A robot contact force stability control method based on the influence of environmental position and stiffness

Abstract

The present invention discloses a robot contact force stabilization control method based on the influence of environmental position and stiffness, comprising the following steps: step 1: collecting the contact force between the robot execution end and the processing target and predicting the environmental stiffness through a trained extreme learning machine; step 2: using iterative learning control to estimate the environmental position and calculate the environmental position estimation error, and compensate the robot arm; step 3: using the MDPSO algorithm to calculate the impedance parameters m, b, k, and the learning coefficient k p , k p , k d Optimization is performed; the present invention can take into account different influencing factors in high-precision processing and manufacturing to meet the needs of processing and manufacturing.

Description

Technical Field

[0001] The invention belongs to the technical field of robot control, and in particular relates to a robot contact force stabilization control method based on the influence of environmental position and stiffness. Background Art

[0002] The working tasks of the robot arm have changed from non-contact tasks to contact tasks, which puts higher requirements on its functions and performance. In work scenarios that require interaction with the environment, in order to ensure that the target or the robot arm itself is not damaged and the task is carried out smoothly, the robot arm must be controlled compliantly.

[0003] The implementation methods of the compliance control of the robot arm are divided into passive compliance and active compliance; the force control accuracy of passive compliance control is low and its adaptability to the environment is limited, so active compliance control is the main research direction. The main research hotspot is impedance control. Its core idea is to treat the force and displacement deviation between the robot arm and the environment as a dynamic system to achieve indirect control of the contact force at the end of the robot arm. The position-based impedance controller consists of an external impedance control outer loop and a position control inner loop; the contact force error is converted into a correction amount of the reference trajectory through a given impedance relationship; then the end of the robot arm is controlled to move along the corrected path to achieve control of the contact force.

[0004] There are three main problems with the position-based impedance controller in the control process: First, if the steady-state force error of the impedance control is expected to be 0, accurate environmental position information and stiffness information are required, but these two pieces of information are difficult to obtain accurately in the actual production process. Second, the determination of the impedance parameters of the impedance control is the core issue of the algorithm. Only by determining the appropriate impedance parameters can accurate force control of the interaction between the robot and the environment be achieved. Third, in the design process of general impedance controllers, it is assumed that there is no error in the motion control loop of the robot, and the robot can perfectly track the compensated trajectory. However, due to the geometric parameter factors of the robot itself and the influence of non-geometric factors in the motion process, the position control of the robot will have errors and cannot track the compensated trajectory. Summary of the invention

[0005] In order to solve the above problems, the purpose of the present invention is to provide a robot contact force stabilization control method based on the influence of environmental position and stiffness, which can take into account different influencing factors in high-precision machining and meet the needs of machining and manufacturing.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A robot contact force stabilization control method based on the influence of environmental position and stiffness comprises the following steps:

[0008] Step 1: Collect the contact force between the robot execution end and the processing target and predict the environmental stiffness through the trained extreme learning machine;

[0009] Step 2: Utilize estimated environment position Get the estimated ideal trajectory of the robot arm k e is the environmental stiffness, is the estimated environmental position, f d is the desired contact force;

[0010] Calculate the error between the ideal trajectory and the estimated ideal trajectory

[0011] The error between the ideal trajectory and the estimated ideal trajectory is brought into the impedance model to obtain the environmental position estimation error. m, b, and k are impedance parameters;

[0012] Establishing a robotic arm model based on iterative learning control:

[0013]

[0014] Where k is the number of iterations, the output of the k+1th iteration is equal to the output of the kth iteration plus the error compensation term, k p , k p , k d is the learning coefficient, τ is the integral at a certain time between 0 and t, e f is the force error, is the differential of the force error;

[0015] The sampling period and the current number of iterations are set, and the robot model is discretized to obtain the final environmental position estimation error; the final environmental position estimation error is used to compensate the end position of the robot, so that the robot moves according to the compensated trajectory to achieve stable force control;

