Force impedance control method combining feedforward compensation and variable damping modeling

Abstract

The present invention discloses a force impedance control method integrating feedforward compensation and variable damping modeling. Under the framework of position impedance control, adaptive feedforward PI compensation and deep neural network variable damping model are integrated, wherein the adaptive PI control effectively compensates for reference position information through feedforward control and reduces the force steady-state error; the deep neural network variable damping model adjusts the damping coefficient online in real time to dynamically adapt to the unknown contact environment and realize effective control of the contact force.

Description

Technical Field

[0001] The present invention relates to the field of robotics, and in particular to a force impedance control method integrating feedforward compensation with variable damping modeling. Background Art

[0002] Robot technology has entered the intelligent stage and occupies an important position in the fields of manufacturing, transportation, and service industries. When robots perform operations such as grinding, deburring, clamping objects, precision assembly, and human-machine interaction, it is often necessary to control the contact force within a certain expected range. However, due to different tasks, stiffness and position change with environmental information, etc., it is more difficult to maintain good tracking of contact force and control dynamic response indicators. Since most industrial robots have high position accuracy, the force control of robots in contact with unknown environments relies on position control. The existing methods are: 1) Directly predict the reference trajectory or estimate environmental information. Some literature obtains the reference trajectory and external force estimation formula from the necessary conditions of the force steady-state error and the environmental model, and uses Lyapunov stability to estimate environmental information in real time. Some literature constructs the reference position adaptive law from the difference between the current position and the reference position, and introduces PID (Proportion-Integral-Derivative) into impedance control to form a new impedance relationship. Some literature introduces the idea of ​​fractional order into the adaptive impedance control framework, showing flexibility to the contact environment. Some literatures construct an impedance relationship with adaptive damping coefficient from the two aspects of unknown environmental stiffness and contact position, and prove the stability of this method. 2) Use intelligent control technology to adjust the impedance coefficient. Some literatures introduce fuzzy control on the basis of adaptive impedance control, adjust the impedance coefficient in real time according to force measurement information, and improve the dynamic response capability of the system. Some literatures propose a fuzzy adaptive hybrid impedance control strategy to compensate for the force control error caused by the real-time change of the stiffness of the contact environment. Some literatures use neural networks to compensate for the reference trajectory, so that the final system effect shows impedance characteristics. Some literatures introduce wavelet neural networks to adjust the damping terms online, compensate for the unknown terms of the impedance relationship, and improve the robustness of the control system. Although existing studies have improved the force control performance from the perspectives of predicting reference trajectories, estimating environmental information, and adjusting the impedance coefficient to compensate for system uncertainty, they have not comprehensively considered the steady-state performance and dynamic characteristics of the robot when in contact with the environment, which directly affects the further improvement of the force control performance. Summary of the invention

[0003] The present invention aims to solve the problem that the existing robot force impedance control cannot effectively adapt to the dynamic changes of the random and unknown contact environment, thereby limiting the further improvement of the static and dynamic characteristics of the industrial robot force control, and provides a force impedance control method that integrates feedforward compensation and variable damping modeling.

[0004] To solve the above problems, the present invention is achieved through the following technical solutions:

[0005] The force impedance control method integrating feedforward compensation and variable damping modeling includes the following steps:

[0006] Step 1: Collect the robot's contact force F through the force sensor e , and according to the expected force F r Contact force F with the collection e The force deviation ΔF is obtained, where ΔF = F r -F e ;

[0007] Step 2: Construct a deep neural network variable damping model. The deep neural network variable damping model consists of an input layer, a hidden layer, a convolutional layer, and an output layer. The input of the deep neural network variable damping model is the absolute value of the force deviation |ΔF|, and the output is the damping coefficient B. d ;

[0008] Step 3: Send the absolute value of the force deviation |ΔF| to the deep neural network variable damping model constructed in step 2 to learn the network parameters and obtain the damping coefficient B d ;

