Measurement robot kinetic parameter identification method based on IABC-ANN neural network
Through the method based on IABC-ANN neural network, combined with swarm algorithm and filtering processing, the problems of low accuracy and poor generalization ability of traditional dynamic parameter identification methods are solved, and higher robot control performance and recognition accuracy are achieved.
Patent Information
- Application Number
- CN202411051686.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-15
AI Technical Summary
Traditional robot dynamic parameter identification methods have problems such as low recognition accuracy, poor generalization ability and easy to fall into local optimality, which is difficult to meet the requirements of high-performance control.
Using an IABC-ANN neural network method, a robot dynamics model is established through the Lagrangian equation, an excitation trajectory and verification trajectory are designed, and the neural network weights and thresholds are optimized by combining the swarm algorithm, and the joint is compensated by using the Coulon-viscosity friction model. Friction, removes no-influence parameters, and filtering and Monte Carlo method are used to optimize the excitation trajectory.
It improves the accuracy and robustness of robot dynamic parameters recognition, enhances control performance, avoids local optimal problems, and improves the prediction accuracy of neural networks.
Smart Images

Figure CN118952190A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for identifying robot dynamic parameters, specifically a method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network, and relates to a multi-degree-of-freedom serial industrial robot widely used in the fields of workpiece measurement, intelligent manufacturing equipment, industrial production, etc. Accurate dynamic parameters can control the robot with higher precision. Background Art
[0002] At present, in many industries, there are very high requirements for the measurement accuracy of workpieces. Since pneumatic measurement has the characteristics of non-contact, simple structure, and easy operation, it is widely used in production fields such as machinery manufacturing and automobile manufacturing. The traditional method of measuring parts by hand can no longer meet the requirements of automated lines, which seriously restricts production efficiency. With the rapid development and widespread application of industrial robot technology research in recent years, the combination of robot technology and pneumatic measurement system has retained the stability and easy operation characteristics of pneumatic measurement technology, and realized automated measurement, thus realizing the automated pneumatic measurement technology of workpieces.
[0003] When measuring, the robot first grabs the workpiece to be measured on the conveyor belt through a mechanical claw, and puts the workpiece into the V-shaped probe of the pneumatic measuring instrument for measurement. The pneumatic measuring instrument processes the measurement data and uploads it to the host computer, so that the host computer analyzes the measurement results and determines whether the measurement results are qualified. The accuracy of the measurement results depends on the measurement accuracy, and higher measurement accuracy requires that the workpiece can be accurately placed on the V-shaped probe, so it requires higher control performance for the robot, and accurate robot dynamic model parameters are the basis for studying the robot's motion control performance, force control performance and safety performance. With the increasing application of industrial robots, for example, the realization of functions such as robot collision detection, force-based impedance control and trajectory tracking control all rely on the dynamic parameter identification value to predict the robot's joint driving torque. At present, the common methods for obtaining robot dynamic parameters are as follows: disassembly measurement method, CAD method, and experimental identification method. Compared with the parameters obtained by disassembly measurement method and CAD method, the experimental identification method is more accurate, so it is also the widely used method at present. Although the dynamic parameters of the robot can be derived from the quality attribute module of Solidworks software through its three-dimensional model, this method ignores accessories such as joint motors, bearings and reducers, and cannot obtain the joint friction parameters of the robot. The parameter identification accuracy is difficult to meet the growing high-performance control requirements. When using traditional learning algorithms for dynamic parameter identification, local optimal problems are usually encountered, which means that the optimization algorithm may converge to a local optimum rather than a global optimum. At the same time, since traditional neural networks lack generalization ability when processing data, the constructed neural network model is difficult to effectively promote to new data. Affected by the initial weights and thresholds of the neural network, different initial parameters may lead to different results. In order to solve the above-mentioned technical problems, the present invention provides a method for identifying the dynamic parameters of a measurement robot based on an IABC-ANN neural network. Summary of the invention
[0004] In order to solve the problems of low identification accuracy, poor generalization ability and easy falling into local optimum when using traditional learning algorithms for dynamic parameter identification in the prior art, the present invention provides a method for dynamic parameter identification of a measuring robot based on an IABC-ANN neural network.
