Milling robot low dynamic stiffness pose prediction method based on GPR
Through the GPR-based method, the frequency response function is fitted by the force hammer excitation test and the Levy method, and the prediction model of the opposite sex rational quadratic core is established, which solves the problem of predicting the weak dynamic stiffness posture of the milling robot, and realizes the designation of the milling robot Accurate prediction at rotation speed improves machining stability and workpiece surface quality.
Patent Information
- Application Number
- CN202411095151.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-10
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-08-10
AI Technical Summary
The prior art is difficult to accurately predict the weak dynamic stiffness posture of milling robots in high load processing, resulting in flutter instability problems, affecting the surface quality of the workpiece and causing economic losses.
Using GPR-based method, the frequency response data of the milling robot is obtained through the force hammer excitation test, and the Levy method is used to construct a rational polynomial fit frequency response function, establish a GPR prediction model of the opposite sex rational quadratic core, weigh the influence of each subsystem, and quickly obtain milling The main modal natural frequencies of the robot under different postures achieve accurate prediction of weak dynamic stiffness postures.
This method can more accurately consider the complex structure and material parameters of the robot and assembly area, comprehensively analyze the impact of each subsystem on processing vibration, reduce model complexity, and improve processing stability and workpiece surface quality.
Smart Images

Figure CN119036440A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of robot milling processing, and more specifically, relates to a method for predicting weak dynamic stiffness posture of a milling robot based on GPR (Gaussian regression process). Background Art
[0002] In recent years, articulated industrial robots have been widely used in aerospace, automotive and other manufacturing fields, with the advantages of low cost and flexible processing mode. However, in high-load milling tasks, the weak stiffness posture of the robot can easily cause chatter instability in robot milling, significantly reduce the surface quality of the workpiece, and even cause high-value-added parts to be scrapped, resulting in huge economic losses. Therefore, avoiding the posture of the milling robot with weak end stiffness is an important research topic in the field of robot milling. Stiffness can be divided into static stiffness and dynamic stiffness. Dynamic stiffness can reflect the stiffness performance of the milling robot under dynamic excitation of processing, and is a key factor that directly affects milling stability.
[0003] At present, the dynamic stiffness of robots is mainly derived based on multi-body dynamics analysis. For example, Pan et al. in the United States modeled the chatter process by abstracting the robot milling system as a spring-mass-damper system, thereby studying the mechanism of stability and chatter suppression during processing; Baglioni et al. in Italy established a dynamic model of a milling robot with rigid links and flexible joints based on the standard DH parameter method, determined the mass matrix by calculating the moment of inertia of the model, and determined the joint elastic parameters by the technical specifications of the harmonic drive; Mousavi et al. in France established a dynamic model of the robot by considering the influence of the stiffness and damping of each joint through the Euler-Bernoulli unit. However, the above studies all consider the mechanism model. Due to the complex structure of the robot and the assembly area, it is difficult to fully consider its geometric model, material parameters and internal assembly method in the mechanism model, so there is a large deviation between the calculation and the true value; at the same time, the mechanism model is mainly a single-order model of the robot body or the end tool, and often ignores the influence of the actual assembly of the milling system on the vibration of the end of the processing. It is difficult to weigh the influence of each subsystem in the simplified analysis, and the inclusion of all in the analysis will greatly increase the complexity of the model.
[0004] Therefore, the field of robotic milling processing urgently needs an efficient and accurate method for predicting the weak dynamic stiffness posture of milling robots. Summary of the invention
[0005] The present invention provides a GPR-based method for predicting the weak dynamic stiffness posture of a milling robot. According to the variation characteristics of a dynamic stiffness frequency response curve, multi-order modal natural frequencies are used as judgment factors of the weak dynamic stiffness of the milling robot. By carrying out a force hammer excitation test, three-way terminal frequency response data of 100 groups of robot postures are obtained. A rational polynomial fitting frequency response function is constructed based on the Levy method, and the influence of each subsystem on the machining vibration is weighed, so that 8 main modes in the X / Y / Z directions are selected to form a multi-modal natural frequency training set. A GPR prediction model of anisotropic rational quadratic kernels is established based on the test data. The main modal natural frequencies of the milling robot in different postures can be quickly obtained, and the weak dynamic stiffness posture of the milling robot can be accurately predicted under a specified milling spindle speed.