[0009] Step 4: Calculate the position error Δx using the impedance model of the robot and the environment; the impedance model of the robot and the environment is:

[0010]

[0011] Step 5: At the initial reference position x r0 Based on the above, the current reference position x is compensated online using the feedforward PI r ,Right now:

[0012] x r =x r0 +k p ΔF+k i ∫ΔFdt

[0013] Step 6: Based on the position error Δx obtained in step 4 and the current reference position x obtained in step 5 r Get the robot's position control value x c , where x c =x r +Δx;

[0014] Step 7: Collect the actual position x of the robot through the position sensor, and adjust the position control value x obtained in step 6. c Deviation x from the actual position x collected c-x After PID control, the control torque of the robot is obtained to control the position of the robot to achieve the purpose of control force;

[0015] Where ΔF represents the force deviation; B d Represents the damping coefficient, M d Indicates the quality coefficient, K d represents the stiffness coefficient; Δx represents the position error, and They represent the first and second differentials of the position error Δx respectively; x r Indicates the current reference position, x r0 represents the initial reference position, k p represents the scale parameter, k i represents the integration parameter.

[0016] The deep neural network variable damping model constructed in step 2 above is:

[0017] The input layer input is:

[0018] |ΔF|

[0019] The activation function of the hidden layer is:

[0020]

[0021] The convolution result of the convolution layer is:

[0022]

[0023] The output of the output layer is:

[0024]

[0025] Where |ΔF| represents the absolute value of the force deviation; h(i) represents the activation function of the i-th neuron in the hidden layer, c i and b i Respectively represent the height and width of the activation function of the i-th neuron in the hidden layer; a j represents the convolution result of the jth neuron in the convolution layer, B j Represents the bias of the jth neuron in the convolutional layer; represents the shared weight vector of the hidden layer and the convolutional layer, h(j) represents the activation function of the jth neuron in the hidden layer, h(j+1) represents the activation function of the j+1th neuron in the hidden layer, and h(j+2) represents the activation function of the j+2th neuron in the hidden layer; w j represents the weight from the jth neuron in the convolutional layer to the output layer, B drepresents the damping coefficient; i = 1, 2, ..., I, I is the number of neurons in the hidden layer; j = 1, 2, ..., J, J is the number of neurons in the convolutional layer.

[0026] As an improvement, in step 5, the scale parameter k p is a fixed value; the integral parameter k i is an adaptive value that has a nonlinear relationship with the force deviation ΔF, where the integral parameter k i for:

[0027]

[0028] Where ΔF represents the force deviation, h 1 It represents the height of the nonlinear function, and w represents the width of the nonlinear function.

[0029] Compared with the prior art, the present invention has the following characteristics:

[0030] 1. The introduced deep neural network variable damping model. In the deep neural network variable damping model, according to the force error ΔF and the damping coefficient B d Mechanism characteristics, design the activation function of the hidden layer, and at the same time, the convolution layer uses shared weights and local connections to extract useful features of the data. Compared with other neural networks, it has fewer weight parameters and can better extract the data features of the hidden layer, speed up the update of parameters, and reduce the complexity of the model.

[0031] 2. The feedforward control of PI structure is introduced to compensate the reference position online, and a nonlinear function with force deviation as the independent variable is designed to change the integral intensity, so as to realize the adaptive adjustment of the integral in the feedforward PI compensation and achieve the purpose of dynamically correcting the reference position information.

[0032] 3. The online acquisition of variable damping coefficient and adaptive feedforward PI compensation improve the static and dynamic tracking performance of force control, reduce the overshoot of force control, and realize the compliant control of the robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the robot force impedance control system of the present invention.

[0034] Figure 2 It is the activation function curve of the hidden layer.

[0035] Figure 3 This is the structural diagram of the deep neural network variable damping model.

[0036] Figure 4 It is a nonlinear function curve graph. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific examples.

