control and robust control, although they can improve the stability and accuracy of the system to a certain extent, are often difficult to achieve ideal control effects when faced with the complex dynamic characteristics of flexible joints; especially in high dynamic, high precision and high response speed application scenarios, traditional control methods have problems such as low prediction accuracy, limited control performance, and difficulty in adapting to nonlinear characteristics; therefore, the study of efficient and intelligent flexible joint predictive control methods has become an important topic in the field of robotic control.

[0003] In recent years, the rise of artificial intelligence and deep learning technologies has provided new solutions for robot control. For example, patent publication number CN119396008A discloses a dual neural network model predictive control method for flexible joints of robots. The method includes: modeling the nonlinear dynamic system of the flexible joint robot to construct a dynamic model; setting a GRU neural network model to approximate the system's dynamic model; constructing a hybrid control NMNPC controller containing position, velocity, and torque based on the GRU neural network model; optimizing the NMNPC controller based on a nonlinear optimization solver; and setting a NARX neural network model to approximate the solution of the NMNPC controller, forming a dual neural network model predictive control method. This method can reduce the computational overhead incurred by traditional methods in high-dimensional nonlinear systems, thereby achieving real-time control.

[0004] However, although the above technology can realize predictive control of robot flexible joints, its control strategy relies on NMNPC optimization solution, which has a large amount of calculation and leads to a long solution time in high-dimensional nonlinear control, affecting real-time performance. At the same time, the above technology approximates the system dynamics through the GRU neural network, but still needs to use the NARX neural network to learn the optimal solution of the NMNPC controller, which increases the complexity of model training and optimization. In addition, the combination of GRU and NARX neural networks is relatively fixed, which makes it difficult to quickly adapt to new working conditions, reducing the generalization ability and adaptability of the model.

[0005] In view of this, the present invention proposes a robot flexible joint predictive control method based on a dual neural network model to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a predictive control method for robot flexible joints based on a dual neural network model, which can give full play to the advantages of deep learning in nonlinear dynamic system modeling and optimization control, effectively improve the control accuracy, response speed and adaptability of flexible joints, and enhance the real-time performance and prediction accuracy of the control system.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A robot flexible joint predictive control method based on a dual neural network model, the robot flexible joint predictive control method based on a dual neural network model comprising:

[0009] Collect the dynamic parameters of each flexible joint of the robot under various working conditions and build a motion state database;

[0010] Based on the motion state database, the dynamic parameters are divided into observation parameters and response parameters, and a feedforward neural network structure is used to build a real-time prediction model;

[0011] The observation parameters of each flexible joint are collected in real time, and the joint labels are set respectively and then input into the real-time prediction model to obtain the predicted response parameters as real-time characteristic parameters;

[0012] Based on the motion state database, the dynamic parameters of a flexible joint under the same working condition are taken as a set of parameters, and a recurrent neural network structure is used to build a future prediction model;

[0013] Input the real-time characteristic parameters into the future prediction model to predict the future dynamic parameters of each flexible joint;

[0014] An intelligent decision-making algorithm is used to dynamically optimize the future dynamic parameters of each flexible joint to obtain optimized motion parameters, and each flexible joint is controlled based on the optimized motion parameters.

[0015] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution. They are merely further supplements or optimizations. Under the premise that there are no technical or logical contradictions, each optional method can be combined separately for the above-mentioned overall solution, or multiple optional methods can be combined.

[0016] Preferably, the method is based on a motion state database, divides the dynamic parameters into observation parameters and response parameters, and adopts a feedforward neural network structure to construct a real-time prediction model, including:

[0017] Each kinetic parameter in the motion state database is divided into observation parameters and response parameters;

[0018] Obtain multiple sets of different observation parameters and corresponding response parameters from the motion state database, and set corresponding joint labels for each set of observation parameters according to the joint type corresponding to each set of observation parameters, where the joint label is the label corresponding to the joint type;

[0019] The observation parameters after setting the joint labels are used as the input of the feedforward neural network structure, and the corresponding response parameters are used as the prediction targets of the feedforward neural network structure. The feedforward neural network structure is trained and updated, and after the training and updating is completed, the feedforward neural network structure is output as a real-time prediction model.

[0020] Preferably, the method is based on a motion state database, takes the dynamic parameters of a flexible joint under the same working condition as a set of parameters, and adopts a recurrent neural network structure to construct a future prediction model, including:

[0021] Obtain the dynamic parameters corresponding to a flexible joint in the motion state database and mark them as training parameters. The training parameters corresponding to the same working condition are used as a set of parameter sets. The training parameters in each parameter set are sorted from early to late according to the corresponding collection time to generate a training set corresponding to each parameter set.