[0005] In order to achieve the above object, the specific scheme adopted by the present invention is: A method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network mainly includes the following steps: S1. Establish the dynamic model of the n-joint robot according to the Lagrange equation. According to the dynamic model, derive the linear expression of the robot. According to the minimum inertia principle, optimize the linear expression, and sort out the dynamic minimum inertia parameters and corresponding observation matrix that need to be identified. S2. Design the robot's excitation trajectory and optimize the excitation trajectory; S3, the robot moves according to the optimized excitation trajectory, collects the angles, angular velocities and angular accelerations of n joints during the robot movement, obtains multiple groups of observation matrices as test samples, and obtains joint feedback currents at the same time, and calculates the joint torques of n joints according to the torque coefficients; S4, filtering the collected data; S5, constructing an IABC-ANN neural network, optimizing the weights and thresholds of the constructed IABC-ANN neural network by using a bee colony algorithm, taking the multiple groups of observation matrices collected in step S3 as inputs of the neural network, and the joint torques of n joints as outputs, training the constructed IABC-ANN neural network to obtain a trained IABC-ANN neural network, and identifying the dynamic minimum inertia parameter by using the trained IABC-ANN neural network; S6. Design a verification trajectory that is different from the excitation trajectory, and make the robot move along the verification trajectory. Collect the robot's dynamic parameters and joint feedback current for multiple cycles, compare the actual joint torque values obtained with the theoretical joint torque values obtained based on the identified dynamic minimum inertia parameters, and use the torque residual value to evaluate the accuracy of the desired dynamic parameters.
[0006] Furthermore, in step S1, the linear expression is optimized into the following form according to the minimum inertia principle: in: represents the reorganized joint torque vector, represents the simplest observation matrix after reorganization, Represents the dynamic minimum inertia parameter.
[0007] Furthermore, in step S1, when n is 6, according to the geometric relationship, the 72 dynamic parameters of the six joints of the robot are eliminated according to the principle of minimum inertia to eliminate the parameters that have no effect on the dynamic model, and 48 dynamic minimum inertia parameters that need to be identified are obtained, among which the minimum inertia parameters of joint one of the six-joint robot are Izz, fc, fv; the minimum inertia parameters of joint two and joint three are Ixx, Ixy, Ixz, Iyz, Izz, mx, my, fc, fv; the minimum inertia parameters of joint four and joint five are Ixx, Ixy, Ixz, Iyy, Iyz, mx, mz, fc, fv; the minimum inertia parameters of joint six are Ixx, Ixy, Ixz, Iyz, Izz, mx, my, fc, fv.
[0008] Further, in step S2, the excitation trajectory for dynamic parameter identification is established based on the fifth-order Fourier series, and the reorganized simplest observation matrix condition number As the objective function of the excitation trajectory optimization, a certain motion range of the robot is taken as a constraint condition to obtain the optimal excitation trajectory.
[0009] Furthermore, in step S4, the specific method for filtering the collected data is: first, large errors are removed from multiple sets of observation matrix data according to the differential criterion, and then the joint encoder signal is filtered using a second-order Butterworth filter, and then the data after the differential operation is subjected to zero-phase digital filtering using a filtfilt filter to reduce the errors caused by data noise.