[0038] Position-based force tracking impedance control is to equate the mass-damper-spring model to the contact force model between the robot end and the environment. When there is a force deviation, there is a corresponding position error at the robot end. The relationship between these two errors is adjusted by the impedance coefficient of the model, thereby indirectly and flexibly controlling the contact force.

[0039] The impedance model between the robot and the environment is:

[0040]

[0041] Where Δx is the position error, Δx = x c -x r , x c is the robot position control quantity, x r is the reference position corresponding to the desired force, and are the first and second differentials of the position error Δx; M d is the known mass coefficient, B d is the damping coefficient, K d is the known stiffness coefficient; F r is the expected force, F e The contact force is obtained through the force sensor.

[0042] Force sensor description:

[0043]

[0044] Where x is the actual position, x e is the environmental position, k e is the environmental stiffness.

[0045] 1) Steady-state performance

[0046] Force deviation ΔF = F r -F e , it is known from the existing technology that when the system is stable, the force deviation ΔF corresponds to the steady-state error e fss for:

[0047]

[0048] In order to achieve a steady-state error of 0, the following conditions must be met:

[0049]

[0050] Formula (4) Explanation: 1) The premise of precise force control is to obtain accurate reference position x in real time through environmental informationr , or the stiffness coefficient K d =0. When K d = 0, impedance control degenerates into damping control; 2) When the reference position x r If the contact environment changes and no adaptive adjustment is made, the system will produce a large force deviation.

[0051] In practice, the information of the contact environment cannot be accurately known, that is, the stiffness and contact surface position are unknown and dynamically changing, which makes x r It is changing all the time, get accurate x in advance r Very difficult.

[0052] 2) Dynamic performance

[0053] The impedance model of the robot and the environment shows its damping ratio for:

[0054]

[0055] Natural frequency ω n for:

[0056]

[0057] From the factors affecting the damping ratio and natural frequency, it can be seen that when the contact environment changes, the environmental stiffness k e Real-time changes, resulting in the damping ratio and the natural frequency ω n changes with it, directly affecting the dynamic performance of the system. In addition, because the environmental stiffness k e and stiffness coefficient K d It is determined by the characteristics of the system itself and cannot be adjusted manually. Therefore, the target damping coefficient B is adjusted in real time. d , which can achieve the purpose of indirectly adjusting the damping ratio and reducing the impact of environmental changes on the dynamic performance of the system.

[0058] In the framework of position impedance control, the force impedance control method proposed in the present invention integrates feedforward compensation and variable damping modeling, such as Figure 1 As shown, its setting value is obtained through adaptive feedforward compensation, and the position error Δx is obtained by the deep neural network variable damping model to obtain the damping ratio B d Substitute into formula (1) and solve formula (1) to obtain.

[0059] (1) Deep Neural Network Variable Damping Model

[0060] In the robot force impedance control, the force deviation ΔF and the damping coefficient B d The mechanism characteristics are as follows: 1) When the force deviation ΔF is large, the expected damping coefficient B dbecomes smaller in order to achieve a quick response; 2) When the force deviation ΔF is small, the damping coefficient B is expected to d Increase and reduce the system overshoot in order to maintain stability. To this end, according to the force deviation ΔF and the damping coefficient B d Mechanism characteristics, build a deep neural network to describe its characteristics, and the activation function of the hidden layer is designed as:

[0061]

[0062] Where h(i) is the activation function of the i-th neuron in the hidden layer, which is used to represent the force deviation ΔF and the damping coefficient B d The nonlinear mechanism relationship of h(i) is as follows Figure 2 As shown; c i and b i are the height and width of the activation function of the ith neuron in the hidden layer; ΔF is the force deviation, ΔF = F r -F e , F r is the expected force, F e It is the contact force.