[0022] Set the sliding step size and sliding window length , the training parameters in each training set are converted into multiple training samples using the sliding window method. A training sample contains Set training parameters, use the training samples as the input of the recurrent neural network structure, and The training parameters are used as prediction targets of the recurrent neural network structure, the recurrent neural network structure is trained and updated, and after the training and updating, the recurrent neural network structure is output as a motion prediction model of the corresponding flexible joint;

[0023] The motion prediction models of all flexible joints are used as future prediction models.

[0024] Preferably, the inputting of the real-time characteristic parameters into the future prediction model to predict the future dynamic parameters of each flexible joint includes:

[0025] The observation parameters of a flexible joint collected in real time and the corresponding response parameters in the real-time characteristic parameters are used as real-time parameters;

[0026] From the motion state database, obtain the corresponding real-time parameters Group of continuous dynamic parameters; among them, real-time parameters and corresponding The acquisition time of the continuous dynamic parameters of the group is continuous, and the real-time parameters are consistent with the corresponding The continuous dynamic parameters of each group correspond to the same flexible joint;

[0027] The real-time parameters are compared with the corresponding A set of continuous dynamic parameters are input together into the motion prediction model of the corresponding flexible joint to predict the future dynamic parameters of the current flexible joint.

[0028] Preferably, the intelligent decision-making algorithm is used to dynamically optimize the future dynamic parameters of each flexible joint to obtain the optimized motion parameters, including:

[0029] Obtain the earliest dynamic parameters of each flexible joint corresponding to the future moment and mark them as parameters to be optimized;

[0030] According to the parameters to be optimized, each flexible joint is constructed Group candidate set, is an integer greater than 1;

[0031] From each flexible joint The best set corresponding to each flexible joint is selected from the candidate set; the best set corresponding to all flexible joints is used as the optimized motion parameter.

[0032] As a preference, the method of constructing each flexible joint according to the parameters to be optimized is as follows: Group candidate set, including:

[0033] A preset perturbation set includes the perturbation value corresponding to each parameter in the dynamic parameters;

[0034] Obtain the dynamic parameters corresponding to a flexible joint among the parameters to be optimized and mark them as disturbance parameters;

[0035] Add the disturbance value of the corresponding parameter in the disturbance set to each parameter in the disturbance parameter to obtain the maximum disturbance value;

[0036] Subtract the disturbance value of the corresponding parameter in the disturbance set from each parameter in the disturbance parameter to obtain the minimum disturbance value;

[0037] According to the maximum and minimum disturbance values corresponding to each parameter, the disturbance range value corresponding to each parameter in the disturbance parameter is constructed;

[0038] A value is randomly selected from each perturbation range value, and a set of candidate sets is constructed. Group candidate set, The candidate sets of groups are all different.

[0039] As a preference, the The best set corresponding to each flexible joint is selected from the candidate set, including:

[0040] Set an increasing numerical label for each candidate set in turn and mark it as a set label. The range of the set label is , and according to Determine the initial screening center and initial screening radius;

[0041] Define the iterative process, which is: generate within the range of the set label candidate solutions, each candidate solution corresponds to a set label one by one. ; Calculate the control effect corresponding to each candidate solution in turn and generate the candidate solution boundary; Based on the candidate solution boundary, calculate the boundary solution corresponding to each candidate solution, and calculate the control effect corresponding to each boundary solution in turn; Candidate solutions and boundary solutions are collectively referred to as calculated solutions, compare the control effects of all calculated solutions, move the screening center to the calculated solution with the highest control effect, and adjust the screening radius;

[0042] The iterative process is executed. When the number of executions of the iterative process is greater than or equal to the preset iteration threshold, the iterative process ends, the calculated solution corresponding to the screening center is marked as the best solution, and the candidate set corresponding to the set label corresponding to the best solution is obtained and marked as the best set.

[0043] As an advantage, the method for generating the candidate solution is: Generate a random coefficient, multiply the screening radius by the random coefficient and add the screening center as a candidate solution;

[0044] The method for adjusting the screening radius is as follows: A random adjustment coefficient is generated in , the screening radius is multiplied by the adjustment coefficient to obtain the adjustment radius, and the screening radius is adjusted to the adjustment radius.

[0045] Preferably, the control effect calculation process is as follows:

[0046] Based on the motion state database, a deep neural network structure is used to build an effect calculation model;

[0047] The observation parameters of a flexible joint collected in real time and the corresponding response parameters in the real-time characteristic parameters are used as real-time parameters, the real-time parameters and the calculation solutions belonging to the same flexible joint are input into the effect calculation model, and the effect calculation model outputs the control effect.

[0048] Preferably, the method for generating candidate solution boundaries includes:

[0049] Obtain the set label corresponding to each candidate solution and mark it as an edge label; compare each edge label separately, mark the edge label with the largest value as the maximum label, and mark the edge label with the smallest value as the minimum label; generate the candidate solution boundary based on the maximum label and the minimum label, that is, take the maximum label as the maximum value of the candidate solution boundary, and take the minimum label as the minimum value of the candidate solution boundary;

[0050] The step of calculating the boundary solution corresponding to each candidate solution according to the candidate solution boundary includes:

[0051] From the interval Randomly generated numerical values, and use them as the calculation coefficients corresponding to each candidate solution in turn; add the maximum label to the minimum label to obtain the label coefficient; multiply the label coefficient by each calculation coefficient to obtain the boundary coefficient corresponding to each candidate solution; subtract the corresponding candidate solution from each boundary coefficient to obtain the boundary solution corresponding to each candidate solution.