[0016] Step 3: Use the MDPSO algorithm to calculate the impedance parameters m, b, k, and the learning coefficient k p , k p , k d Optimize; specifically include the following steps:

[0017] First, the initial particle swarm is randomly generated, and each particle i=1,2…N, N is the number of particles in the initial population;

[0018] Secondly, establish the fitness function; fitness function e f is the contact force error, e x is the position error, Tmax is the running time;

[0019] Then, using the current individual extreme value P best , the current global optimal value G best , the individual with the best response speed P v , the optimal individual P for steady-state error ss , optimal individual overshoot P os and the oscillating optimal individual P o Update particles and obtain the particle update rate;

[0020] Next, the inertia weight is dynamically changed. [a,b] is the variation interval, I and It max Represent the current number of iterations and the maximum number of iterations respectively; w i is the initial value of the inertia weight;

[0021] Finally, mutation operation is used to enhance the global search capability; after each iteration, according to Select a particle for mutation operation, and update the rate when the calculated γ is less than the set value Perform mutation operations;

[0022]

[0023] in, is the update rate after the mutation operation, λ is the coefficient of variation, ζ is a random number between 0 and 1, which is used to determine the mutation direction, Ub is the upper limit of mutation, and Lb is the lower limit of mutation.

[0024] Furthermore, the extreme learning machine in step 1 is trained by the following steps:

[0025] First, an initial extreme learning machine model is constructed, wherein the extreme learning machine model includes an input layer, a hidden layer with q nodes, and an output layer;

[0026] Secondly, construct a training sample set, control the robot movement so that its execution end contacts the processing target, and obtain the contact force in a single direction. and displacement data d i , calculate the environmental stiffness under the current contact force through two adjacent sets of data Construct the training sample set as {f ei ,k ei |i=1,2,…,n};

[0027] Finally, the constructed training sample set is used to train the initial extreme learning machine model to generate a trained extreme learning machine.

[0028] Furthermore, the number of nodes q in the hidden layer of the extreme learning machine is determined by collecting a test contact force test set F t And calculate the corresponding environmental stiffness K t , the contact force data set F t Input into the trained extreme learning machine to predict the environmental stiffness Calculate the average prediction error of the extreme learning machine m is the number of data in the test set, is the environmental stiffness K t The i-th value in The environmental stiffness The i-th value in ;

[0029] The variation range of q is set to 10% to 100% of the training sample set. For the average error value of the extreme learning machine corresponding to each q, the q corresponding to the optimal average error value is selected as the number of hidden layer nodes of the ELM.

[0030] Furthermore, the update rate of the particles obtained in step 3 specifically includes:

[0031] First, according to the current individual extreme value P best and the current global optimal value G best Update speed and location is the number of particles, k is the current number of iterations, is the current velocity of the particle, is the original speed, w is the inertia weight, is the current position of the particle, is the original position, c1 is a single learning factor, c2 is a global learning factor, R1 and R2 are random values ​​in the interval [0,1];

[0032] Then, the optimal individual P in response speed is obtained. v , the optimal individual P for steady-state error ss , optimal individual overshoot P os and the oscillating optimal individual P o ;

[0033] Finally, calculate the current individual and the individual with the best response speed P v , the optimal individual P for steady-state error ss , optimal individual overshoot P os and the oscillating optimal individual P o The total error The particle update rate is

[0034] Furthermore, the individual with the best response speed P v It refers to the fastest response speed of force error within one cycle. The smaller the response speed, the faster the speed, thus determining the individual with the best response speed P v ; T is the adoption period, f e is the contact force;

[0035] The optimal individual P for steady-state error ss It refers to the optimal individual of steady-state error in one sampling period, and the steady-state error e ss =f e (Td max ), f e (T max ) is the contact force at the maximum operating time, e ss The smallest one is the optimal individual P of the steady-state error ss ;

[0036] The optimal individual overshoot P os It refers to the optimal individual overshoot in one adoption cycle, the overshoot e os =f emax -f d , f emax is the maximum value of the actual contact force, f d is the expected contact force, e ss The smallest individual is the individual with the best overshoot;

[0037] The oscillating optimal individual P o It means the minimum number of oscillations in one cycle, satisfying [f d -f e (tT)][f d -f e (t+T)]<0 is an oscillation, and the individual with the least oscillation times is the optimal oscillation individual P o .