[0010] Furthermore, in step S5, adjusting the weights and thresholds of the constructed IABC-ANN neural network based on the bee colony algorithm includes the following steps: (1) Initialization parameters: the maximum number of iterations of the bee colony, the queen bee life span, the upper limit of pheromone retention, and the population size; (2) Initialize the population parameters, randomly initialize the positions of all individuals in the population, and calculate the threat level and speed of the positions; check whether the queen bee exists. If not, initialize the queen bee position as the current highest threat point, and initialize the queen bee speed and pheromone vector at the same time, check the queen bee lifespan, and destroy it if the set lifespan value is reached; (3) Input pheromones according to the current highest global threat level and update the queen bee position and speed; (4) Individuals in the population update their own positions and speeds based on their current speeds and the current position of the queen bee; (5) Calculate the threat level of all individuals to update the global threat level and determine whether the number of iterations has been reached. If so, terminate the optimization and output the final population state and target point. If not, continue the optimization. (6) Input the optimized weights and thresholds into the constructed IABC-ANN neural network.
[0011] Beneficial effects:
[0012] Compared with the prior art, the present invention uses the Coulomb-viscous friction model to compensate for the joint friction of the robot when modeling the dynamic model, which can reduce the identification error caused by the unmodeled friction force, so that the identification method has higher robustness under different motion conditions; in order to simplify the identification process, according to the minimum inertia principle, the terms in the robot inertia parameters that have no effect on the dynamic model are eliminated; when optimizing the excitation trajectory, the Monte Carlo method is used to solve the robot workspace, which can avoid the concentration of data in certain specific areas, prevent data shortage or data deviation, and more comprehensively reflect the dynamic characteristics of the robot, thereby improving the accuracy of parameter identification; the condition number of the observation matrix is used as the optimization target of the excitation trajectory, and the coefficients of the sine and cosine terms of the excitation trajectory are optimized by the fmincon function in Matlab; on the basis of the traditional neural network identification method, the bee colony algorithm is introduced to optimize the weights and thresholds of the neural network. Traditional optimization methods (such as gradient descent) are prone to fall into local optimality on the surface of complex loss functions. The bee colony algorithm can more comprehensively explore the search space of weights and thresholds through group intelligence and random search, thereby finding a more optimal parameter configuration, so that the obtained neural network prediction model has a higher prediction accuracy. The present invention is helpful to improve the robot control performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a flow chart of neural network dynamics parameter identification in the present invention.
[0014] Figure 2 It is a schematic diagram of the optimal excitation trajectory of the joint in the present invention.
[0015] Figure 3 It is a cloud map of the robot workspace in the present invention.
[0016] Figure 4 It is a schematic diagram of the neural network model in the present invention.
[0017] Figure 5 It is the RMSE graph during the hybrid network training process in the present invention.
[0018] Figure 6 It is a comparison diagram of the predicted torque and actual torque of the first three joints in the present invention. DETAILED DESCRIPTION
[0019] The technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0020] The present invention provides a method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network, and each step is described in detail below.
[0021] S1. Establish the dynamic model of the n-joint robot based on the Lagrange equation. According to the dynamic model, derive the linear expression of the robot. According to the minimum inertia principle, optimize the linear expression, and sort out the dynamic minimum inertia parameters and corresponding observation matrix that need to be identified.
[0022] S11. Establish the dynamic model of the n-joint robot based on the Lagrange equation.
[0023] For a robot system with n joints, considering the influence of joint friction and gravity, its dynamic equation can be expressed by a second-order differential equation: Where: the subscripts i and k represent the i-th row and k-th column of the matrix respectively, τ∈R 6×1 is the joint torque vector, M i (q)∈R 6 ×6 is the inertia matrix, C ik ∈R 6×6 is the centrifugal force and Coriolis force matrix, G(q)∈R 6×1 is the gravity vector, is the joint friction torque vector, They represent the joint angle, angular velocity and angular acceleration of the robot respectively.
[0024] Taking the JAKA MiniCobo collaborative robot as an example, the following details the steps of establishing the n-joint robot dynamics model through the Lagrangian equation.
[0025] For an n-joint robot, assume that a set of generalized coordinates θ∈R in the robot configuration space n , for an open kinematic chain where all joints are driven joints, we can always choose the generalized coordinate θ to be a vector consisting of the values of the joint variables and express the generalized force as τ∈R n After selecting the generalized coordinates θ and determining the generalized forces τ, for the Lagrangian function The statement is as follows: in: represents the kinetic energy of the entire robot system, Represents the potential energy of the entire robot system.