[0063] Considering the complexity of the network structure and the amount of computation, the deep neural network variable damping model designed in the present invention consists of an input layer, a hidden layer, a convolutional layer and an output layer. Figure 3 As shown. The hidden layer describes the force deviation ΔF and the damping coefficient B through the activation function h(i) d The convolutional layer can use shared weights and local connections to extract useful features of the data. It has fewer weight parameters, which is conducive to the rapid learning of model weighting. Figure 3 middle, is the neural network input variable, i.e., force deviation ΔF; c i is the height of the activation function of the hidden layer, b i is the activation function width of the hidden layer, h(i) is the nonlinear activation function of the i-th neuron in the hidden layer; is the shared weight vector between the hidden layer and the convolutional layer; z j is the activation function value of the jth neuron in the convolutional layer, a j is the convolution result of the jth neuron in the convolutional layer; w j is the weight from the jth neuron in the convolutional layer to the output layer; is the neural network output, i.e. the damping coefficient B d i = 1, 2, ..., I, I is the number of neurons in the hidden layer. j = 1, 2, ..., J, J is the number of neurons in the convolutional layer.

[0064] The network optimization process consists of two parts: forward propagation and parameter learning. Forward propagation calculates the output value, and parameter learning uses the gradient descent method to update the relevant parameters to approach the target value.

[0065] (1.1) Forward propagation of the network

[0066] 1) Input layer:

[0067]

[0068] 2) Hidden layer:

[0069]

[0070] 3) Convolutional layer:

[0071]

[0072]

[0073] In the formula, B j is the bias of the jth neuron in the convolutional layer.

[0074] 4) Output layer:

[0075]

[0076] (1.2) Network parameter learning

[0077] Assume the performance index function is:

[0078]

[0079] According to the steepest descent method, the update increment formula of the weight coefficient of each layer of the neural network is obtained.

[0080] 1) Weight increment from output layer to convolution layer:

[0081]

[0082] 2) Shared weight increment from hidden layer to convolutional layer:

[0083]

[0084] Where:

[0085] 3) Convolutional layer bias increment:

[0086]

[0087] 4) Hidden layer activation function width parameter increment:

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] Where:

[0095] 5) Hidden layer activation function height parameter increment:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] Where:

[0103] In the neural network iteration process, the weight coefficient update formula is as follows

[0104] w j (k) = w j (k-1)+Δw j +α[w j (k-1)-w j (k-2)] (29)

[0105]

[0106] B j (k) = B j (k-1)+ΔB j +α[B j (k-1)-B j (k-2)] (31)

[0107] b i (k) = b i (k-1)+Δb i +α[b i (k-1)-b i (k-2)] (32)

[0108] c i (k) = c i (k-1)+Δc i +α[c i (k-1)-c i (k-2)] (33)

[0109] Where: w j (k) is the weight from the jth neuron in the hidden layer to the output layer at the current k moment, is the mth shared weight of the convolutional layer at the current k moment, B j (k) is the bias of the jth neuron in the convolutional layer at the current k moment, b i (k) is the activation function width of the i-th neuron in the hidden layer at the current k moment, c i (k) is the activation function height of the i-th neuron in the hidden layer at the current k moment, η is the learning rate, α is the momentum factor. k-1 is the moment before the current k moment.

[0110] (2) Adaptive feedforward PI compensation

[0111] At the initial reference position x r0 Feedforward PI is introduced to compensate the reference position x online r , directly adjust the reference position through the force deviation to make the force steady-state error approach 0. The feedforward PI compensation reference position is:

[0112] x r =x r0 +k p (F r -F e )+k i ∫(F r -F e )dt (34)

[0113] In the formula, k p is the scale parameter, k i is the integration parameter.