[0006] The technical solution adopted by the present invention comprises the following steps:
[0007] (a) For a six-degree-of-freedom milling robot, constructing the terminal frequency response function matrix can describe the system dynamic stiffness;
[0008] (b) Carry out force hammer excitation test to obtain frequency response data of milling robot;
[0009] (c) The experimental frequency response data are fitted using the multimodal identification method Levy method to construct a rational fraction polynomial of the theoretical frequency response of the N-order modal viscous damping system;
[0010] (d) Establishing a GPR prediction model with anisotropic rational quadratic kernels to predict the weak dynamic stiffness posture of the milling robot at a specified milling spindle speed. The GPR prediction model contains 8 sub-models. Each sub-model input contains three dimensions, namely, the X, Y coordinate values of the processing TCP point and the redundancy A, and the output is the main modal natural frequency Fn;
[0011] (e) The GPR prediction model composed of 8 sub-models can obtain the natural frequency of the milling robot under any posture. For the specified milling processing conditions, the blade passing frequency of the milling force is calculated. At this time, the X, Y coordinate values and redundancy A corresponding to the blade passing frequency close to the 8 main modal natural frequencies are found, which is the weak dynamic stiffness posture of the milling robot at the specified speed.
[0012] In step (a) of the present invention, the tool coordinate system i-1 The frequency response functions in the three directions of XYZ in P represent the dynamic stiffness of the milling robot. The terminal frequency response H in the tool coordinate system tool(ω) can be composed of direct frequency response and cross-coupled frequency response. In fact, for the frequency response of the articulated milling robot in each direction, compared with the contribution of the excitation in this direction to its vibration, the influence of the excitation in the orthogonal direction on the vibration is smaller. Therefore, only the direct frequency response in each direction is considered, and the cross-coupled frequency response is ignored. The terminal frequency response function can be composed of the direct frequency response function H in each direction x (ω), H y (ω), H z (ω) is expressed as:
[0013] H tool (ω)=diag[H x (ω),H y (ω),H z (ω)]
[0014] The natural angular frequency ω of each mode in the X, Y, and Z directions of the articulated milling robot n It is the core judgment factor of whether the system is in a weak dynamic stiffness posture.
[0015] In step (b) of the present invention, a hammer method is used to obtain frequency response data of 100 sets of milling robot postures. The milling robot posture of the test is selected with reference to the TCP relative base coordinate system in the robot end tool coordinate system. In the specified processing plane, points are taken at equal intervals in the X and Y directions within the processing range, and a total of 20 points are selected. At each point, 5 postures are selected at equal intervals within the range of redundancy A. In each set of postures, frequency response data are measured in the X, Y, and Z directions respectively. A total of 300 sets of milling robot end frequency response data are obtained, and the selected postures basically cover the changes of the robot in the end milling processing plane.
[0016] In step (c) of the present invention, the rational fraction polynomial for constructing the theoretical frequency response of the N-order modal viscous damping system is:
[0017]
[0018] Let the theoretical frequency response function H L (ω) and the variance of the test frequency response data are minimized, so as to identify the coefficient vectors of the numerator and denominator of the rational fraction and obtain the natural angular frequency ω of the v-th order mode of the milling robot in one direction nv ,ω nv =2πf nv , for subsequent calculations, to form the natural frequency f nv Training set of data corresponding to the poses of the milling robot.