[0038] The present invention has the following beneficial effects:

[0039] (1) Aiming at the acquisition of environmental information, a method for acquiring environmental information based on extreme learning machine and iterative learning control is proposed, and the ideal trajectory of the robot arm movement is established based on the obtained environmental information.

[0040] (2) Aiming at determining the parameters of the impedance controller, a multi-directional particle swarm optimization algorithm is proposed to obtain more accurate impedance control parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of the present invention.

[0042] Figure 2 This is a training flow chart of the extreme learning machine of the present invention.

[0043] Figure 3 This is a flow chart of the environmental position estimation error compensation of the present invention.

[0044] Figure 4 It is a schematic diagram of the flow of the MDPSO algorithm of the present invention. DETAILED DESCRIPTION

[0045] The position-based impedance controller achieves indirect control of the contact force of the end effector of the industrial robot by analyzing the relationship between the robot and the environment. The impedance model can usually be regarded as a system composed of mass, damping and spring; the impedance model is:

[0046]

[0047] Among them, m, b, k are impedance parameters, x, and are the actual values ​​of the position, velocity and acceleration of the robot end, respectively, r , and are the ideal values ​​of the position, velocity and acceleration of the robot end; f e is the actual contact force, f d is the expected contact force, e f Force error.

[0048] When an industrial robot moves in free space and is not in contact with the environment, the robot performs trajectory planning based only on position errors. When the robot comes into contact with the environment, contact forces are generated, and the impedance controller responds to the contact forces to achieve a balance between position control and force control, thereby achieving force tracking. In order to accurately track the force of the robot arm, the actual trajectory of the robot arm Among them, k e is the environmental stiffness, x e is the environmental position, f d is the expected contact force. Only after obtaining accurate environmental parameters can the ideal reference trajectory of the industrial robot be obtained, so that the steady-state contact force error converges to 0.

[0049] like Figure 1 As shown, this embodiment provides a method for stabilizing contact force of an industrial robot under the influence of environmental position and stiffness, comprising the following steps:

[0050] (1) Environmental stiffness prediction;

[0051] The stiffness of an industrial robot system mainly includes the stiffness of the robot body, fixtures, and tools, etc., which has an important influence on the contact force and position control between the industrial robot and the environment, especially in the impedance control of a flexible robot. This embodiment uses an extreme learning machine (ELM) to obtain the environment stiffness.

[0052] The structure of the extreme learning machine ELM includes an input layer, a hidden layer with q nodes, and an output layer; the training of the extreme learning machine ELM is as follows Figure 2 As shown, the input of the extreme learning machine ELM is set to the contact force The output is the environmental stiffness k ei .

[0053] Extreme Learning Machine ELM training sample acquisition: Control the movement of the industrial robot so that its end effector contacts the processing target, with an interval displacement of d, and obtain the contact force in a single direction and displacement data d i ; Use two adjacent sets of data to calculate the environmental stiffness under the current contact force

[0054] The training sample is {f ei ,k ei |i=1,2,…,n}, the input vector is F=[f e1 f e2 … f en ], the output vector is K = [k e1 k e2 … k en ], the weight vector of the input layer is ω=[ω1 ω2 … ω q ] T , the bias vector is b = [b1 b2… b q ] T , the output of the hidden layer is H = G (f e1 ,f e2 ,…,f en ,ω1,ω2,…,ω q ,b1,b2,…,b q ), q is the number of nodes in the hidden layer, and the output layer weight β is a q-dimensional vector. The environmental stiffness predicted by ELM is Hβ=K T , H and K are known, so β ​​can be calculated.