[0114] Considering that the changing contact environment will make the system contact process fluctuate, the fixed proportional parameter k p and the integration parameter k i Therefore, the present invention uses the nonlinear function f(ΔF) as the integral parameter k i , so that the integral parameter k iThe speed is dynamically adjusted according to the force tracking during the contact process. When the force error is small, the integral effect is increased, and when the force error is large, the integral effect is weakened, in order to better adapt to the changing contact environment and improve the stability of the control system. The adjustment rule of the nonlinear function f(ΔF) is: when the force deviation is large, f(ΔF) should be reduced to prevent the system from being saturated and causing large overshoot; when the force deviation is small, f(ΔF) should be increased to eliminate the steady-state error, such as Figure 4 As shown. Figure 4 The target curve in , the design force deviation is the independent variable, and f(ΔF) is a nonlinear function of the dependent variable:

[0115]

[0116] In the formula, h 1 represents the height of the set nonlinear function, and w represents the width of the set nonlinear function. When the force deviation becomes smaller, f(ΔF) gradually approaches 1, and the integral effect becomes stronger; conversely, when f(ΔF) gradually approaches 0, the integral effect becomes weaker, achieving dynamic adjustment k i The purpose of speed is to better reduce the steady-state error of force.

[0117] At this time, the impedance control of the adaptive feedforward PI compensation reference position is:

[0118] x r =x r0 +k p ΔF+f(ΔF)∫ΔFdt (36)

[0119] Based on the above analysis, the present invention proposes a force impedance control method integrating feedforward compensation and variable damping modeling, which includes the following steps:

[0120] Step 1: Collect the robot's contact force F through the force sensor e , and according to the expected force F r Contact force F with the collection e The force deviation ΔF is obtained, where ΔF = F r -F e ;

[0121] Step 2: Construct a deep neural network variable damping model. The deep neural network variable damping model consists of an input layer, a hidden layer, a convolutional layer, and an output layer. The input of the deep neural network variable damping model is the absolute value of the force deviation |ΔF|, and the output is the damping coefficient B. d ;

[0122] The input layer input is:

[0123] |ΔF|

[0124] The activation function of the hidden layer is:

[0125]

[0126] The convolution result of the convolution layer is:

[0127]

[0128] The output of the output layer is:

[0129]

[0130] Step 3: Send the absolute value of the force deviation |ΔF| to the deep neural network variable damping model constructed in step 2 to learn the network parameters and obtain the damping coefficient B d ;

[0131] Step 4: Calculate the position error Δx using the impedance model of the robot and the environment; the impedance model of the robot and the environment is:

[0132]

[0133] Step 5: At the initial reference position x r0 Based on the adaptive feedforward PI, the current reference position x is compensated online. r ,Right now:

[0134] x r =x r0 +k p ΔF+k i ∫ΔFdt

[0135] Step 6: Based on the position error Δx obtained in step 4 and the current reference position x obtained in step 5 r Get the robot's position control value x c , where x c =x r +Δx;

[0136] Step 7: Collect the actual position x of the robot through the position sensor, and adjust the position control value x obtained in step 6. c Deviation x from the actual position x collected c -x After PID control, the control torque of the robot is obtained to control the position of the robot to achieve the purpose of control force;

[0137] Where |ΔF| represents the absolute value of the force deviation; h(i) represents the activation function of the i-th neuron in the hidden layer, c i and b i Respectively represent the height and width of the activation function of the i-th neuron in the hidden layer; a j represents the convolution result of the jth neuron in the convolution layer, B jRepresents the bias of the jth neuron in the convolutional layer; represents the shared weight vector of the hidden layer and the convolutional layer, h(j) represents the activation function of the jth neuron in the hidden layer, h(j+1) represents the activation function of the j+1th neuron in the hidden layer, and h(j+2) represents the activation function of the j+2th neuron in the hidden layer; w j represents the weight from the jth neuron in the convolutional layer to the output layer, B d represents the damping coefficient; i = 1, 2, ..., I, I is the number of neurons in the hidden layer; j = 1, 2, ..., J, J is the number of neurons in the convolutional layer; ΔF represents the force deviation; M d Indicates the quality coefficient, K d represents the stiffness coefficient; Δx represents the position error, and They represent the first and second differentials of the position error Δx respectively; x r Indicates the current reference position, x r0 represents the initial reference position, k p Indicates the set proportional parameter, which is a set fixed value. i It represents the set integral parameter, which can be a set fixed value or an adaptive value that has a nonlinear relationship with the force deviation ΔF, that is, h 1 represents the height of the set nonlinear function, and w represents the width of the set nonlinear function.