[0019] In step (d) of the present invention, the kernel function of the GPR model directly determines the model regression effect, and a rational quadratic kernel is used to describe the complex characteristics of the data. At the same time, combined with the automatic correlation determination technology in machine learning, a rational quadratic kernel of various anisotropies in the model is constructed to adapt to the multi-scale and multi-dimensional complex data feature peaks between the natural frequency and posture of the milling robot, which is expressed as:
[0020]
[0021] In the formula, σ 2 is the output scale parameter, i.e., the output spatial variance, M h It is a diagonal matrix of kernel function length scales, and each element on the diagonal is the length scale l of each input dimension X, Y, and A. x , l y , l A The square inverse of , this matrix is used to control the overall smoothness of the kernel function, α is the mixed scale parameter, which determines the sensitivity of the kernel function to multiple scales;
[0022] After setting the kernel function, the prior Gaussian process needs to be trained with the model data set to search for the optimal function hyperparameters. The optimization is performed by minimizing the likelihood negative logarithm function. Assuming that the data has normal noise, the Gaussian likelihood function is used for prediction. Its negative logarithm function can be written as:
[0023]
[0024] In the formula, σ n 2 is the noise variance, which indicates the deviation between the measured data output value and the true value of the data. By solving the partial derivative of this formula through the quasi-Newton method, we can get the optimized function hyperparameters to match the data characteristics between the natural frequency and the robot posture.
[0025] The function settings in a single sub-model are clarified, and the rational quadratic kernel function, constant mean function and Gaussian likelihood function with different anisotropy are selected for construction. The hyperparameter group composed of the function parameters in the v-th order modal sub-model is
[0026]
[0027] In step (d) of the present invention, the redundant angle is adjusted by keeping the point position unchanged at the weak dynamic stiffness position so that the natural frequency of the milling robot is far away from the natural frequency of the main mode, thereby avoiding the weak dynamic stiffness position, effectively reducing the processing vibration, and improving the surface quality of the processed workpiece. The present invention has the following beneficial effects:
[0028] 1. The data-driven GPR prediction model is closer to the real structure of the system, and can take into account the complex geometric models, material parameters and internal assembly methods of the robot and assembly area, avoiding the deviation of the mechanism model.
[0029] 2. The model weighs the influence of each subsystem through test results and includes all major modes in the analysis. Compared with the current simplified analysis of the single-order mode of the robot body or the end tool, it more comprehensively considers the influence of the structure on the vibration of the processing end. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The drawings described herein are used to provide further understanding of the present invention and constitute a part of this application. The illustrative examples of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0031] Figure 1 is a flow chart of the method of the present invention;
[0032] Figure 2 These are the eight main modal diagrams of the end frequency response of the milling robot of the present invention;
[0033] Figure 3 It is a schematic diagram of the GPR weak dynamic stiffness prediction model of the present invention;
[0034] Figure 4 It is a diagram of the prediction of the weak dynamic stiffness model under the redundancy control of the present invention;
[0035] Figure 5 It is a redundancy diagram of the present invention. DETAILED DESCRIPTION
[0036] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0037] The present invention provides a GPR-based weak dynamic stiffness posture prediction method for a milling robot. The method is based on the change characteristics of a dynamic stiffness frequency response curve, takes multi-modal natural frequencies as weak dynamic stiffness prediction factors, constructs a terminal three-way frequency response function through a hammering method, identifies the main modal natural frequencies that play a leading role in terminal vibration to form a training set, and establishes a GPR prediction model with an anisotropic rational quadratic kernel to achieve accurate prediction of the weak dynamic stiffness posture of the milling robot at a specified speed.
[0038] The present invention is described in detail below with reference to the accompanying drawings.