[0055] The setting of the number of nodes q in the hidden layer has a great impact on the prediction performance of ELM. A too small value will increase the error of the fitting model, while a too large value will lead to overfitting and reduce the efficiency of model training. In order to determine the value of q, this embodiment uses another set of sampled data as a test set. and m is the number of samples in the test set, and F t As the input of the extreme learning machine ELM, the predicted environmental stiffness is obtained

[0056] The prediction error of the extreme learning machine ELM is: ΔK = K t -K.

[0057] The performance of the ELM with the current number of hidden layer nodes being q is calculated using the error average The range of q is set from 10% of the training set to the same as the training set. The average error of the ELM network corresponding to each q is obtained, and the q corresponding to the optimal average error is selected as the number of hidden layer nodes of the ELM. After the value of q is determined, the environmental stiffness is predicted in real time based on the trained ELM network.

[0058] (2) Environmental position compensation;

[0059] The ideal trajectory of the industrial robot in the force control task needs to be designed based on the environmental position, but it is difficult to obtain the accurate environmental position in practical applications. Therefore, in order to eliminate the influence of the environmental position error on the impedance control, this embodiment adopts iterative learning control ILC to repeatedly estimate and compensate for the environmental position error, thereby improving the performance of the impedance controller.

[0060] Using estimated environmental position Get the estimated ideal trajectory of the robot arm k e is the environmental stiffness, is the estimated environmental position, f d is the desired contact force.

[0061] The error between the ideal trajectory and the estimated ideal trajectory is Δx e is the error between the environmental positions.

[0062] Since the actual trajectory of the robot arm x r cannot be obtained accurately, so the ideal trajectory will be estimated Imported into the impedance model, we get further Among them, m, b, k are impedance parameters, x, and are the actual values ​​of the position, velocity and acceleration of the robot end, and are the estimated ideal values ​​of the position, velocity and acceleration of the robot end respectively; f e is the actual contact force, fd is the expected contact force, e f is the force error;

[0063] according to Defining the error in the estimation of the environmental position in impedance control

[0064] In order to obtain the environmental position estimation error, the robotic arm model is established based on iterative learning control ILC k is the number of iterations. The output of the k+1th iteration is equal to the output of the kth iteration plus the error compensation term. p , k p , k d is the learning coefficient, τ is the integral at a certain time between 0 and t, e f is the force error, is the differential of the force error.

[0065] Set T as the sampling period, I k is the current iteration number, and the discretized robot model obtains the final environmental position estimation error;

[0066]

[0067] By calculating L(t), the compensation value of the end position of the robot arm can be obtained, so that the robot arm moves according to the compensated trajectory, thereby achieving stable force control.

[0068] (3) Optimization of impedance parameters and learning system based on MDPSO algorithm,

[0069] In impedance control, appropriate impedance parameters m, b, and k can achieve better control effects. In order to ensure the stability of the impedance control system, the learning coefficient k in the iterative learning ILC is also required. p , k p , k d This embodiment improves the PSO algorithm and proposes an MDPSO algorithm for the impedance parameters m, b, k, and the learning coefficient k p , k p , k d Make optimizations.

[0070] First, generate the initial particle swarm; the MDPSO algorithm randomly generates the initial particle swarm, and each particle The first three parameters correspond to three impedance parameters, and the last three parameters correspond to three learning coefficients in iterative learning control ILC. N is the number of particles in the initial population.

[0071] Secondly, establish the fitness function; in the impedance control of industrial robots, in addition to ensuring that the contact force reaches the expected value in a stable state, the position error of the robot end effector must also be taken into account. Comprehensively considering the force control requirements and position control requirements, the direction with the smallest sum of the contact force error and the position error is selected for particle optimization, so the fitness function e f is the contact force error, e x is the position error, T max is the running time.

[0072] Then, the particle update rate in the MDPSO algorithm is determined; all particles in the PSO algorithm are updated according to the current individual extreme value P best and the current global optimal value G best Update your speed and location is the number of particles, k is the current number of iterations, is the current velocity of the particle, is the original speed, w is the inertia weight, is the current position of the particle, that is, the specific values ​​of all parameters, is the original position, c1 is a single learning factor, c2 is a global learning factor, and R1 and R2 are random values ​​in the interval [0,1].

