CF / PEEK milling surface roughness modeling method based on L1-LSSVR algorithm

Abstract

The invention provides a CF / PEEK milling surface roughness modeling method based on an L1-LSSVR algorithm. The method comprises the steps that S1, static milling process parameters serve as experimental variables, an experimental scheme is designed, dynamic signals in the milling process are collected in the experimental process, and after the experiment is finished, the CF / PEEK three-dimensional surface roughness is measured; s2, time domain features of the dynamic signals are extracted, and the time domain features are used for L1-LSSVR model training after being combined with static milling process parameters; s3, analyzing the correlation between the features through a Pearson's correlation coefficient, and eliminating redundant features with similar influence rules on the roughness of the three-dimensional surface; and performing feature selection by using L1 norm regularization, and screening out features having high correlation with the three-dimensional surface roughness. S4, establishing a CF / PEEK milling surface roughness model by taking the features screened in the step S3 as the input of an LSSVR algorithm and the CF / PEEK milling three-dimensional surface roughness as the output; according to the algorithm designed by the invention, modeling of the milling three-dimensional surface roughness of the CF / PEEK material can be effectively realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of production and manufacturing, and particularly relates to a CF / PEEK milling surface roughness modeling method based on an L1-LSSVR algorithm. BACKGROUND

[0002] Carbon fiber reinforced polyether ether ketone (CF / PEEK) material is widely used in the fields of aerospace, medical devices and automobiles due to its excellent mechanical properties, heat resistance, chemical resistance and low density. In the background of lightweight and high performance requirements, CF / PEEK material plays an important role in key components in various fields, for example, the F-22 fighter main landing gear door, Boeing 787 ceiling parts, automobile engine covers and suspension system components all use CF / PEEK material. When processing the material, milling is one of the main methods, and the surface roughness has strict requirements due to long-term service in extreme conditions such as high temperature and high pressure. Due to the high strength and wear resistance of CF / PEEK, adhesion between the tool and the material is easily generated during processing, which leads to a decrease in processing performance. In addition, its low thermal conductivity, anisotropy and non-uniformity further increase the difficulty of milling processing, and these characteristics make it extremely complex to control the surface roughness of CF / PEEK processing.

[0003] Surface roughness modeling plays a key role in evaluating the surface roughness after processing. Milling processing is a complex thermal-mechanical coupling process, and the processing process is not only affected by process parameters, but also affected by dynamic factors such as machine tool power fluctuation, cutting force and tool wear, and each factor affects each other in a nonlinear way. The traditional modeling method based on theoretical physics and experimental analysis is based on physical principles or mathematical formulas to describe the processing process, but the nonlinear characteristics in CF / PEEK milling processing are difficult to accurately describe by mathematical formulas. The machine learning method can bypass the underlying mechanism of machine tool milling processing, learn the processing data, and mine the complex relationship between process parameters and processing results, so as to establish a more accurate and stable model. SUMMARY

[0004] The application provides a CF / PEEK milling surface roughness modeling method based on an L1-LSSVR algorithm, which can learn processing data, mine the complex relationship between process parameters and processing results, and establish a more accurate and stable model.

[0005] The application adopts the following technical scheme.

[0006] A CF / PEEK milling surface roughness modeling method based on an L1-LSSVR algorithm comprises the following steps.

[0007] Step S1: Design the experimental scheme with static milling process parameters as experimental variables, collect dynamic signals during the milling process during the experiment, and measure the three-dimensional surface roughness of CF / PEEK after the experiment;

[0008] Step S2: Extract the time domain features of dynamic signals (electrical signals, cutting force signals), and use them in combination with static milling process parameters for L1-LSSVR model training;

[0009] Step S3: Analyze the correlation between features by Pearson correlation coefficient, eliminate redundant features with similar influence law on three-dimensional surface roughness, and use L1 norm regularization for feature selection to select features with high correlation with three-dimensional surface roughness.

[0010] Step S4: Use the features selected in step S3 as the input of LSSVR algorithm, and the three-dimensional surface roughness of CF / PEEK milling as the output to establish the CF / PEEK milling surface roughness model.

[0011] In step S1, the static milling process parameters include cutting speed, cutting depth and feed per tooth, and the dynamic signals include electrical signals and cutting force signals.

[0012] Step S1 includes the following steps:

[0013] Step S1.1, with cutting speed, cutting depth, feed per tooth and fiber direction angle as experimental factors, set the levels of each factor and design orthogonal experiment, and the constraints of each factor are set as:

[0014] v min ≤v c ≤v max

[0015] f min ≤f z ≤f max

[0016] a pmin ≤a p ≤a pmax

[0017] θ=u1、v1

[0018] Among them,

[0019] v c , f z , a p , θ respectively represent cutting speed, cutting depth, feed per tooth and fiber direction angle;

[0020] v min , v max respectively represent the minimum and maximum values of cutting speed;

[0021] f min 、f max respectively represent the minimum and maximum values of the cutting depth;

[0022] a pmin 、a pmax respectively represent the minimum and maximum values of the feed per tooth;

[0023] u1, v1 are 0°, 90° fiber direction angles;

[0024] Step S1.2, CF / PEEK milling experiments are carried out according to the experimental scheme, and dynamic signal data in the milling process is collected during the experiment;

[0025] Specifically, it includes 18 groups of electrical signal parameters (as shown in Table 1) and three-direction cutting force signals in the milling process (as shown in Table 2). After the experiment, the three-dimensional surface roughness of CF / PEEK after milling is measured and recorded.