[0052] The present invention provides a robot flexible joint predictive control method based on a dual neural network model, which adopts two neural network structures, a real-time prediction model and a future prediction model, to operate in parallel. The former can capture the nonlinear coupling relationship of the flexible joint in static and transient states, and the latter can capture the timing dependence and dynamic evolution law of the flexible joint during the movement process, thereby realizing comprehensive modeling and accurate prediction of the flexible joint motion characteristics; at the same time, an intelligent decision-making algorithm is adopted to dynamically optimize the future motion parameters, and the control parameters are updated in real time through a rolling time domain strategy, so that the control system can quickly respond to the real-time state changes of the flexible joint, effectively improving the control accuracy, response speed and adaptability of the flexible joint; giving full play to the advantages of deep learning in nonlinear dynamic system modeling and optimization control, not only improving the adaptability of the robot in a dynamic environment, but also reducing the computational burden of traditional control methods, and improving the real-time performance and prediction accuracy of the control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a robot flexible joint predictive control method based on a dual neural network model of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0056] Example 1: Figure 1 As shown, this embodiment provides a robot flexible joint predictive control method based on a dual neural network model, comprising the following steps:

[0057] Step S1: Collect the dynamic parameters of each flexible joint in the robot under various working conditions and build a motion state database, which helps to establish a comprehensive and accurate basic data support covering multiple working states and provides a reliable basis for subsequent modeling and prediction.

[0058] A robot's flexible joints include revolute and translatory joints. The dynamic parameters of flexible joints under various operating conditions refer to the dynamic parameters exhibited by the robot's flexible joints during different operations (such as assembly, welding, and handling). These dynamic parameters include rotational and translatory dynamic parameters. Rotational dynamic parameters correspond to the revolute joints within the flexible joints (such as the robot's elbow and shoulder joints), while translatory dynamic parameters correspond to the translatory joints within the flexible joints (such as the robot's leg extension joints and the gripper's opening and closing joints). Rotational dynamic parameters include joint angle, angular velocity, angular acceleration, and torque. Torque includes driving torque, elastic torque, and gravitational torque. Translocation dynamic parameters include linear displacement, linear velocity, linear acceleration, and applied force. Forces include driving force, elastic force, and gravity. Dynamic parameters are obtained through actual measurements during the flexible joint's historical operation.

[0059] The joint angle is the position representation of the rotating joint, which is used to describe the amount of rotation of the rotating joint around its rotation axis; the angular velocity is the rate of change of the rotating joint position, which represents the angle of rotation of the rotating joint per unit time; the angular acceleration is the rate of change of the angular velocity of the rotating joint, which represents the change in angular velocity per unit time; the driving torque is the torque applied by the control system on the rotating joint, the elastic torque is the torque generated by the elastic deformation inside the rotating joint, and the gravity torque is the torque generated by the deadweight of the rotating joint; the linear displacement is the position representation of the moving joint, which is used to represent the translation distance of the moving joint in a certain direction; the linear velocity is the rate of change of the moving joint position, which represents the distance the moving joint translates per unit time; the linear acceleration is the rate of change of the linear velocity of the moving joint, which represents the change in linear velocity per unit time; the driving force is the force applied by the control system on the moving joint, the elastic force is the force generated by the elastic deformation inside the moving joint, and the gravity is the force generated by the deadweight of the moving joint.

[0060] Step S2: Based on the motion state database, a feedforward neural network structure is used to construct a real-time prediction model to achieve rapid modeling of complex dynamic characteristics, which can reflect the system behavior of different flexible joints in different motion states in real time and efficiently.

[0061] The feedforward neural network structure is an artificial neural network structure, which is characterized by information starting from the input layer, passing through one or more hidden layers in sequence, and finally reaching the output layer. The neurons in each layer are only connected to the previous and next layers. Information flows in one direction and there is no loop or feedback connection.

[0062] The method for building a real-time prediction model provided in this embodiment includes:

[0063] In the rotational dynamics parameters, the joint angle and angular velocity are used as rotational observation parameters, and the angular acceleration and torque are used as rotational response parameters; in the movement dynamics parameters, the linear displacement and linear velocity are used as movement observation parameters, and the linear acceleration and force are used as movement response parameters; the rotation observation parameters and the movement observation parameters are respectively referred to as observation parameters, and the rotation response parameters and the movement response parameters are respectively referred to as response parameters, that is, the observation parameters can be either rotation observation parameters or movement observation parameters, and the response parameters can be either rotation response parameters or movement response parameters.