[0073] This embodiment adds four additional indicators based on the update rate of PSO and proposes a multi-directional particle swarm optimization algorithm MDPSO; the MDPSO algorithm introduces the optimal individuals of four indicators as multi-directional optimization information, namely, the optimal individual P for response speed, v , the optimal individual P for steady-state error ss , optimal individual overshoot P os and the oscillating optimal individual P o .

[0074] Among them, the individual with the best response speed P v , v is the response speed. The smaller its value is, the faster the speed is, so as to determine the individual P with the best response speed. v .

[0075] The optimal individual P for steady-state error ss , e ss =f e (Td max ), e ss is the steady-state error, f e (T max ) is the contact force at the maximum operating time, f d is the expected contact force, e ss The smallest one is the optimal individual P of the steady-state error ss.

[0076] The optimal individual overshoot P os , overshoot e os =f d e max , e os is the overshoot, f emax is the maximum value of the actual contact force, e ss The smallest individual is the individual with the best overshoot.

[0077] Oscillation optimal individual P o , satisfying [f d -f e (tT)][f d -f e (t+T)]<0 is an oscillation, and the individual with the least oscillation times is the optimal oscillation individual P o .

[0078] After obtaining the four best individuals, calculate the total error between the current individual and these four individuals The particle update rate is

[0079] Next, the inertia weight changes dynamically; the larger the inertia weight, the more conducive it is to global search capabilities, so that the algorithm will not fall into the local optimum; the smaller the inertia weight, the more conducive it is to local search, so that the algorithm can quickly converge to the optimal solution. In order to strike a balance between search speed and accuracy, this embodiment dynamically changes the inertia weight. [a,b] is the variation interval, It and It max Represent the current number of iterations and the maximum number of iterations respectively; w i is the initial value of the inertia weight. To avoid the inertia weight being too small, w m Set to the minimum value of the inertia weight.

[0080] Finally, the particle mutation operation, this embodiment uses the mutation operation to enhance the global search capability; after each iteration, according to Select a particle for mutation operation, and update the rate when the calculated γ is less than the set value Perform mutation operation.

[0081]

[0082] in, is the update rate after the mutation operation, λ is the coefficient of variation, ζ is a random number between 0 and 1, which is used to determine the mutation direction, Ub is the upper limit of mutation, and Lb is the lower limit of mutation.

[0083] The above description is only a preferred implementation manner of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.

Claims

1. A robot contact force stabilization control method based on the influence of environmental position and stiffness, characterized in that: The steps include: Step 1: Collect the contact force between the robot execution end and the processing target and predict the environmental stiffness through the trained extreme learning machine; The extreme learning machine is trained by the following steps: first, constructing an initial extreme learning machine model, the extreme learning machine model comprising an input layer, a hidden layer with q nodes, and an output layer; Secondly, construct a training sample set, control the robot movement so that its execution end contacts the processing target, and obtain the contact force in a single direction. and displacement data d i , calculate the environmental stiffness under the current contact force through two adjacent sets of data Construct the training sample set as {f ei ,k ei |i=1,2,…,n}; Finally, the constructed training sample set is used to train the initial extreme learning machine model to generate a trained extreme learning machine; Step 2: Utilize estimated environment position Get the estimated ideal trajectory of the robot arm k e is the environmental stiffness, is the estimated environmental position, f d is the desired contact force; Calculate the error between the ideal trajectory and the estimated ideal trajectory The error between the ideal trajectory and the estimated ideal trajectory is brought into the impedance model to obtain the environmental position estimation error. m, b, and k are impedance parameters; Establishing a robotic arm model based on iterative learning control: Where k is the number of iterations, the output of the k+1th iteration is equal to the output of the kth iteration plus the error compensation term, k p , k i , k d is the learning coefficient, τ is the integral at a certain time between 0 and t, e f is the force error, is the differential of the force error; The sampling period and the current number of iterations are set, and the robot model is discretized to obtain the final environmental position estimation error; the final environmental position estimation error is used to compensate the end position of the robot, so that the robot moves according to the compensated trajectory to achieve stable force control; Step 3: Use the MDPSO algorithm to calculate the impedance parameters m, b, k, and the learning coefficient k p , k i , k d Optimize; specifically include the following steps: First, the initial particle swarm is randomly generated, and each particle N is the number of particles in the initial population; Secondly, establish the fitness function; fitness function e f is the contact force error, e x is the position error, T max is the running time; Then, using the current individual extreme value P best , the current global optimal value G best , the individual with the best response speed P v , the optimal individual P for steady-state error ss , the optimal individual P for overshoot os and the oscillating optimal individual P o Update particles and obtain the particle update rate; Next, the inertia weight is dynamically changed. [a,b] is the variation interval, It and It max Represent the current number of iterations and the maximum number of iterations respectively; w i is the initial value of the inertia weight; Finally, mutation operation is used to enhance the global search capability; After each iteration, according to γ ​​= Select a particle for mutation operation, and update the rate when the calculated γ is less than the set value Perform mutation operations; in, is the update rate after the mutation operation, λ is the coefficient of variation, ζ is a random number between 0 and 1, which is used to determine the mutation direction, Ub is the upper limit of mutation, and Lb is the lower limit of mutation.