[0026] The kinetic energy of the entire robot system It is defined as the sum of the translational kinetic energy of the entire object contracting to the center of mass and the rotational kinetic energy about the center of mass, expressed as: Where: I is the moment of inertia matrix of the rigid body, m represents the total mass of the rigid body, ω is the angular velocity of the connecting rod in the inertial coordinate system, and v is the angular velocity of the connecting rod in the inertial coordinate system.
[0027] In the rigid body of the robot rod, assume that the mass of the point mass i is The total mass of the rigid body is m, let r i =(x i ,y i ,z i ) is the fixed position of the point mass i in the inertial coordinate system {b}, and R is used to represent the attitude transformation between the inertial coordinate system and the connecting rod coordinate system. The relationship between the total mass m of the rigid body and the moment of inertia matrix I of the rigid body is obtained as follows:
[0028] Assume that the robot joint velocity v i and the joint angular velocity w i It is known that through the Jacobian matrix, the linear velocity and angular velocity on the connecting rod are: in: are the Jacobian matrices of the joint velocity and joint angular velocity in the form of the joint fixed coordinate system, respectively.
[0029] By using formulas (3)-(5), the total kinetic energy of the n-link robot can be obtained: for: in: is the point mass, m ij (q) is the (i,j)th element of the n×n mass matrix M(q), I i is the generalized equivalent inertia, R i (q) Homogeneous coordinates of a point on the connecting rod with respect to coordinate system {i}.
[0030] In rigid body dynamics, assuming that the mass of each link of the robot is concentrated at the center of mass, the total potential energy of the robot is the sum of the potential energies of each link, expressed as: Where: g is the gravity vector in the inertial coordinate system, vector Represents the centroid coordinates.
[0031] According to the Lagrangian function, The dynamic equation of each joint is: Among them: P represents the The potential energy of the joint, K represents the The kinetic energy of each joint, is the generalized force corresponding to the coordinates, is the generalized representation of the centripetal term and the Coriolis term, represents the generalized velocity, represents the generalized component of potential energy, represents generalized acceleration.
[0032] The robot dynamics equation is obtained by rearranging formula (8): in: are the vector of the force and centrifugal force, and G(q) is the gravity vector.
[0033] The Coulomb-viscous friction model is selected to compensate the joint friction of the robot, based on the linear relationship between the robot inertia parameters and the joint torque.
[0034] The Coulomb-viscous friction between joints can be expressed as:
[0035] Where: τ f represents the Coulomb-viscous friction between the joints, f c is the Coulomb friction coefficient, f v is the viscous friction coefficient, is a symbolic function.
[0036] Therefore, the dynamic equation of formula (9) is expressed as follows based on considering joint friction:
[0037] S12. Based on the dynamic model, derive the linear expression of the robot.
[0038] The kinetic model (Formula (11)) can be linearized into the following form: Where: τ∈R 6×1 is the joint torque vector, p d is the complete inertia parameter, that is, the set of inertia parameters of n joints, is the corresponding observation matrix, the complete inertial parameter p d For a 10n×1 vector: in: For the robot Inertia parameters of the joint. No. The vector form of the inertia parameters of the joint is: in, is the mass of the connecting rod, is the first moment of mass, Respectively represent connecting rod The inertia tensor of the xx axis, yy axis, zz axis, xy axis, xz axis, and yz axis, is the joint friction coefficient.
[0039] S13. According to the minimum inertia principle, the complete dynamic parameter set of all joints that has no influence on the dynamic model is eliminated to optimize expression (12) and sort out the minimum dynamic inertia parameters and corresponding observation matrix to be identified.
[0040] Since the observation matrix Not full rank, some column vectors are linearly correlated, for any are all zero, then ψ T ψ is not reversible and the least squares method cannot be used directly.