[0039] (a) In order to obtain the dynamic stiffness model of the milling robot, it is necessary to perform dynamic modeling and perform Laplace transform on the dynamic equation to obtain the frequency response function of the single degree of freedom system:
[0040]
[0041] Where k is the stiffness, ω n is the natural angular frequency of the system (distinguished from the natural frequency f n ,ω n =2πf n ), ζ is the damping ratio. For a six-degree-of-freedom milling robot, constructing the terminal frequency response function matrix can describe the system dynamic stiffness. In the tool coordinate system i-1 The frequency response functions in the three directions of XYZ in P represent the dynamic stiffness of the milling robot. The terminal frequency response H in the tool coordinate system tool (ω) can be composed of direct frequency response and cross-coupled frequency response. In fact, for the frequency response of the articulated milling robot in each direction, compared with the contribution of the excitation in this direction to its vibration, the influence of the excitation in the orthogonal direction on the vibration is smaller. Therefore, assuming that only the direct frequency response in each direction is considered and the cross-coupled frequency response is ignored, the terminal frequency response function can be composed of the direct frequency response function in each direction H x (ω), H y (ω), H z (ω) is expressed as:
[0042]
[0043] The single-direction frequency response function reflects the modes of different subsystems of the milling robot. According to the modal superposition theory, the natural angular frequency, damping ratio and stiffness of the v-th order independent mode in a single direction are set as ω nv , v , k v , combined with formula (1), the frequency response function of a single direction under a fixed posture of the robot can be expressed as the superposition of multiple modes:
[0044]
[0045] In the formula, r v is the natural frequency ratio r v =ω / ω nv , v = 1, 2, 3, ..., M;
[0046] In summary, the changes in the modal parameters of each order independent mode of the system under different postures can be obtained through the experimental modal, and the direct frequency response functions H of the robot milling system dependent on the posture can be derived and calculated. x (ω), H y (ω), H z(ω), the end dynamic stiffness frequency response model of the articulated milling robot can be obtained, which is used to measure the dynamic stiffness performance of the robot at different positions during the milling process.
[0047] Analysis of the dynamic stiffness frequency response curve shows that at the natural frequency, the dynamic stiffness will be determined by damping. At this time, the system is in the weak dynamic stiffness area, that is, the ability of the structure to resist deformation is the smallest, and it is easy to be excited by similar vibration frequencies. It can be determined that the natural frequency of each mode in the articulated milling robot is the core factor in determining whether the system is in a weak dynamic stiffness posture;
[0048] (b) The force hammer method is used to obtain the frequency response data of 100 sets of milling robot postures in the processing space. The milling robot posture selection described in the test is based on the tool center point TCP relative to the base coordinate system in the robot end tool coordinate system. In the specified processing plane, points are taken at equal intervals in the X and Y directions within the processing range, and a total of 20 points are selected. At each point, 5 postures are selected at equal intervals within the range with redundancy A. In each set of postures, the frequency response data of the X, Y, and Z directions are measured respectively. A total of 300 sets of milling robot end frequency response data are obtained. The selected postures basically cover the changes of the robot in the vertical milling processing plane;
[0049] (c) The Levy method, a multimodal identification method, was used to fit 300 sets of experimental frequency response data, and the rational fraction polynomial of the theoretical frequency response of the M-order modal viscous damping system was constructed as follows:
[0050]
[0051] In the formula, the frequency response of the M-order modal system corresponds to a 2M-order rational fraction polynomial, a k and b k are the numerator coefficient vector and denominator coefficient vector of the transfer function sorted in ascending power, k = 0, 1, 2, …, 2M;
[0052] Let the theoretical frequency response function H L The sum of the variances of (ω) and the experimental frequency response data is minimized, thereby identifying the numerator and denominator coefficient vectors of the rational polynomial;
[0053] Convert (4) to a transfer function using the Laplace operator, set the denominator of the power polynomial to 0, and obtain M pairs of complex roots, one of which is expressed as:
[0054]
[0055] Furthermore, the natural angular frequency of the v-th order mode of the milling robot in one direction can be obtained as:
[0056]
[0057] Wherein, v = 1, 2, 3, ..., M; k = 0, 1, 2, ..., 2M;
[0058] Obtain the natural angular frequency ω of the vth-order mode of the milling robot in one direction nv (ω nv =2πf nv ), for subsequent calculations, forms the natural frequency f nv A training set of data corresponding to the milling robot poses;
[0059] (d) Establish the GPR prediction model of anisotropic rational quadratic kernels such as Figure 3 , to predict the weak dynamic stiffness posture of the milling robot at a specified milling spindle speed. It has been verified that there are 8 main modes in the end frequency response function of the robot that have a large influence on the end, such as Figure 2 The model uses the natural frequencies of 8 main modes as the model training set, and uses Gaussian process regression to predict the weakest dynamic stiffness area of each main mode of the robot body and milling processing system in the milling robot under different postures, that is, the natural frequency damping control area;
[0060] The natural frequencies of the eight main modes in the experimental modal analysis (called modes XL, XH, YL1, YL2, YH, ZL1, ZL2, and ZH according to their natural frequencies from low to high in the X / Y / Z directions) are used as training sets to establish a GPR prediction model. The model contains a total of eight sub-models. Each sub-model input contains three dimensions, namely, the X and Y coordinate values of the TCP point of the milling robot and the redundancy A, and the output is the natural frequency of the main mode.