[0064] Get from the motion state database A set of different observation parameters and corresponding response parameters, The group observation parameters include Group rotation observation parameters and A group of moving observation parameters, that is, a group of observation parameters includes a rotation observation parameter or a moving observation parameter obtained at a sampling moment, and , and Each set of observation parameters is assigned a corresponding joint label based on the joint type corresponding to each set of observation parameters. Joint types include rotational joints and translational joints, and joint labels are numerical labels corresponding to the joint type. Different joint types have different joint labels. Each set of observation parameters and the corresponding joint labels are considered a set of joint observation parameters, and the joint observation parameters correspond one-to-one with the observation parameters.

[0065] A set of joint observation parameters and corresponding response parameters are converted into a corresponding set of feature vectors, and each set of feature vectors is used as the input of the real-time prediction model. The real-time prediction model uses a set of predicted response parameters corresponding to each set of joint observation parameters as output, and the actual response parameters corresponding to each set of joint observation parameters as the prediction target. The actual response parameters are the response parameters corresponding to the joint observation parameters in the motion state database. Different sets of observation parameters are divided into training and test sets, both of which include rotation and motion observation parameters. The real-time prediction model is trained based on the training set. After each training round, the model is tested based on the test set, and the prediction error is calculated. Based on the prediction error, the error change value is calculated. The error change value is compared with a preset change threshold. If the error change value is greater than or equal to the change threshold, training continues. If the error change value is less than the change threshold, training stops, completing the construction of the real-time prediction model, which is a deep neural network model. The change threshold is preset based on actual conditions.

[0066] The calculation method of the prediction error is as follows: subtract the corresponding data in the corresponding predicted response parameter from each data in the actual response parameter corresponding to each group of joint observation parameters in the test set, and square each difference to obtain the square difference corresponding to each group of joint observation parameters; preset a weight set, the weight set includes the weight coefficient corresponding to each parameter in the response parameter, and the weight set is pre-set according to the actual situation; multiply the square difference corresponding to each group of joint observation parameters by the corresponding weight coefficient in the weight set to obtain the difference weight; add the difference weights corresponding to each group of joint observation parameters in turn to obtain the error value corresponding to each group of joint observation parameters; add the error values corresponding to all joint observation parameters in the test set in turn to obtain the prediction error.

[0067] The error change value is calculated by subtracting the prediction error calculated after the previous training round from the prediction error calculated after the current training round to obtain the error change value.

[0068] Step S3: collect real-time state parameters and input them into the real-time prediction model. The real-time prediction model captures the nonlinear coupling relationship of the flexible joint under static and transient conditions and predicts the real-time characteristic parameters, which helps to capture the nonlinear and dynamic characteristics of the flexible joint in real time, enhance the sensitivity and responsiveness to changes in actual working conditions, and improve control accuracy.

[0069] The real-time state parameters include the observed parameters of each flexible joint on the robot during real-time operation, acquired through the inertial measurement unit integrated into each flexible joint. The joint label corresponding to each flexible joint is added to the corresponding observed parameters within the real-time state parameters to obtain the real-time observed parameters for each flexible joint. Each set of real-time observed parameters is input into the real-time prediction model to predict the response parameters for each flexible joint. The response parameters for all flexible joints are then used as real-time characteristic parameters.

[0070] It should be noted that the reason for predicting real-time characteristic parameters is that torque sensors are large in size and difficult to integrate into all flexible joints. In addition, torque sensors usually have bandwidth limitations and cannot accurately capture transient torque during high-speed dynamic changes. At the same time, acceleration measurements are easily affected by noise interference, especially in high-frequency vibration environments, and acceleration sensors also have measurement range and accuracy limitations. Therefore, it is necessary to predict real-time characteristic parameters based on real-time observation parameters to obtain accurate real-time characteristic parameters.

[0071] It should be understood that deep neural networks have powerful nonlinear function approximation capabilities and multi-layer feature extraction capabilities. They can learn complex mapping relationships between variables from multi-dimensional perception data, and thus accurately capture the complex nonlinear coupling relationships of flexible joints in static and transient states.

[0072] Step S4: Based on the motion state database, a recurrent neural network structure is used to construct a future prediction model. By introducing temporal memory capabilities, the historical dependencies and dynamic evolution laws of joint motion are mined to provide accurate predictions for future motion trends.

[0073] The recurrent neural network structure is a neural network structure specifically used to process sequence data. It is characterized by its "memory ability" and can transmit information in the time dimension through hidden states, realizing the joint modeling of current input and historical information, thereby capturing dynamic changes in time series.

[0074] Future prediction models include A motion prediction model, is the number of flexible joints on the robot, that is, each flexible joint corresponds to a motion prediction model, The motion prediction models are all RNN neural network models, and The construction methods of the motion prediction models are consistent.