[0041] From the above, we can see that not all inertial parameters have an impact on the dynamic model. If the observation matrix There exists a vector ψ m The following relationship is satisfied: ψ m =a 1m ψ 1m +a 2m ψ 2m …+a jm ψ jm (15) Where: 1m ,ψ 2m ,…ψ jm Represents the observation matrix and the vector ψ m Linearly dependent columns, a 1m ,a 2m ,…a jm Represent their linear correlation coefficients respectively.
[0042] From the above formula, we can know that ψ 1m ,ψ 2m ,…ψ jm is m A maximal linearly independent vector group, at this time, the corresponding column in the observation matrix can be eliminated without affecting the accuracy of the model. By using this method to eliminate some inertia parameters, the remaining parameters are the minimum inertia parameters after reorganization.
[0043] According to the above method, the linear expression in step 2 is adjusted to obtain the optimized kinetic model as follows: in: represents the reorganized joint torque vector, Represents the simplest observation matrix after reorganization. For the convenience of description, replace, Represents the dynamic minimum inertia parameter.
[0044] The reorganized observation matrix The column vectors are linearly independent, so at this time any sampling point Both can make ψ m Full rank, then the least squares method can be used to solve
[0045] When n is 6, the collaborative robot has a total of 72 parameters. After eliminating the parameters that have no effect on the dynamic model and reorganizing them according to the above method, a total of 48 identifiable inertial parameters are obtained. The set of its minimum dynamic inertial parameters is shown in Table 1.
[0046] Table 1 Collection of dynamic minimum inertia parameters
[0047] S2. Design the robot's excitation trajectory and optimize the excitation trajectory.
[0048] In this step, the excitation trajectory for dynamic parameter identification is established based on the fifth-order Fourier series, and the reorganized simplest observation matrix condition number As the objective function of the excitation trajectory optimization, a certain motion range of the robot is taken as a constraint condition to obtain the optimal excitation trajectory.
[0049] The fifth-order Fourier series is selected as the excitation trajectory for dynamic parameter identification: in: is the coefficient to be optimized; is a constant term, are the coefficients of the quintic polynomial, t is the sampling period, l is the order of the Fourier series, t f is the trajectory operation period, ω f is the fundamental frequency, j = [t / t f ] is the rounding function. According to the starting position of the excitation trajectory, the following relationship can be used to obtain The value of Values: in: For the The joint start angle.
[0050] The objective function of the incentive trajectory optimization is: in: Represents the simplest observation matrix after reorganization The spectral norm of is the arithmetic square root of the largest eigenvalue of the corresponding matrix, is the condition number of the reorganized simplest observation matrix, which determines the sensitivity of the system to noise. The smaller it is, the less the solution of its system of equations is disturbed by noise, and the smaller the error is; The constraints are: Condition one: Condition 2: q(0)=q(t′ f )=q0 Condition three: -q max ≤q(t′)≤q max Where: S represents the Cartesian space of the robot end, s(q(t′)) is the end trajectory; q0 is the joint angle at the initial moment of the cycle, q(0), They represent the joint angle, joint angular velocity and joint angular acceleration at the initial moment respectively. They represent the joint angle, joint angular velocity and joint angular acceleration at the termination time respectively. They represent the joint angle, joint angular velocity and joint angular acceleration at time tˊ respectively, is the motion constraint value of each joint.
[0051] The method to determine the Cartesian space S of the robot end in condition 3 is as follows: Taking a 6-joint robot as an example, first calculate the maximum value of the robot's joint angle q max and the minimum value q min , and combine with the random function rand in Matlab to generate a random function, then you will get a random angle q i (q i =q min +(q max -q min )*rand), and then a set of joint angles are randomly given. According to the transformation matrix of the robot end joint in the base coordinates, the center position of the end can be obtained [p x ,p y ,p z]; After that, each joint randomly generates 10,000 variables within the working range, calculates the position of the end point, and draws a cloud map of the working space, such as Figure 3 As shown in the figure, through the workspace cloud map, we can get the approximate working range of the robot: X-axis (-580~580), Y-axis (-580~580), Z-axis (-400~770), the unit is mm.