[0061] Fn in (in=XL,XH,YL1,YL2,YH,ZL1,ZL2,ZH), the sub-model can be expressed as:
[0062] f(x)~N(μ(x),κ(x,x))(7)
[0063] Where μ(x) is the mean function and κ(x,x) is the covariance function, also known as the kernel function, which can determine a Gaussian process.
[0064] The GPR prediction model function will be set based on the characteristics of the training set data. By analyzing the data results, that is, analyzing the change rules of the natural frequency of the milling robot with the posture, it can be found that the natural frequency has both the local characteristics of subtle changes and the global characteristics of trend changes in a single input dimension. At the same time, the change characteristics of the natural frequency in the three input dimensions are quite different.
[0065] The variation law of the natural frequency of the milling robot with the posture is analyzed, and the rational quadratic kernel (RQ) is introduced to describe the complex characteristics of the data. At the same time, combined with the automatic correlation determination technology in machine learning, the rational quadratic kernel of various anisotropies in the model is constructed to adapt to the complex data of multiple scales and dimensions, which can be expressed as:
[0066]
[0067] In the formula, σ 2 is the output scale parameter, i.e., the output spatial variance, which determines the overall variation of the function. h It is a diagonal matrix of kernel function length scales, and each element on the diagonal is the length scale l of each input dimension X, Y, and A. x , l y , l A The square inverse of is used to control the overall smoothness of the kernel function. α is the mixed scale parameter, which determines the sensitivity of the kernel function to multiple scales.
[0068] Compared with the widely used radial basis function kernel, the anisotropic rational quadratic kernel has higher flexibility, which is equivalent to the weighted sum of infinite radial basis function kernels of different length scales. By adjusting the mixed scale parameter, it can adapt to the changing characteristics at different scales, so that the model can achieve a balance between data overfitting and generalization, taking into account the local and global characteristics of the natural frequency changing with a single posture parameter. At the same time, it ensures that each input dimension has an independent length scale to adjust, so as to capture the changing characteristics of different input dimensions more finely. In summary, the anisotropic rational quadratic kernel is more advantageous in capturing the complex changing relationship between the natural frequency of the robot and the posture.
[0069] After setting the kernel function, the model data set needs to be used to train the prior Gaussian process to search for the optimal function hyperparameters. The optimization is performed by minimizing the likelihood negative logarithm function. Assuming that the data has normal noise, the Gaussian likelihood function is used for prediction, and its negative logarithm function can be written as:
[0070]
[0071] In the formula, is the noise variance, which indicates the deviation between the measured data output value and the true value of the data. By solving the partial derivative of this formula through the quasi-Newton method, the optimized function hyperparameters can be obtained to match the data characteristics between the natural frequency and the robot posture.