[0138] The present invention is an industrial robot force impedance control method based on the fusion of adaptive feedforward PI compensation and deep neural network variable damping model under the framework of position impedance control. The adaptive PI control effectively compensates the reference position information through feedforward control and reduces the force steady-state error; the deep neural network variable damping model adjusts the damping coefficient online in real time to dynamically adapt to the unknown contact environment and realize effective control of the contact force.

[0139] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A force impedance control method combining feedforward compensation with variable damping modeling, characterized in that: The steps include: Step 1: Collect the robot's contact force F through the force sensor e , and according to the expected force F r Contact force F with the collection e The force deviation ΔF is obtained, where ΔF = F r -F e ; Step 2: Construct a deep neural network variable damping model. The deep neural network variable damping model consists of an input layer, a hidden layer, a convolutional layer, and an output layer. The input of the deep neural network variable damping model is the absolute value of the force deviation |ΔF|, and the output is the damping coefficient B. d ; The deep neural network variable damping model constructed above is: The input layer input is: |ΔF| The activation function of the hidden layer is: The convolution result of the convolution layer is: The output of the output layer is: Step 3: Send the absolute value of the force deviation |ΔF| to the deep neural network variable damping model constructed in step 2 to learn the network parameters and obtain the damping coefficient B d ; Step 4: Calculate the position error Δx using the impedance model of the robot and the environment; the impedance model of the robot and the environment is: Step 5: At the initial reference position x r0 Based on the above, the current reference position x is compensated online using the feedforward PI r ,Right now: x r =x r0 +k p ΔF+k i ∫ΔFdt Step 6: Based on the position error Δx obtained in step 4 and the current reference position x obtained in step 5 r Get the robot's position control value x c , where x c =x r +Δx; Step 7: Collect the actual position x of the robot through the position sensor, and adjust the position control value x obtained in step 6. c After PID control of the deviation from the actual position x collected, the control torque of the robot is obtained to control the position of the robot to achieve the purpose of control force; Where |ΔF| represents the absolute value of the force deviation; h(i) represents the activation function of the i-th neuron in the hidden layer, c i and b i Respectively represent the height and width of the activation function of the i-th neuron in the hidden layer; a j represents the convolution result of the jth neuron in the convolution layer, B j Represents the bias of the jth neuron in the convolutional layer; represents the shared weight vector of the hidden layer and the convolutional layer, h(j) represents the activation function of the jth neuron in the hidden layer, h(j+1) represents the activation function of the j+1th neuron in the hidden layer, and h(j+2) represents the activation function of the j+2th neuron in the hidden layer; w j represents the weight from the jth neuron in the convolutional layer to the output layer, B d represents the damping coefficient; i = 1, 2, ..., I, I is the number of neurons in the hidden layer; j = 1, 2, ..., J, J is the number of neurons in the convolutional layer; ΔF represents the force deviation; M d Indicates the quality coefficient, K d represents the stiffness coefficient; Δx represents the position error, and They represent the first and second differentials of the position error Δx respectively; x r Indicates the current reference position, x r0 represents the initial reference position, k p represents the scale parameter, k i Represents the integration parameter.

2. The force impedance control method integrating feedforward compensation and variable damping modeling according to claim 1 is characterized in that: In step 5, the scale parameter k p is a fixed value; the integral parameter k i It is an adaptive value that has a nonlinear relationship with the force deviation ΔF.

3. The force impedance control method integrating feedforward compensation and variable damping modeling according to claim 2 is characterized in that: Integration parameter k i for: Wherein, ΔF represents the force deviation, h1 represents the set nonlinear function height, and w represents the set nonlinear function width.

Images