[0052] Finally, the excitation trajectory expression (18) is solved by the fmincon function in the Matlab optimization algorithm toolbox. Solution, and Substituting the value of into equation (19) yields the fifth-order polynomial coefficient The optimal excitation trajectory is obtained by using the optimal excitation trajectory. The optimal excitation trajectory diagram of the six joints is obtained, as shown in Figure 2 The Matlab solution record shows that the condition coefficient of the excitation trajectory corresponding to the measurement matrix is 95.2431, indicating that the measurement matrix is well-conditioned and has good noise resistance.
[0053] S3. The robot moves according to the optimized excitation trajectory, collects the angles, angular velocities and angular accelerations of n joints during the robot's movement, obtains multiple groups of observation matrices as test samples, and obtains joint feedback currents at the same time. According to the torque coefficient, the joint torques of the n joints are calculated.
[0054] S4. Filter the collected data.
[0055] The data collected in the experiment are preprocessed by filtering method to reduce the influence of data noise on the experiment. The specific steps are as follows: first, the multiple sets of data obtained from the experiment are filtered according to the differential criterion to remove large errors, and the joint encoder signal is filtered using a second-order Butterworth filter. Then, the data after differential operation is subjected to zero-phase digital filtering using a filtfilt filter to reduce the error caused by data noise.
[0056] S5, such as Figure 1 As shown, an IABC-ANN neural network is constructed, and the weights and thresholds of the constructed IABC-ANN neural network are optimized by a bee colony algorithm. The multiple groups of observation matrices collected in step S3 are used as inputs of the neural network, and the joint torques of n joints are used as outputs. The constructed IABC-ANN neural network is trained to obtain a trained IABC-ANN neural network, and the dynamic minimum inertia parameters are identified by the trained IABC-ANN neural network.
[0057] S51. Constructing IABC-ANN neural network
[0058] like Figure 4The figure shows a schematic diagram of the constructed neural network model. The constructed IABC-ANN neural network model expression is as follows: Where: Lλ represents the weight of the neural network training model input λ at the Lth neuron, b Lλ is the threshold of the input λ at the Lth neuron, f λ is the transfer function.
[0059] The Relu function is used as the activation function of the model; the mean square error (RMSE) is selected as the loss function, and the loss function of the network is defined as follows: in: Indicates the actual output, Y M is the predicted value of the network, M represents the number of test sample sequences, and k represents the total number of test set samples.
[0060] In order to reduce the dependence of the constructed neural network on a certain eigenvalue and simplify the model, Dropout is selected for regularization.
[0061] S52, using a bee colony algorithm to optimize the initial weights and thresholds of the IABC-ANN neural network, mainly including the following steps: (1) Initialization parameters: the maximum number of iterations of the bee colony, the queen lifespan, the upper limit of pheromone retention, and the population size. The queen is assumed to be born in the Mth generation, and the pheromone it releases disappears after Nth generation, that is, the population can retain the global optimal position p for Nth generation. Best , then the pheromone produced by the queen bee in each iteration is the optimal position: P=[P Best (M′),P Best (M′+1)…P Best (X′)…P Best (M′+N′-1)] T (twenty three) Where: P represents the optimal position of the queen bee in each iteration, P Best (X′) represents the global optimal position of the swarm after Xˊ iterations; (2) Initialize the population parameters, randomly initialize the positions of all individuals in the population, and calculate the threat level and speed of the positions; check whether the queen bee exists. If not, initialize the queen bee position as the current highest threat point (optimal value point), and initialize the queen bee speed and pheromone vector at the same time, check the queen bee lifespan, and destroy it if the set lifespan value is reached; (3) Input pheromones according to the current global highest threat level and update the queen bee position and speed. In the Tth generation, the current queen bee position is x(T′). According to p in P Best (M′) to p Best (T′-1) to update the next moment position x(T′+1), the position update formula is as follows: Where: ξ⊙v(T′) represents the speed of the queen bee at the Tˊth iteration, Indicates the first Individuals, T′ max is the maximum number of iterations, a1, a2…a n represents the pheromone retention coefficient, a2, a3…a n-1 a1~a n The equal points of (4) The position x(T′) of the queen bee in the colony and the current speed of the queen bee Update its next position x(T′+1) and speed The update formula is as follows: Among them: ρ1, ρ2 are control parameters; (5) Calculate the threat level of all individuals to update the global threat level (global optimal value) and determine whether the number of iterations has been reached. If so, terminate the optimization output; if not, continue the optimization; (6) The optimized weights and thresholds are input as parameters of the neural network.