[0072] So far, the function settings in a single sub-model have been clarified. The model is constructed using various anisotropic rational quadratic kernel functions, constant mean functions, and Gaussian likelihood functions. The hyperparameter group composed of the function parameters in the v-th order modal sub-model is expressed as:
[0073]
[0074] The initial value is set to β0 = {(0; 0; 0); 0; 0; 0; 1}, and the optimization is performed by minimizing the negative logarithmic marginal likelihood to obtain the optimized model hyperparameter group. The sub-models have large differences in different length scales, the change characteristics of the natural frequency with different dimensions are different, and the different mixed scales indicate that there are multi-scale change characteristics in a single input dimension. It can be determined that it is necessary to select anisotropic rational quadratic kernel functions.
[0075] Furthermore, it is necessary to exclude the overfitting of the GPR prediction model, test the prediction effect and generalization ability of the model, and perform internal verification of the GPR multimodal weak stiffness prediction model based on the holdout method. The optimized hyperparameter group is set as the initial value, the experimental data set is randomly shuffled and divided into 80% training set and 20% test set, and the mean relative error MRE and determination coefficient R are used to calculate the optimal hyperparameter group. 2 As an evaluation indicator of the regression model, it can evaluate the accuracy and fitting precision of the regression model;
[0076] (e) During the vertical milling process, the spindle speed is selected to obtain the theoretical cutting edge frequency of the milling force. The main modal natural frequency is predicted by the model as it changes with the posture. When the main modal natural frequency is close to the excitation frequency, the TCP coordinates and redundancy A are output, and the weak dynamic stiffness posture of the milling robot can be accurately determined.
[0077] At this time, the redundancy angle can be adjusted to make the multi-mode natural frequency away from the excitation frequency, such as Figure 4 , so that the redundant angle of the milling robot avoids the weak dynamic stiffness posture, effectively reducing the end processing vibration and improving the surface quality of the processed workpiece.
[0078] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A GPR-based method for predicting the weak dynamic stiffness posture of a milling robot, characterized in that: The following steps are involved: (a) For a six-degree-of-freedom milling robot, constructing the terminal frequency response function matrix can describe the system dynamic stiffness; (b) Carry out force hammer excitation test to obtain frequency response data of milling robot; (c) The experimental frequency response data are fitted using the multimodal identification method Levy method to construct a rational fraction polynomial of the theoretical frequency response of the N-order modal viscous damping system; (d) Establishing a GPR prediction model with anisotropic rational quadratic kernels to predict the weak dynamic stiffness posture of the milling robot at a specified milling spindle speed. The GPR prediction model contains 8 sub-models. Each sub-model input contains three dimensions, namely, the X, Y coordinate values of the processing TCP point and the redundancy A, and the output is the main modal natural frequency Fn; (e) The GPR prediction model composed of 8 sub-models can obtain the natural frequency of the milling robot under any posture. For the specified milling processing conditions, the blade passing frequency of the milling force is calculated. At this time, the X, Y coordinate values and redundancy A corresponding to the blade passing frequency close to the 8 main modal natural frequencies are found, which is the weak dynamic stiffness posture of the milling robot at the specified speed.
2. The method for predicting weak dynamic stiffness and posture of a milling robot based on GPR according to claim 1, characterized in that: In the step (a), the tool coordinate system i-1 The frequency response functions in the three directions of XYZ in P represent the dynamic stiffness of the milling robot. The terminal frequency response H in the tool coordinate system tool (ω) can be composed of direct frequency response and cross-coupled frequency response. In fact, for the frequency response of the articulated milling robot in each direction, compared with the contribution of the excitation in this direction to its vibration, the influence of the excitation in the orthogonal direction on the vibration is smaller. Therefore, only the direct frequency response in each direction is considered, and the cross-coupled frequency response is ignored. The terminal frequency response function can be composed of the direct frequency response function H in each direction x (ω), H y (ω), H z (ω) is expressed as: H tool (ω)=diag[H x (ω),H y (ω),H z (oh)] The natural angular frequency ω of each mode in the X, Y, and Z directions of the articulated milling robot n It is the core judgment factor of whether the system is in a weak dynamic stiffness posture.