[0026] Table 1 Electrical signal parameter table

[0027]

[0028] Table 2 Cutting force signal parameter table

[0029]

[0030]

[0031] Step S2 includes the following steps:

[0032] Step S2.1, before the time domain feature extraction of the dynamic signal, the signal is truncated to extract the time domain feature of each group of dynamic signal, at the beginning of milling, the dynamic signal data fluctuates in a large range, the data in the milling process is cut off from the fluctuation, and the cut-off signal is used as the data for subsequent feature extraction;

[0033] Step S2.2, 11 time domain features of each group of dynamic signal are extracted, 231 (21x11) dynamic characteristics are calculated by extracting 11 indexes including peak-to-peak value, mean value, peak value, standard deviation, skewness, kurtosis, root mean square, waveform factor, margin factor, kurtosis factor and pulse factor,

[0034] The index calculation formula is shown in Table 3. Combined with the static milling process parameter characteristics, 234 corresponding characteristics are obtained, which are used to construct the training data set of L1-LSSVR model.

[0035] Table 3 Time domain feature calculation formula

[0036]

[0037] Step S2.3, after the above feature truncation processing and time domain feature extraction operation are completed, the dynamic features obtained through feature extraction and static milling process parameters are combined as input feature variables of the model, and the corresponding CF / PEEK milling three-dimensional surface roughness value is taken as the output variable, to constitute the training set D={(x i ,y i )|i=1, 2,..., n} of the L1-LSSVR model, wherein x i represents the input feature vector of the training data, y i represents the output value of the training data, i.e. the CF / PEEK milling three-dimensional surface roughness value, and n represents the sample number of the training data set; the training data set D is divided into training set D train and test set D text in proportion respectively;

[0038] Step S2.4, the input feature vector of the training data set D={(x i ,y i )|i=1, 2,..., n} is normalized, and the normalization method is as follows:

[0039]

[0040] wherein,

[0041] is the normalized input feature vector;

[0042] xt i is the feature value of the original input feature vector of the i-th training sample;

[0043] xt min is the minimum value in the input feature vector;

[0044] xt max is the maximum value in the input feature vector.

[0045] Step S3 includes the following steps:

[0046] Step S3.1, the correlation between features is analyzed by using Pearson correlation coefficient, which aims to eliminate redundant features with similar influence law on three-dimensional surface roughness. The calculation formula of Pearson correlation coefficient is the ratio of the covariance between variables to the standard deviation, and the calculation formula is as follows:

[0047]

[0048] wherein,

[0049] n represents the sample number; x iColumn data representing input features; y i Column data representing output; Column data representing features x i Mean value of x; Mean value of y i Mean value of y

[0050] Static process parameters do not participate in the calculation of Pearson correlation coefficient, and 11 time domain features extracted from each dynamic signal data need to be calculated for Pearson correlation coefficient. One of the two features with PCC value (i.e. correlation) greater than 0.8 is selected to reduce information redundancy;

[0051] Step S3.2, the feature selection principle of L1 norm obtains the weight of the feature by assuming a linear regression fitting model. The squared error of the predicted value and the actual value is taken as the loss function, and an L1 regularization penalty term is introduced to establish the objective function of the model; the features are sparsified by using the regularization operation in the process of solving the objective function, and the sparse solution is obtained, i.e. the features with zero model coefficients, so that the features related to the target are selected. The objective function includes two parts: the loss function and the regularization penalty term. The squared error of the linear regression model is the loss function; the L1 regularization penalty term is introduced to the feature weight vector;

[0052] The objective function is defined as follows:

[0053]

[0054] Where b is the intercept; the first term of the objective function L(w, b) is the sum of squared prediction errors; w j is the feature weight coefficient matrix; p represents the number of features, j = 1, 2,..., p; λ is the regularization penalty parameter, which is used to control the L1 regularization penalty strength;

[0055] Step S3.3, in order to solve the optimization problem of the objective function with L1 regularization term, the non-continuity problem of the loss function (the absolute value function is not continuous at zero) needs to be solved. The alternating direction multiplier method (ADMM) is used for optimization, which decomposes the complex problem into multiple relatively simple sub-problems for solving the objective function containing L1 regularization. The z variable is introduced to represent ω, so that each sub-problem can be optimized independently, which is convenient for solving the sparsity problem of L1 regularization. Compared with directly optimizing ω, the convergence speed is faster. The objective function is expressed in the form of minimization as shown below:

[0056]

[0057] Where the auxiliary variable z = w represents the sparse output of w;

[0058] Step S3.4, a Lagrange multiplier u is introduced in the solving process to construct a Lagrange function, as shown in the following formula:

[0059]

[0060] wherein, ρ is a control parameter of ADMM, used to adjust the weight of the Lagrange multiplier, usually taking the value of 1;

[0061] Step S3.5, the original optimization problem is decomposed into two or more sub-problems which are easy to optimize, and the solution of the sub-problems is calculated by using the Lagrange multiplier, so that the function finally converges to the minimum value of the objective function; the update step of ADMM is divided into three sub-problems which are optimized alternately; fixing z and u, w is updated, as shown in the following formula:

[0062]

[0063] Fixing w and u, the update of z is carried out, as shown in the following formula:

[0064]

[0065] The Lagrange multiplier u is updated, as shown in the following formula:

[0066] u (k+1) =u (k) +z (k+1) -w (k+1) ;

[0067] Step S3.6, in the iterative calculation process of the objective function, the variable z update formula can be converted into a form of directly applying soft threshold operation to realize sparsity, and some feature weight coefficients are compressed to zero to realize feature selection. The function of soft threshold operation is derived through the optimal solution of L1 norm regularization, and the soft threshold operation expression and threshold operation function are as shown in the following formula:

[0068]

[0069]

[0070] Wherein, λ / ρ represents the threshold adjusted for the optimal solution iteration;

[0071] Step S3.7, after ADMM iteration, the objective function is minimized, and L1 norm regularization is achieved; the final feature vector is its sparse solution, and the penalty term will compress part of the feature weight coefficient to zero, representing that the feature is irrelevant to the target. For the remaining non-zero solution selected, the feature weight coefficient can be calculated according to the soft threshold operation to obtain the weight coefficient size of the feature. In the case of L1 regularization, the absolute value of the non-zero feature weight coefficient can reflect the strength of the relationship between the feature and the target variable to some extent. Its absolute value reflects the influence degree of different features on the three-dimensional surface roughness in the ADMM convergence state.

[0072] Step S4 includes the following steps:

[0073] Step S4.1, the features selected in step S3 are included in the training data set D of the LSSVR model together with the milled three-dimensional surface roughness L1-g , and the data set is divided into a training set D L1-train and a test set D L1-text according to the required 8:2 ratio.

[0074] Step S4.2, the kernel function of LSSVR is a multi-kernel function combining radial basis kernel function and polynomial kernel function, and a linear combination strategy is used for weighted summation. The calculation formula of the multi-kernel function is as follows:

[0075]

[0076] wherein,

[0077] is a multi-kernel function; x i and are two input feature vectors; γ1>0 is the kernel parameter of the radial basis kernel function, which controls the width of the kernel function; represents the Euclidean distance between the two feature vectors; α p >0, c p and d are the kernel parameters of the polynomial kernel function, which represent the scaling factor, the bias and the order of the polynomial respectively;

[0078] Step S4.3, the selection of the hyperparameters of the LSSVR algorithm will affect the accuracy of the established model, so selecting appropriate hyperparameter combinations (regularization penalty parameter λ, kernel parameter γ1 of the radial basis kernel function, scaling factor α p of the polynomial kernel function) can enhance the performance of the algorithm. In view of the instability of manual parameter adjustment,

[0079] The fish eagle optimization algorithm improved by using Tent chaotic mapping and convex lens imaging reverse learning strategy is used for hyperparameter optimization. Specifically, the fish eagle population N is set, and the iteration number T is set. The constraint of the parameters to be optimized is set. The fish eagle population is initialized as {Xi i = 1, 2, …, N}, N represents the size of the fish eagle population, and the individual fish eagle is X i = {x i,1 , x i,2 , x i,3 , …, x i,j}, j represents the dimension of the fish eagle position vector;

[0080] Step S4.4, the Tent chaotic mapping improves the population initialization step as follows:

[0081]

[0082] x i,j = lb j +(ub j -lb j )×Z k+1

[0083] Wherein, Z k+1 represents the chaotic sequence of the Tent mapping, and Z k is a pseudo-random number between [0, 1]; a z is a set coefficient, taking the value of 0.4; x i,j represents the value of the i-th fish eagle in the j-th dimension; lb j is the j-th dimension of the lower bound of the search, and ub j is the j-th dimension of the upper bound of the search;

[0084] Step S4.5, the fish eagle optimization algorithm searches for other fish eagle positions with higher fitness values in the solution space in the first stage, and performs hunting at this position, and the position update formula is as follows:

[0085] P = {X k | k ∈ {1, 2, …, N} ∧ F e k <F e j}∪{X best}

[0086]

[0087]

[0088]

[0089] Wherein, P represents the individual set of the underwater fish population; represents the fitness value of the k-th fish eagle; represents the fitness value of the j-th fish eagle; X k represents the position of the k-th fish eagle individual with better fitness; X bestIt is the location of the best individual osprey; is the new position of the i-th osprey in this stage The value in the jth dimension; r i,j Represents a random number between the interval [0,1]; SP i is the position of the fish randomly selected by the i-th osprey from its underwater fish school P, set to a random number between 0 and 1; I i，j is any number in the set {1,2}; X i represents the individual position of the i-th osprey, Indicates the individual osprey at this stage; F ei represents the individual fitness value of the i-th osprey, is the fitness value of the i-th osprey in this stage;

[0090] Step S4.6: In the second stage, after the osprey catches a fish, it will move it to a suitable location for feeding. Here, a reverse learning strategy based on the convex lens imaging principle is introduced to improve the osprey optimization algorithm. Its mathematical expression is shown as follows:

[0091]

[0092] Where r is a random number in the interval [0,1]; t is the current iteration number; is the new position of the i-th Osprey in the second phase The value in the jth dimension; represents the inverse solution generated by the convex lens imaging principle;

[0093] The generated reverse solution is added to the position update judgment condition of the second stage of the Osprey optimization algorithm to find the optimal individual. The position update formula is as follows:

[0094]

[0095] in, Reverse individuals for the new Osprey individuals of the second phase; Represents the fitness value of the reverse individual;

[0096] Step S4.7, set the osprey population N, the number of iterations T and the hyperparameter constraints; initialize the osprey population, and the individual osprey position vector is composed of the regularization penalty parameter γ, the kernel parameter γ1 of the radial basis kernel function, and the scaling factor α of the polynomial kernel function. p The hyperparameters are updated and the LSSVR model is trained. The root mean square error (RMSE) of the test set is fed into the Osprey optimization algorithm to calculate the fitness of the individual Ospreys. After completing the iteration of the Osprey optimization algorithm, the optimal hyperparameter combination is determined within the hyperparameter constraints.