[0075] Specifically, the method for constructing a motion prediction model includes:

[0076] The dynamic parameters corresponding to a flexible joint in the motion state database are obtained and marked as training parameters. Specifically, a group of dynamic parameters is marked as a set of training parameters. All training parameters corresponding to the same operation process are grouped as a parameter set, with each parameter set corresponding to the operation process. The training parameters in each parameter set are sorted from earliest to latest according to their corresponding collection time, generating a training set corresponding to each parameter set. Based on each training set, a motion prediction model is trained to predict the training parameters at future moments.

[0077] Preset sliding step size based on actual experience And the sliding window length , the training parameters in each training set are converted into multiple training samples using the sliding window method, and each training sample includes continuous Set training parameters. Take each training sample as the input of the motion prediction model to predict the sliding step length The training parameters after each training sample are taken as output. The training parameters of a set are used as prediction targets, and the model accuracy is evaluated using the mean absolute percentage error (MAPE) of the prediction results. When the calculated MAPE is less than the preset error threshold, the motion prediction model training is complete, generating a motion prediction model that predicts future training parameters based on multiple sets of continuous training parameters. The error threshold is pre-set based on the prediction accuracy requirements of the motion prediction model.

[0078] The calculation formula for the mean absolute percentage error is:

[0079] ;

[0080] Where, is the mean absolute percentage error, For the The first prediction target corresponding to the training parameters predicted by the group parameters, For the The first of the training parameters predicted by the group parameters, The total number of predicted training parameter groups is related to the number of predicted targets and the number of training parameter groups in each predicted target. is the number of parameters in the training parameters, is the training parameter The proportional coefficient corresponding to each parameter in the training parameters is preset according to the actual situation.

[0081] For example, the training set Contains 10 sets of training parameters, , define the sliding window length as 3, the sliding step size as 1, use the sliding window to construct 7 training samples, each training sample contains 3 consecutive sets of training parameters, and the next set of training parameters of the 3 consecutive sets of training parameters is used as the prediction target; if the training sample , then the training samples The corresponding prediction target is ; If the training sample , then the training samples The corresponding prediction target is ; And so on, it is used to train the motion prediction model corresponding to the training parameters.

[0082] Step S5: input the real-time characteristic parameters into the future prediction model. The future prediction model captures the timing dependency and dynamic evolution law of the flexible joint during the movement process, and predicts the future movement parameters. This helps to combine the current state of the flexible joint with the time evolution law, predict the future state of the flexible joint in advance, and improve the foresight and initiative of the control system.

[0083] Each set of observation parameters in the real-time state parameters and the corresponding response parameters in the real-time characteristic parameters are respectively regarded as a set of real-time parameters, and the real-time parameters correspond to the observation parameters in the real-time state parameters one by one; the corresponding parameters of each set of real-time parameters are obtained from the motion state database. Group of continuous dynamic parameters; among them, real-time parameters and corresponding The acquisition time of each set of continuous dynamic parameters is continuous, and the real-time parameters and the corresponding continuous dynamic parameters correspond to the same flexible joint; each set of real-time parameters is compared with the corresponding A set of continuous dynamic parameters is used as a set of joint parameters; each set of joint parameters is input into the corresponding motion prediction model to predict the future Dynamic parameters at the moment; all future corresponding flexible joints The dynamic parameters at each moment are used as the future motion parameters.

[0084] It should be noted that in this embodiment, the motion state database is updated in real time. That is, after real-time acquisition of dynamic parameters, the newly acquired dynamic parameters are added to the motion state database so that the real-time parameters can be combined to predict dynamic parameters at future moments. Furthermore, the real-time prediction model and the future prediction model can be updated periodically.

[0085] It should be understood that the RNN neural network introduces a recurrent connection of hidden states in the time dimension, so that the output at each moment not only depends on the current input, but also inherits the state information of the previous moment. It can effectively "memorize" and "process" historical state information within the model, thereby capturing the time dependency and dynamic evolution law of flexible joints during movement.

[0086] Step S6: Use an intelligent decision-making algorithm to dynamically optimize future motion parameters, obtain optimized motion parameters, and control the flexible joints based on the optimized motion parameters through a rolling time domain optimization strategy to achieve closed-loop optimization and rolling updates, thereby enhancing the robustness, stability and adaptability of the control system.

[0087] The method for obtaining optimized motion parameters includes:

[0088] (1) Obtain the earliest dynamic parameters of each flexible joint in the future motion parameters corresponding to the future moment and mark them as parameters to be optimized.

[0089] (2) According to the parameters to be optimized, each flexible joint is constructed Group candidate set, is an integer greater than 1.