3. The method for predicting weak dynamic stiffness and posture of a milling robot based on GPR according to claim 1, characterized in that: In the step (b), a hammer method is used to obtain frequency response data of 100 sets of milling robot postures. The milling robot posture of the test is selected with reference to the TCP relative base coordinate system in the robot end tool coordinate system. In the specified processing plane, points are taken at equal intervals in the X and Y directions within the processing range, and a total of 20 points are selected. At each point, 5 postures are selected at equal intervals within the range of redundancy A. In each set of postures, frequency response data are measured in the X, Y, and Z directions respectively. A total of 300 sets of milling robot end frequency response data are obtained, and the selected postures basically cover the changes of the robot in the end milling processing plane.
4. The method for predicting weak dynamic stiffness and posture of a milling robot based on GPR according to claim 1, characterized in that: In the step (c), the rational fraction polynomial for constructing the theoretical frequency response of the N-order modal viscous damping system is: Let the theoretical frequency response function H L (ω) and the variance of the test frequency response data are minimized, so as to identify the coefficient vectors of the numerator and denominator of the rational fraction and obtain the natural angular frequency ω of the v-th order mode of the milling robot in one direction nv ,ω nv =2πf nv , for subsequent calculations, to form the natural frequency f nv Training set of data corresponding to the poses of the milling robot.
5. The method for predicting weak dynamic stiffness and posture of a milling robot based on GPR according to claim 1, characterized in that: In the step (d), the kernel function of the GPR model directly determines the model regression effect. A rational quadratic kernel is used to describe the complex characteristics of the data. At the same time, combined with the automatic correlation determination technology in machine learning, a rational quadratic kernel of various anisotropies in the model is constructed to adapt to the multi-scale and multi-dimensional complex data feature peaks between the natural frequency and posture of the milling robot, which is expressed as: In the formula, σ 2 is the output scale parameter, i.e., the output spatial variance, M h It is a diagonal matrix of kernel function length scales, and each element on the diagonal is the length scale l of each input dimension X, Y, and A. x , l y , l A The square inverse of , this matrix is used to control the overall smoothness of the kernel function, α is the mixed scale parameter, which determines the sensitivity of the kernel function to multiple scales; After setting the kernel function, the prior Gaussian process needs to be trained with the model data set to search for the optimal function hyperparameters. The optimization is performed by minimizing the likelihood negative logarithm function. Assuming that the data has normal noise, the Gaussian likelihood function is used for prediction. Its negative logarithm function can be written as: In the formula, σ n 2 is the noise variance, which indicates the deviation between the measured data output value and the true value of the data. By solving the partial derivative of this formula through the quasi-Newton method, we can get the optimized function hyperparameters to match the data characteristics between the natural frequency and the robot posture. The function settings in a single sub-model are clarified, and the rational quadratic kernel function, constant mean function and Gaussian likelihood function with different anisotropy are selected for construction. The hyperparameter group composed of the function parameters in the v-th order modal sub-model is 6. The method for predicting weak dynamic stiffness and posture of a milling robot based on GPR according to claim 1, characterized in that: In the step (d), the redundant angle is adjusted by keeping the point position unchanged at the weak dynamic stiffness posture so that the natural frequency of the milling robot is far away from the natural frequency of the main mode, thereby avoiding the weak dynamic stiffness posture, effectively reducing the processing vibration, and improving the surface quality of the processed workpiece.
Citation Information
Patent Citations
Method for optimizing reliability of milling-cutting machining process parameters
CN107480352A
Binary tree robot milling system frequency response predication method based on RCSA (Risk and Control Self-Assessment)
CN108268745A
Robot tool nose frequency response prediction method and system based on pose dependence characteristics and cross coupling terms
CN112765863A
Incremental self-excitation method for identifying low-frequency frequency-response function of milling robot
CN116861658A
CONTROL SYSTEM USING GAUSSIAN PROCESS REGRESSION
DE102021213112A1