[0097] The modeling method is used for the cutting speed range 150-300 m / min, and the milling processing environment is the milling processing scene simulation of dry milling.

[0098] In the milling processing scene, the tool is selected as a polycrystalline diamond tool, and a carbon fiber reinforced polyether ether ketone CF / PEEK plate is milled.

[0099] In step S1, the three-dimensional surface roughness of each milling experiment is obtained after the experiment, and the three-dimensional surface roughness S is calculated by using the surface information of the test piece measured by the three-dimensional profiler. q The calculation formula is as follows:

[0100]

[0101] Wherein, A is the total area of the measured surface; Z(x,y) is the deviation height of the measured surface.

[0102] The model performance of the CF / PEEK milling surface roughness model is evaluated by using the mean absolute percentage error MAPE and the determination coefficient R 2 The fitting accuracy of the model is evaluated.

[0103] The machine learning method is used, the underlying mechanism of the machine tool milling processing can be bypassed, the complex relationship between the process parameters and the processing results is mined by learning the processing data, so that a more accurate and stable model is established, and the algorithm designed in the application can effectively realize the modeling of the three-dimensional surface roughness of CF / PEEK material milling. BRIEF DESCRIPTION OF DRAWINGS

[0104] The application will be further described in detail in combination with the drawings and specific embodiments:

[0105] The Figure 1 It is the L1-LSSVR algorithm flowchart in the embodiment;

[0106] The Figure 2 It is the Pearson correlation coefficient and scatter plot matrix diagram of the tangential force part time domain characteristics in the embodiment. DETAILED DESCRIPTION

[0107] As shown in the figure, a CF / PEEK milling surface roughness modeling method based on L1-LSSVR algorithm comprises the following steps:

[0108] Step S1: taking the static milling process parameters as the experimental variables, designing the experimental scheme, collecting the dynamic signals in the milling process during the experiment, and measuring the three-dimensional surface roughness of CF / PEEK after the experiment;

[0109] Step S2: Extract the time domain features of the dynamic signals (electrical signals, cutting force signals), combine with static milling process parameters, and use for L1-LSSVR model training;

[0110] Step S3: Analyze the correlation between features by Pearson correlation coefficient, eliminate redundant features that have similar influence rules on three-dimensional surface roughness; and use L1 norm regularization for feature selection, and select features with high correlation with three-dimensional surface roughness.

[0111] Step S4: Use the features selected in step S3 as the input of LSSVR algorithm, and the three-dimensional surface roughness of CF / PEEK milling as the output, to establish the surface roughness model of CF / PEEK milling.

[0112] In step S1, the static milling process parameters include cutting speed, cutting depth and feed per tooth, and the dynamic signals include electrical signals and cutting force signals.

[0113] Step S1 includes the following steps:

[0114] Step S1.1, taking cutting speed, cutting depth, feed per tooth and fiber direction angle as experimental factors, setting the levels of each factor and designing orthogonal experiment, the constraints of each factor are set as:

[0115] v min ≤v c ≤v max

[0116] f min ≤f z ≤f max

[0117] a pmin ≤a p ≤a pmax

[0118] θ=u1、v1

[0119] Wherein,

[0120] v c , f z , a p , θ respectively represent cutting speed, cutting depth, feed per tooth and fiber direction angle;

[0121] v min , v max respectively represent the minimum and maximum values of cutting speed;

[0122] f min , f max respectively represent the minimum and maximum values of cutting depth;

[0123] apmin , a pmax are minimum and maximum values of the feed per tooth, respectively;

[0124] u1, v1 are 0°, 90° fiber direction angles;

[0125] Step S1.2, CF / PEEK milling experiments are carried out according to the experimental scheme, and dynamic signal data in the milling process is collected during the experiment;

[0126] Specifically, 18 groups of electrical signal parameters (as shown in Table 1) and three-direction cutting force signals in the milling process (as shown in Table 2) are included. After the experiment, the three-dimensional surface roughness of CF / PEEK after milling is measured and recorded.

[0127] Table 1 Electrical signal parameter table

[0128]

[0129] Table 2 Cutting force signal parameter table

[0130]

[0131] Step S2 includes the following steps:

[0132] Step S2.1, before the time domain feature extraction of the dynamic signal, the signal is truncated to extract the time domain features of each group of dynamic signals. At the beginning of milling, the dynamic signal data fluctuates in a large range, and the data in the milling process stage is cut off from the fluctuation. The cut-off signal is used as the data for subsequent feature extraction.

[0133] Step S2.2, 11 time domain features of each group of dynamic signals are extracted. Through the extraction of peak-to-peak value, mean value, peak value, standard deviation, skewness, kurtosis, root mean square, waveform factor, margin factor, kurtosis factor and pulse factor, a total of 11 indexes are calculated to obtain corresponding 231 (21x11) dynamic characteristics,

[0134] The index calculation formula is shown in Table 3. Combined with the static milling process parameter characteristics, 234 corresponding characteristics are obtained, which are used to construct the training data set of L1-LSSVR model.

[0135] Table 3 Time domain feature calculation formula

[0136]

[0137]

[0138] Step S2.3, after the above feature truncation processing and time domain feature extraction operation are completed, the dynamic features obtained through feature extraction and static milling process parameters are combined as input feature variables of the model, and the corresponding CF / PEEK milling three-dimensional surface roughness value is taken as the output variable, to form the training set D={(x i ,y i )|i=1, 2,..., n} of the L1-LSSVR model, wherein x i represents the input feature vector of the training data, y i represents the output value of the training data, i.e. the CF / PEEK milling three-dimensional surface roughness value, and n represents the sample number of the training data set; the training data set D is divided into training set D train and test set D text in proportion respectively.