[0090] A perturbation set is preset, and the perturbation set includes the perturbation value corresponding to each parameter in the dynamic parameters. The perturbation set is preset according to the actual situation. The dynamic parameters corresponding to a flexible joint in the parameters to be optimized are obtained and marked as perturbation parameters. Each parameter in the perturbation parameter is added with the corresponding perturbation value in the perturbation set to obtain the maximum perturbation value. Each parameter in the perturbation parameter is subtracted from the corresponding perturbation value in the perturbation set to obtain the minimum perturbation value. According to the maximum and minimum perturbation values corresponding to each parameter, the perturbation range value corresponding to each parameter in the perturbation parameter is constructed. A value is randomly selected from each perturbation range value, and a set of candidate sets is constructed to construct a total of Group candidate set, The candidate sets of groups are all different.

[0091] (3) From each flexible joint corresponding The best set corresponding to each flexible joint is screened out from the candidate set; the best set corresponding to all flexible joints is used as the optimized motion parameter.

[0092] Set increasing numerical labels for each candidate set in turn, and mark each numerical label as a set label. The range of the set label is ;Will Add 1 and divide by 2 as the initial screening center. Subtract 1 and divide by 2 to obtain the initial screening radius. In other embodiments, the settings of the initial screening center and the initial screening radius may have other variations.

[0093] Define the iterative process, which is: generate within the range of the set label candidate solutions, each candidate solution corresponds to a set label one by one. ; Calculate the control effect corresponding to each candidate solution in turn, and generate the candidate solution boundary; According to the candidate solution boundary, calculate the boundary solution corresponding to each candidate solution, and calculate the control effect corresponding to each boundary solution in turn; The candidate solution and the boundary solution are collectively referred to as the calculated solution, and the control effects of all the calculated solutions are compared. The screening center is moved to the calculated solution with the highest control effect, and the screening radius is adjusted.

[0094] The iterative process is executed. When the number of executions of the iterative process is greater than or equal to the preset iteration threshold, the iterative process ends, the calculated solution corresponding to the screening center is marked as the best solution, and the candidate set corresponding to the set label corresponding to the best solution is obtained and marked as the best set; the iteration threshold is pre-set according to the actual situation.

[0095] The method for generating candidate solutions is: Generate a random coefficient in the interval, multiply the screening radius by the random coefficient and add the screening center to form a candidate solution; the method of adjusting the screening radius is: An adjustment coefficient is randomly generated in , the screening radius is multiplied by the adjustment coefficient to obtain the adjustment radius, and the screening radius is adjusted to the adjustment radius.

[0096] The control effect is calculated by constructing an effect calculation model, using a candidate set corresponding to real-time parameters and calculation solutions as calculation data, inputting this data into the trained effect calculation model, and predicting the corresponding control effect. The control effect is a comprehensive representation of the robot's flexible joint's trajectory tracking accuracy and flexible vibration suppression capability. Trajectory tracking accuracy is the error between the flexible joint's actual trajectory and the target trajectory, while flexible vibration suppression capability is the vibration level of the flexible joint. Higher trajectory tracking accuracy and stronger flexible vibration suppression capability—that is, smaller error between the flexible joint's actual trajectory and the target trajectory and lower vibration level—result in higher control effect.

[0097] The effect calculation model is also a deep neural network model. In the training of the effect calculation model, when constructing a training data set, real-time parameters and calculation solutions are generated according to the above method, and the flexible joint movement is simulated or actually controlled according to the calculation solution to obtain new dynamic parameters. The new dynamic parameters and the calculation solution are compared to annotate the control effect of the calculation data in the training data set (for example, based on the response data in the calculation solution and the response data in the new dynamic parameters, the error value calculation method in the real-time prediction model is used to calculate the control effect) to complete the construction of the training data set. Based on the training data set, a feedforward neural network structure is used to construct the effect calculation model. Except for the different input and output of the model, the construction process of the effect calculation model is consistent with the construction process of the real-time prediction model, which will not be repeated in this embodiment.

[0098] Methods for generating candidate solution boundaries include:

[0099] Obtain the set label corresponding to each candidate solution and mark it as an edge label; compare each edge label separately, mark the edge label with the largest value as the maximum label, and mark the edge label with the smallest value as the minimum label; based on the maximum label and the minimum label, generate the candidate solution boundary, that is, take the maximum label as the maximum value of the candidate solution boundary, and take the minimum label as the minimum value of the candidate solution boundary.

[0100] Methods for computing boundary solutions include:

[0101] From the interval Randomly generated numerical values, and use them as the calculation coefficients corresponding to each candidate solution in turn; add the maximum label to the minimum label to obtain the label coefficient; multiply the label coefficient by each calculation coefficient to obtain the boundary coefficient corresponding to each candidate solution; subtract the corresponding candidate solution from each boundary coefficient to obtain the boundary solution corresponding to each candidate solution.