2. The robot contact force stabilization control method based on environmental position and stiffness influence according to claim 1 is characterized in that: Determination of the number of nodes q in the hidden layer of the extreme learning machine: Collecting the test contact force test set F t And calculate the corresponding environmental stiffness K t , the contact force data set F t Input into the trained extreme learning machine to predict the environmental stiffness Calculate the average prediction error of the extreme learning machine m is the number of data in the test set, is the environmental stiffness K t The i-th value in The environmental stiffness The i-th value in ; The variation range of q is set to 10% to 100% of the training sample set. For the average error value of the extreme learning machine corresponding to each q, the q corresponding to the optimal average error value is selected as the number of hidden layer nodes of the ELM.

3. The robot contact force stabilization control method based on environmental position and stiffness influence according to claim 1 is characterized in that: The update rate of the particles obtained in step 3 specifically includes: First, according to the current individual extreme value P best and the current global optimal value G best Update speed and location i=1,2…N is the number of particles, k is the current iteration number, is the current velocity of the particle, is the original speed, w is the inertia weight, is the current position of the particle, is the original position, c1 is a single learning factor, c2 is a global learning factor, R1 and R2 are random values ​​in the interval [0,1]; Then, the optimal individual P in response speed is obtained. v , the optimal individual P for steady-state error ss , the optimal individual P for overshoot os and the oscillating optimal individual P o ; Finally, calculate the current individual and the individual with the best response speed P v , the optimal individual P for steady-state error ss , the optimal individual P for overshoot os and the oscillating optimal individual P o The total error The particle update rate is 4. The robot contact force stabilization control method based on environmental position and stiffness influence according to claim 3 is characterized in that: The optimal response speed individual P v It refers to the fastest response speed of force error within a sampling period. The smaller the response speed, the faster the speed, thus determining the individual with the best response speed P v ; T is the sampling period, f e is the contact force; The optimal individual P for steady-state error ss It refers to the optimal individual of steady-state error within a sampling period. The steady-state error e ss =f e (T max ), f e (T max ) is the contact force at the maximum operating time, e ss The smallest is the optimal individual P of the steady-state error ss ; The optimal individual overshoot P os It refers to the optimal individual overshoot in a sampling period, the overshoot e os =f emax -f d , f emax is the maximum value of the actual contact force, f d is the expected contact force, e os The smallest individual is the individual with the best overshoot; The oscillating optimal individual P o It means the minimum number of oscillations in a sampling period, satisfying [f d -f e (tT)][f d -f e (t+T)]<0 is an oscillation, and the individual with the least oscillation times is the optimal oscillation individual P o .

Images