[0062] S53, taking the multiple groups of observation matrices collected in step S3 as the input of the neural network and the joint torques of the n joints as the output, training the constructed IABC-ANN neural network to obtain the trained IABC-ANN neural network, such as Figure 5 The figure shows the RMSE graph during the hybrid network training process. The trained IABC-ANN neural network is then used to predict the joint torque, and the dynamic minimum inertia parameters are identified based on the predicted joint torque.
[0063] S6. Design a verification trajectory that is different from the excitation trajectory, and make the robot move along the verification trajectory. Collect the robot's dynamic parameters and joint feedback current for multiple cycles, compare the actual joint torque values obtained with the theoretical joint torque values obtained based on the identified dynamic minimum inertia parameters, and use the torque residual value to evaluate the accuracy of the desired dynamic parameters.
[0064] Since the gravity and friction of the robot have a greater impact on its front three joints during movement, the proposed identification algorithm is more obvious in the front three joints. Therefore, in order to reduce the difficulty of the identification experiment, the rear three joints of the robot are fixed and only the front three joints are driven.
[0065] It should be noted that in order to obtain the actual value of joint torque and the theoretical value of joint torque, the values of each parameter are as follows: execute the designed optimal excitation trajectory, preset the operation cycle of the robot excitation trajectory to 20s, the sampling frequency to 1KHz, the harmonic number of the excitation trajectory to 5, divide the collected data into a training set (80%) and a test set (20%), and import them into the constructed neural network for training. When the weights and thresholds of the neural network are optimized using the bee colony optimization algorithm, the number of individuals in the bee colony is set to 50, the maximum number of iterations is 100, the initial network weight ω = 0.5, the threshold b = 0.02, and the maximum values in the iteration process are set to 0.9 and 0.3 respectively. The number of neurons in the neural network is set to 10, the sampling interval is 0.5s, the sampling period is 30s, the weights and thresholds of the neural network to be optimized are set between -1 and 1, and the maximum number of iterations of the neural network is set to 1000, and the maximum error is 0.1.
[0066] Figure 6 The comparison diagrams of the predicted torque and actual torque of the first three joints (a) are joint 1, (b) are joint 2, and (c) are joint 3. It can be seen from the figure that the predicted torque curve trained by the model has a good fit with the actual torque curve and can well predict the joint torque of the robot.
[0067] At this point, the fusion of the improved bee swarm algorithm and the neural network can be completed. In view of the shortcomings of the traditional identification method, this paper proposes a model based on IABC-ANN, which solves the problems of low identification accuracy, poor generalization ability and easy falling into local optimality when using traditional neural network models for identification.