[0102] This embodiment adopts two neural network structures, a real-time prediction model and a future prediction model, to operate in parallel. The former can capture the nonlinear coupling relationship of the flexible joint in static and transient states, and the latter can capture the timing dependency and dynamic evolution law of the flexible joint during the movement process, thereby realizing comprehensive modeling and accurate prediction of the motion characteristics of the flexible joint. At the same time, an intelligent decision-making algorithm is used to dynamically optimize the future motion parameters, and the control parameters are updated in real time through a rolling time domain strategy, so that the control system can quickly respond to the real-time state changes of the flexible joint, effectively improving the control accuracy, response speed and adaptability of the flexible joint. The advantages of deep learning in nonlinear dynamic system modeling and optimization control are fully utilized, which not only improves the adaptability of the robot in a dynamic environment, but also reduces the computational burden of traditional control methods, and improves the real-time performance and prediction accuracy of the control system.

[0103] Example 2: The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the robot flexible joint predictive control method based on the dual neural network model in Example 1 are implemented.

[0104] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0105] Example 3: The present invention also provides a computer device, including a processor and a memory storing a plurality of computer instructions. When the computer instructions are executed by the processor, the steps of the robot flexible joint predictive control method based on the dual neural network model in Example 1 are implemented.

[0106] The memory and processor are electrically connected, directly or indirectly, to enable data transmission or interaction. For example, these components may be electrically connected via one or more communication buses or signal lines. The memory stores a computer program executable on the processor, and the processor implements the method of the present invention by executing the computer program stored in the memory.

[0107] The memory may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0108] The processor may be an integrated circuit chip with data processing capabilities. Such a processor may be a general-purpose processor, including a central processing unit (CPU) or a network processor (NP). It may implement or execute the methods, steps, and logic diagrams disclosed in the embodiments of the present invention. The general-purpose processor may be a microprocessor or any conventional processor.

[0109] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A robot flexible joint predictive control method based on a dual neural network model, characterized in that: The robot flexible joint predictive control method based on the dual neural network model includes: Collect the dynamic parameters of each flexible joint of the robot under various working conditions and build a motion state database; Based on the motion state database, the dynamic parameters are divided into observation parameters and response parameters, and a feedforward neural network structure is used to build a real-time prediction model; The observation parameters of each flexible joint are collected in real time, and the joint labels are set respectively and then input into the real-time prediction model to obtain the predicted response parameters as real-time characteristic parameters; Based on the motion state database, the dynamic parameters of a flexible joint under the same working condition are taken as a set of parameters, and a recurrent neural network structure is used to build a future prediction model; Input the real-time characteristic parameters into the future prediction model to predict the future dynamic parameters of each flexible joint; An intelligent decision-making algorithm is used to dynamically optimize the future dynamic parameters of each flexible joint to obtain optimized motion parameters, and each flexible joint is controlled based on the optimized motion parameters.

2. The robot flexible joint predictive control method based on the dual neural network model according to claim 1 is characterized in that: Based on the motion state database, the dynamic parameters are divided into observation parameters and response parameters, and a feedforward neural network structure is used to build a real-time prediction model, including: Each kinetic parameter in the motion state database is divided into observation parameters and response parameters; Obtain multiple sets of different observation parameters and corresponding response parameters from the motion state database, and set corresponding joint labels for each set of observation parameters according to the joint type corresponding to each set of observation parameters, where the joint label is the label corresponding to the joint type; The observation parameters after setting the joint labels are used as the input of the feedforward neural network structure, and the corresponding response parameters are used as the prediction targets of the feedforward neural network structure. The feedforward neural network structure is trained and updated, and after the training and updating is completed, the feedforward neural network structure is output as a real-time prediction model.

3. The robot flexible joint predictive control method based on the dual neural network model according to claim 1 is characterized in that: Based on the motion state database, the dynamic parameters of a flexible joint under the same working condition are taken as a set of parameters, and a recurrent neural network structure is used to construct a future prediction model, including: Obtain the dynamic parameters corresponding to a flexible joint in the motion state database and mark them as training parameters. The training parameters corresponding to the same working condition are used as a set of parameter sets. The training parameters in each parameter set are sorted from early to late according to the corresponding collection time to generate a training set corresponding to each parameter set. Set the sliding step size and sliding window length , the training parameters in each training set are converted into multiple training samples using the sliding window method. A training sample contains Set training parameters, use the training samples as the input of the recurrent neural network structure, and The training parameters are used as prediction targets of the recurrent neural network structure, the recurrent neural network structure is trained and updated, and after the training and updating, the recurrent neural network structure is output as a motion prediction model of the corresponding flexible joint; The motion prediction models of all flexible joints are used as future prediction models.

4. The robot flexible joint predictive control method based on the dual neural network model according to claim 3 is characterized in that: The inputting of the real-time characteristic parameters into the future prediction model to predict the future dynamic parameters of each flexible joint includes: The observation parameters of a flexible joint collected in real time and the corresponding response parameters in the real-time characteristic parameters are used as real-time parameters; From the motion state database, obtain the corresponding real-time parameters Group of continuous dynamic parameters; among them, real-time parameters and corresponding The acquisition time of the continuous dynamic parameters of the group is continuous, and the real-time parameters are consistent with the corresponding The continuous dynamic parameters of each group correspond to the same flexible joint; The real-time parameters are compared with the corresponding A set of continuous dynamic parameters are input together into the motion prediction model of the corresponding flexible joint to predict the future dynamic parameters of the current flexible joint.