[0139] Step S2.4, the input feature vector of the training data set D={(x i ,y i )|i=1, 2,..., n} is normalized, and the normalization method is as follows:

[0140]

[0141] wherein,

[0142] is the normalized input feature vector;

[0143] x i is the feature value of the original input feature vector of the i-th training sample;

[0144] x min is the minimum value in the input feature vector;

[0145] x max is the maximum value in the input feature vector.

[0146] Step S3 includes the following steps:

[0147] Step S3.1, the correlation between features is analyzed by using Pearson correlation coefficient, which aims to eliminate redundant features with similar influence law on three-dimensional surface roughness. The calculation formula of Pearson correlation coefficient is the ratio of the covariance between variables to the standard deviation, and the calculation formula is as follows:

[0148]

[0149] wherein,

[0150] n represents the sample number; x i represents the column data of the input feature; y iColumn data representing output; representative feature x i average value of y representative y i average value of y

[0151] Static process parameters do not participate in the calculation of Pearson correlation coefficient, and the Pearson correlation coefficient of 11 time domain features extracted from each dynamic signal data needs to be calculated. One of the two features with a PCC value (i.e. correlation) greater than 0.8 is selected to reduce information redundancy;

[0152] Step S3.2, the feature selection principle of L1 norm obtains the weight of the feature by assuming a linear regression fitting model. The squared error of the predicted value and the actual value is taken as the loss function, and an L1 regularization penalty term is introduced to establish the objective function of the model; the features are sparsified by using the regularization operation in the process of solving the objective function, and the sparse solution is obtained, that is, the features with zero model coefficients are selected, so that the features related to the target are selected. The objective function includes two parts: the loss function and the regularization penalty term. The squared error of the linear regression model is the loss function; the L1 regularization penalty term is introduced to the feature weight vector;

[0153] The objective function is defined as follows:

[0154]

[0155] Where b is the intercept; the first term of the objective function L(w, b) is the sum of squared prediction errors; w j is the feature weight coefficient matrix; p represents the number of features, j = 1, 2,..., p; λ is the regularization penalty parameter, which is used to control the L1 regularization penalty strength;

[0156] Step S3.3, in order to solve the optimization problem of the objective function with L1 regularization term, the non-continuity problem of the loss function (the absolute value function is not continuous at zero) needs to be solved. The alternating direction multiplier method (ADMM) is used for optimization, which decomposes the complex problem into multiple simpler sub-problems for solving the objective function containing L1 regularization. The z variable is introduced to represent w, so that each sub-problem can be optimized independently, which is convenient for solving the sparsity problem of L1 regularization. Compared with directly optimizing w, the convergence speed is faster. The objective function is expressed in the form of minimization as shown below:

[0157]

[0158] Where the auxiliary variable z = w represents the sparse output of w;

[0159] Step S3.4, in the solving process, the Lagrange multiplier u is introduced to construct the Lagrange function as shown below:

[0160]

[0161] wherein p is a control parameter of ADMM, used to adjust the weight of the Lagrange multiplier, usually taking the value of 1;

[0162] Step S3.5, the original optimization problem is decomposed into two or more sub-problems which are easy to optimize, and the solution of the sub-problems is calculated by using the Lagrange multiplier, so as to finally make the function converge to the minimum value of the objective function; the update step of ADMM is divided into three sub-problems which are alternately optimized; fixing z and u, w is updated, as shown in the following formula:

[0163]

[0164] Fixing w and u, the update of z is performed, as shown in the following formula:

[0165]

[0166] The Lagrange multiplier u is updated, as shown in the following formula:

[0167] u (k+1) =u (k) +z (k+1) -w (k+1) ;

[0168] Step S3.6, in the iterative calculation process of the objective function, the variable z update formula can be converted into a form of directly applying soft threshold operation to realize sparsity, and some feature weight coefficients are compressed to zero to realize feature selection. The function of the soft threshold operation is derived through the optimal solution of the L1 norm regularization, and the soft threshold operation expression and the threshold operation function are as shown in the following formula:

[0169]

[0170] wherein λ / p represents the threshold adjusted for the optimal solution iteration;

[0171] Step S3.7, after the ADMM iteration, the objective function reaches the minimum, and the L1 norm regularization is realized; the final obtained feature vector is its sparse solution, and the penalty term compresses part of the feature weight coefficients to zero, representing that the feature is irrelevant to the target. For the remaining non-zero solution screened out, the weight coefficient of the feature can be calculated according to the soft threshold operation, and the weight coefficient size of the feature can be calculated according to the soft threshold operation. In the case of L1 regularization, the absolute value of the non-zero feature weight coefficient can reflect the strength of the relationship between the feature and the target variable to some extent, and the absolute value reflects that, in the ADMM convergence state, the influence degree of different features on the three-dimensional surface roughness.