[0068] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any equivalent changes or modifications made according to the essence of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network, characterized in that: The main steps are as follows: S1. Establish the dynamic model of the n-joint robot according to the Lagrange equation. According to the dynamic model, derive the linear expression of the robot. According to the minimum inertia principle, optimize the linear expression, and sort out the dynamic minimum inertia parameters and corresponding observation matrix that need to be identified. S2. Design the robot's excitation trajectory and optimize the excitation trajectory; S3, the robot moves according to the optimized excitation trajectory, collects the angles, angular velocities and angular accelerations of n joints during the robot movement, obtains multiple groups of observation matrices as test samples, and obtains joint feedback currents at the same time, and calculates the joint torques of n joints according to the torque coefficients; S4, filtering the collected data; S5, constructing an IABC-ANN neural network, optimizing the weights and thresholds of the constructed IABC-ANN neural network by using a bee colony algorithm, taking the multiple groups of observation matrices collected in step S3 as inputs of the neural network, and the joint torques of n joints as outputs, training the constructed IABC-ANN neural network to obtain a trained IABC-ANN neural network, and identifying the dynamic minimum inertia parameter by using the trained IABC-ANN neural network; S6. Design a verification trajectory that is different from the excitation trajectory, and make the robot move along the verification trajectory. Collect the robot's dynamic parameters and joint feedback current for multiple cycles, compare the actual joint torque values obtained with the theoretical joint torque values obtained based on the identified dynamic minimum inertia parameters, and use the torque residual value to evaluate the accuracy of the desired dynamic parameters.
2. The method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network according to claim 1, characterized in that: In step S1, the linear expression is optimized into the following form according to the minimum inertia principle: in: represents the reorganized joint torque vector, represents the simplest observation matrix after reorganization, Represents the dynamic minimum inertia parameter.
3. The method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network according to claim 1, characterized in that: In step S1, when n is 6, according to the geometric relationship, the 72 dynamic parameters of the six joints of the robot are eliminated according to the principle of minimum inertia to eliminate the parameters that have no effect on the dynamic model, and 48 dynamic minimum inertia parameters that need to be identified are obtained. Among them, the minimum inertia parameters of joint one of the six-joint robot are Izz, fc, fv; the minimum inertia parameters of joints two and three are Ixx, Ixy, Ixz, Iyz, Izz, mx, my, fc, fv; the minimum inertia parameters of joints four and five are Ixx, Ixy, Ixz, Iyy, Iyz, mx, mz, fc, fv; the minimum inertia parameters of joints six are Ixx, Ixy, Ixz, Iyz, Izz, mx, my, fc, fv.
4. The method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network according to claim 2, characterized in that: In step S2, the excitation trajectory for dynamic parameter identification is established based on the fifth-order Fourier series, and the reorganized simplest observation matrix condition number As the objective function of the excitation trajectory optimization, a certain motion range of the robot is taken as a constraint condition to obtain the optimal excitation trajectory.
5. The method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network according to claim 1, characterized in that: In step S4, the specific method for filtering the collected data is: first, large errors are removed from multiple groups of observation matrix data according to the differential criterion, then the joint encoder signal is filtered using a second-order Butterworth filter, and then the data after the differential operation is subjected to zero-phase digital filtering using a filtfilt filter to reduce the errors caused by data noise.
6. The method for identifying dynamic parameters of a measuring robot based on an IABC-ANN neural network according to claim 1, characterized in that: In step S5, adjusting the weights and thresholds of the constructed IABC-ANN neural network based on the bee colony algorithm includes the following steps: (1) Initialize the parameters, including the maximum number of iterations of the bee colony, the queen bee life span, the upper limit of pheromone retention, and the population size; (2) Initialize the population parameters, randomly initialize the positions of all individuals in the population, and calculate the threat level and speed of the positions; check whether the queen bee exists. If not, initialize the queen bee position as the current highest threat point, and initialize the queen bee speed and pheromone vector at the same time, check the queen bee lifespan, and destroy it if the set lifespan value is reached; (3) Input pheromones according to the current highest global threat level and update the queen bee position and speed; (4) Individuals in the population update their own positions and speeds based on their current speeds and the current position of the queen bee; (5) Calculate the threat level of all individuals to update the global threat level and determine whether the number of iterations has been reached. If so, terminate the optimization and output the final population state and target point. If not, continue the optimization. (6) Input the optimized weights and thresholds into the constructed IABC-ANN neural network.