5. The robot flexible joint predictive control method based on the dual neural network model according to claim 1 is characterized in that: The intelligent decision-making algorithm is used to dynamically optimize the future dynamic parameters of each flexible joint to obtain optimized motion parameters, including: Obtain the earliest dynamic parameters of each flexible joint corresponding to the future moment and mark them as parameters to be optimized; According to the parameters to be optimized, each flexible joint is constructed Group candidate set, is an integer greater than 1; From each flexible joint The best set corresponding to each flexible joint is selected from the candidate set; the best set corresponding to all flexible joints is used as the optimized motion parameter.

6. The robot flexible joint predictive control method based on the dual neural network model according to claim 5 is characterized in that: According to the parameters to be optimized, each flexible joint is constructed Group candidate set, including: A preset perturbation set includes the perturbation value corresponding to each parameter in the dynamic parameters; Obtain the dynamic parameters corresponding to a flexible joint among the parameters to be optimized and mark them as disturbance parameters; Add the disturbance value of the corresponding parameter in the disturbance set to each parameter in the disturbance parameter to obtain the maximum disturbance value; Subtract the disturbance value of the corresponding parameter in the disturbance set from each parameter in the disturbance parameter to obtain the minimum disturbance value; According to the maximum and minimum disturbance values corresponding to each parameter, the disturbance range value corresponding to each parameter in the disturbance parameter is constructed; A value is randomly selected from each perturbation range value, and a set of candidate sets is constructed. Group candidate set, The candidate sets of groups are all different.

7. The robot flexible joint predictive control method based on the dual neural network model according to claim 5 is characterized in that: The corresponding The best set corresponding to each flexible joint is selected from the candidate set, including: Set an increasing numerical label for each candidate set in turn and mark it as a set label. The range of the set label is , and according to Determine the initial screening center and initial screening radius; Define the iterative process, which is: generate within the range of the set label candidate solutions, each candidate solution corresponds to a set label one by one. ; Calculate the control effect corresponding to each candidate solution in turn and generate the candidate solution boundary; Based on the candidate solution boundary, calculate the boundary solution corresponding to each candidate solution, and calculate the control effect corresponding to each boundary solution in turn; Candidate solutions and boundary solutions are collectively referred to as calculated solutions, compare the control effects of all calculated solutions, move the screening center to the calculated solution with the highest control effect, and adjust the screening radius; The iterative process is executed. When the number of executions of the iterative process is greater than or equal to the preset iteration threshold, the iterative process ends, the calculated solution corresponding to the screening center is marked as the best solution, and the candidate set corresponding to the set label corresponding to the best solution is obtained and marked as the best set.

8. The robot flexible joint predictive control method based on the dual neural network model according to claim 7 is characterized in that: The method for generating the candidate solution is: in the interval Generate a random coefficient, multiply the screening radius by the random coefficient and add the screening center as a candidate solution; The method for adjusting the screening radius is as follows: A random adjustment coefficient is generated in , the screening radius is multiplied by the adjustment coefficient to obtain the adjustment radius, and the screening radius is adjusted to the adjustment radius.

9. The robot flexible joint predictive control method based on the dual neural network model according to claim 7 is characterized in that: The control effect calculation process is as follows: Based on the motion state database, a deep neural network structure is used to build an effect calculation model; The observation parameters of a flexible joint collected in real time and the corresponding response parameters in the real-time characteristic parameters are used as real-time parameters, the real-time parameters and the calculation solutions belonging to the same flexible joint are input into the effect calculation model, and the effect calculation model outputs the control effect.

10. The robot flexible joint predictive control method based on the dual neural network model according to claim 7, characterized in that: The method for generating candidate solution boundaries comprises: Obtain the set label corresponding to each candidate solution and mark it as an edge label; compare each edge label separately, mark the edge label with the largest value as the maximum label, and mark the edge label with the smallest value as the minimum label; generate the candidate solution boundary based on the maximum label and the minimum label, that is, take the maximum label as the maximum value of the candidate solution boundary, and take the minimum label as the minimum value of the candidate solution boundary; The step of calculating the boundary solution corresponding to each candidate solution according to the candidate solution boundary includes: From the interval Randomly generated numerical values, and use them as the calculation coefficients corresponding to each candidate solution in turn; add the maximum label to the minimum label to obtain the label coefficient; multiply the label coefficient by each calculation coefficient to obtain the boundary coefficient corresponding to each candidate solution; subtract the corresponding candidate solution from each boundary coefficient to obtain the boundary solution corresponding to each candidate solution.

Images