[0172] Step S4 includes the following steps:

[0173] Step S4.1, the features screened in step S3 are combined with the milling three-dimensional surface roughness into the training data set D of LSSVR model L1-g In the middle, and according to the required 8:2 ratio, the data set is divided into training set D L1-train And test set D L1-text ;

[0174] Step S4.2, the kernel function of LSSVR is selected as the multi-kernel function combined with radial basis kernel function and polynomial kernel function, and the linear combination strategy is used for weighted summation, and the calculation formula of multi-kernel function is as follows:

[0175]

[0176] Wherein,

[0177] is the multi-kernel function; x i And is two input feature vectors; γ1>0 is the kernel parameter of radial basis kernel function, which controls the width of kernel function; represents the Euclidean distance between two feature vectors; α p >0, c p And d are the kernel parameters of polynomial kernel function, which represent the scaling factor, bias and the order of polynomial respectively;

[0178] Step S4.3, the selection of super parameters of LSSVR algorithm will affect the accuracy of the established model, therefore, selecting appropriate super parameter combination (regularization penalty parameter λ, kernel parameter γ1 of radial basis kernel function, scaling factor α P Of polynomial kernel function) can enhance the performance of algorithm, in view of the instability of manual parameter adjustment;

[0179] The fish eagle optimization algorithm improved by using Tent chaotic mapping and convex lens imaging reverse learning strategy is used for super parameter optimization, specifically: set the fish eagle population N, and the iteration number is set to T, set the constraint of the parameters to be optimized; initialize the fish eagle population as {X i ,i=1,2,...,N}, N represents the size of fish eagle population, and the individual fish eagle is X i ={x i,1 ,x i,2 ,x i,3 ,....,x i,j}, j represents the dimension of fish eagle position vector;

[0180] Step S4.4, the Tent chaotic mapping improved population initialization step is as follows:

[0181]

[0182] xi,j =lb j +(ub j -lb j )×Z k+1

[0183] Among them, Z k+1 represents the chaotic sequence of Tent mapping, Z k is a pseudo-random number between [0,1]; a z is the setting coefficient, the value is 0.4; x i,j represents the value of the i-th osprey in the j-th dimension; lb j is the jth dimension to search for the lower bound, ub j is the jth dimension of the search upper bound;

[0184] Step S4.5: In the first phase, the osprey optimization algorithm searches for other osprey positions with higher fitness values ​​in the solution space and hunts at these positions. The position update formula is as follows:

[0185] P={X k |k∈{1,2,…,N}∧F e k <F e j}∪{X best}

[0186]

[0187]

[0188]

[0189] Among them, P represents the individual set of underwater fish; represents the fitness value of the kth osprey; represents the fitness value of the jth osprey; X k represents the position of the kth osprey with better fitness; X best It is the location of the best individual osprey; is the new position of the i-th osprey in this stage The value in the jth dimension; r i,j Represents a random number between the interval [0,1]; SP i is the position of the fish randomly selected by the i-th osprey from its underwater fish school P, set to a random number between 0 and 1; I i，j is any number in the set {1,2}; X i represents the individual position of the i-th osprey, Indicates the individual osprey at this stage; F ei represents the individual fitness value of the i-th osprey, is the fitness value of the i-th fish eagle in the first stage;

[0190] Step S4.6, after the fish eagle captures a fish in the second stage, it will move to a suitable position for feeding, and the reverse learning strategy based on the convex lens imaging principle is introduced to improve the fish eagle optimization algorithm, and the mathematical expression is as follows:

[0191]

[0192]

[0193] Wherein, r is a random number in the interval [0, 1]; t is the current iteration number; is the new position of the i-th fish eagle in the second stage The value in the jth dimension; Indicates the reverse solution generated by the convex lens imaging principle;

[0194] The generated reverse solution is added to the position update judgment condition of the second stage of the fish eagle optimization algorithm to find the optimal individual, and the position update formula is as follows:

[0195]

[0196] Wherein, The new fish eagle individual in the second stage is the reverse individual; Indicates the fitness value of the reverse individual;

[0197] Step S4.7, set the fish eagle population N, the iteration number T and the hyperparameter constraint; initialize the fish eagle population, and the fish eagle individual position vector is composed of the regularization penalty parameter γ, the kernel parameter γ1 of the radial basis kernel function and the scaling factor α of the polynomial kernel function p The hyperparameters are updated, and the LSSVR model is trained; the root mean square error RMSE of the test set is substituted into the fish eagle optimization algorithm for fitness calculation of the fish eagle individual. After the iteration of the fish eagle optimization algorithm is completed, the optimal hyperparameter combination is determined within the constraint range of the hyperparameters.

[0198] The modeling method is used for the cutting speed range of 150-300 m / min, and the milling processing environment is dry milling.

[0199] In the milling processing scene, the tool is selected as a polycrystalline diamond tool, and a carbon fiber reinforced polyether ether ketone CF / PEEK plate is milled.

[0200] In step S1, the three-dimensional surface roughness of each milling experiment is obtained after the experiment, and the three-dimensional surface roughness S q is calculated by using the surface information of the test piece measured by the three-dimensional profilometer; the calculation formula is as follows:

[0201]

[0202] where A is the total area of the measuring surface; Z(x, y) is the deviation height of the measuring surface.

[0203] The milling experiment results are shown in Table 6.

[0204] Table 6 CF / PEEK milling experiment results

[0205]

[0206] During the milling process, dynamic signals (electrical signals, cutting force signals) are collected. Table 7 and Table 8 respectively show part of the time domain characteristics of the electrical signals and the characteristic table after normalization processing.

[0207] Table 7 Part of the characteristic table extracted from phase A voltage

[0208]

[0209] Table 8 Part of the characteristic table extracted from phase A voltage after normalization processing

[0210]

[0211] After completing the time domain feature extraction, the Pearson correlation coefficient between each feature is calculated and the correlation between the features is analyzed to obtain the scatter matrix diagram. Part of the dynamic characteristics of the tangential force Fx are selected: mean, root mean square, peak-to-peak value, peak value, waveform factor, pulse factor, kurtosis, and the Pearson correlation coefficient and scatter matrix diagram of part of the dynamic characteristics are shown (as shown in Figure 2 The horizontal and vertical coordinates of the scatter diagram respectively represent the values of two characteristic quantities, and the relationship and change trend between the two characteristic quantities are presented. When the scatter diagram tends to be linear, the correlation is high; and the more scattered the data points of the scatter diagram are, the lower the correlation between the characteristic quantities is. In the two features with a Pearson correlation coefficient greater than 0.8, one is selected for subsequent L1 regularization processing.

[0212] The L1 norm is used to select features (process parameters, electrical signal features, cutting force signal features), and features with zero feature weight coefficients are removed, thereby obtaining features related to three-dimensional surface roughness. According to step 3.6, the feature weight matrix can be obtained, and the features after compression of the L1 regularization penalty term of the feature weight matrix can be obtained, in which only the feed per tooth and the cutting speed belong to the milling process parameters, and in the result after ADMM iteration, it can be seen that the weight coefficient of the process parameter is lower than that of most dynamic features. In order to intuitively observe the influence of dynamic features on three-dimensional surface roughness, Table 9 lists the top 10 features in terms of absolute value of feature weight coefficient.

[0213] Table 9 Top 10 feature weight coefficients

[0214]

[0215] Features with non-zero feature coefficients are selected as training dataset D L1-g for CF / PEEK three-dimensional surface roughness model, and the dataset is divided into training set D L1-train and test set D L1-text according to the ratio of 8:2. LSSVR hyperparameters are optimized using improved fish optimization algorithm, with fish population size N L1 = 100, iteration number T L1 = 100, and LSSVR hyperparameter constraints set as follows:

[0216]

[0217] Initialize the fish population, and the fish population individual position vector is composed of LSSVR hyperparameters (regularization penalty parameter γ, kernel parameter γ1 of radial basis kernel function, scaling factor α p of polynomial kernel function). The root mean square error RMSE of the test set is used to calculate the fitness value of the fish individual. After the iteration termination condition is reached, the optimal hyperparameter combination is obtained as shown in Table 10.

[0218] Table 10 L1-LSSVR model hyperparameters

[0219]

[0220] The model performance of the CF / PEEK milling surface roughness model is evaluated using the mean absolute percentage error MAPE and the coefficient of determination R 2 to evaluate the fitting accuracy of the model.

[0221] Specifically, the LSSVR algorithm is used to construct a CF / PEEK milling surface roughness model with the selected features as input and three-dimensional surface roughness as output. The mean absolute percentage error MAPE and the coefficient of determination R 2 are used to evaluate the fitting accuracy of the model and evaluate the performance of the model, as shown in Table 11.

[0222] Table 11 L1-LSSVR model evaluation results

[0223]

[0224] That is: complete the training of the model, and obtain the L1-LSSVR model established by the optimal hyperparameters, in order to further verify the effectiveness of the three-dimensional surface roughness model, three groups of milling process parameter combinations (the 3rd, 9th and 15th groups) are randomly selected from the milling experiment scheme, and five groups of data are obtained by using the cutting speeds of 284.12 and 276.2 m / min, the cutting depths of 1.40 and 1.43 mm, the feed per tooth of 0.046 and 0.049 mm / z, and the fiber direction angle of 0°. The five groups of milling process parameter combinations, the corresponding electric signals and cutting force signals collected during the milling experiment, and the actual milling three-dimensional surface roughness measurement results are used as the verification set of the model, and instance verification is performed, and the model output values obtained by the five groups of verification data are compared with the actual measurement values. The specific verification results are shown in Table 12.

[0225] Table 12 Model verification results

[0226]

[0227] As shown in Table 11 and Table 12, the algorithm designed in the application can effectively realize the modeling of the three-dimensional surface roughness of the CF / PEEK material milling.

[0228] Embodiment:

[0229] The application provides a CF / PEEK milling surface roughness modeling method based on an L1-LSSVR algorithm, data is derived from actual CF / PEEK milling experiments, a 10 mm polycrystalline diamond single-blade milling cutter is selected to mill CF / PEEK plates, and a VMC850Q vertical machining center produced by Shenyang Machine Tool Co., Ltd. is selected as the machining equipment. An orthogonal experiment is designed and carried out, dynamic signals in the machining process are collected, and the machining results are measured. The method uses CF / PEEK static milling process parameters (cutting speed, cutting depth and feed per tooth) and dynamic signals (electric signals and cutting force signals) as the training data set of the CF / PEEK surface roughness model, and an L1-LSSVR algorithm is provided, so that feature selection and CF / PEEK milling surface roughness modeling are realized. Figure 1 The L1-LSSVR algorithm flowchart in the application.

[0230] To achieve the above object, the application adopts the following technical scheme:

[0231] The scheme includes the following specific steps:

[0232] Step S1: Design an experimental scheme with static milling process parameters (cutting speed, cutting depth, and feed per tooth) as experimental variables. Collect dynamic signals (electrical signals and cutting force signals) during the milling process. After the experiment, measure the CF / PEEK three-dimensional surface roughness.

[0233] Step S2: Extract the time-domain features of the dynamic signals (electrical signals and cutting force signals) and combine the static milling process parameters (cutting speed, cutting depth, and feed per tooth) for L1-LSSVR model training.

[0234] Step S3: Analyze the correlation between features by Pearson correlation coefficient, and eliminate redundant features that have similar influence on the three-dimensional surface roughness. Use L1 norm regularization for feature selection to screen features with high correlation with the three-dimensional surface roughness.

[0235] Step S4: Use the features selected in step S3 as the input of the LSSVR algorithm, and the CF / PEEK milling three-dimensional surface roughness as the output to establish the CF / PEEK milling surface roughness model.

[0236] The processing equipment selected in this example is a vertical machining center, the material selected is CF / PEEK laminate, the workpiece fiber direction is fixed, the size is 100 mm x 100 mm x 10 mm, and side milling is performed. The processing tool is a PCD milling cutter.

[0237] The cutting speed, cutting depth, feed per tooth, and fiber direction angle are used as factors to design a four-factor mixed-level CF / PEEK milling orthogonal experiment. The experimental factors and level settings are shown in Table 4, and the orthogonal experiment plan is shown in Table 5.

[0238] Table 4 CF / PEEK milling experimental factors and level settings

[0239]

[0240] Table 5 Orthogonal experiment plan

[0241]

[0242]

Claims

1. A CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm, characterized by: The following steps are included: Step S1: Using static milling process parameters as experimental variables, an experimental plan is designed. During the experiment, dynamic signals during the milling process are collected. After the experiment, the three-dimensional surface roughness of CF / PEEK is measured. Step S2: extracting the time domain features of the dynamic signal and combining them with the static milling process parameters for L1-LSSVR model training; Step S3: The correlation between features is analyzed by the Pearson correlation coefficient, and redundant features with similar influence on the three-dimensional surface roughness are eliminated; and L1 norm regularization is used for feature selection to screen out features with high correlation with the three-dimensional surface roughness. Step S4: Using the features selected in step S3 as the input of the LSSVR algorithm and the CF / PEEK milling three-dimensional surface roughness as the output, a CF / PEEK milling surface roughness model is established.

2. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: In step S1 , the static milling process parameters include cutting speed, cutting depth and feed per tooth, and the dynamic signals include electrical signals and cutting force signals.

3. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 2, characterized in that: Step S1 includes the following steps: Step S1.1: Using cutting speed, cutting depth, feed per tooth, and fiber orientation angle as experimental factors, set the levels of each factor and design an orthogonal experiment. The constraints of each factor are set as follows: in min ≤in c ≤in max f min ≤f z ≤f max a pmin ≤a p ≤a pmax θ=u1、v1 in, v c 、f z 、a p , θ represent cutting speed, cutting depth, feed per tooth and fiber orientation angle, respectively; v min 、v max Represent the minimum and maximum cutting speed respectively; f min 、f max Represent the minimum and maximum values ​​of cutting depth respectively; a pmin 、a pmax are the minimum and maximum values ​​of feed per tooth respectively; u1 and v1 are fiber orientation angles of 0° and 90°; Step S1.2: Conduct a CF / PEEK milling experiment according to the experimental plan, and collect dynamic signal data during the milling process; after the experiment, measure and record the three-dimensional surface roughness of the CF / PEEK after milling.

4. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: Step S2 includes the following steps: Step S2.1: Before extracting the time-domain features of the dynamic signals, the signals are truncated to extract the time-domain features of each group of dynamic signals. At the start of milling, the dynamic signal data fluctuates widely. Data at the milling stage is intercepted from this fluctuation and used as the data for subsequent feature extraction. Step S2.2, extract the time domain features of each group of dynamic signals, calculate the corresponding dynamic features by extracting the peak-to-peak value, mean, peak value, standard deviation, skewness, kurtosis, root mean square, form factor, margin factor, kurtosis factor and impulse factor indicators, combine them with the static milling process parameter characteristics to obtain the corresponding features, and construct the training data set of the L1-LSSVR model; Step S2.3: After completing the above-mentioned feature truncation processing and time domain feature extraction operations, the dynamic features and static milling process parameters obtained through feature extraction are combined as the input feature variables of the model, and the corresponding CF / PEEK milling three-dimensional surface roughness values ​​are used as output variables to form the training set D = {(x i ,y i )|i=1,2,...,n}, where x i Represents the input feature vector of the training data, y i represents the output value of the training data, i.e., the 3D surface roughness value of CF / PEEK milling, and n represents the number of samples in the training data set; the training data set D is divided into training sets D train and the test set D text ; Step S2.4: Set the training data set D = {(x i ,y i )|i=1,2,...,n} is normalized. The normalization method is as follows: in, is the input feature vector after normalization; x t i is the original input eigenvector eigenvalue of the i-th training sample; xt min is the minimum value in the input eigenvector; xt max is the maximum value in the input feature vector.

5. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: Step S3 includes the following steps: Step S3.1: Use the Pearson correlation coefficient to analyze the correlation between features and eliminate redundant features with similar influence on the three-dimensional surface roughness. The calculation formula of the Pearson correlation coefficient is the ratio of the covariance between the variables to the standard deviation, and its calculation formula is as follows: Where n represents the number of samples; x i Column data representing input features; y i Represents the output column data; Representative feature x i The average value of Represents y i The average value of Static process parameters are not included in the Pearson correlation coefficient calculation. The Pearson correlation coefficient is calculated for each time domain feature extracted from the dynamic signal data. One of the two features with a PCC value greater than 0.8 is selected to reduce information redundancy. Step S3.2: The L1-norm feature selection principle obtains feature weights by assuming a linear regression fit model. The squared error between the model's predicted value and the actual value is used as the loss function, and an L1 regularization penalty term is introduced to establish the model's objective function. In solving the objective function, the regularization operation is used to sparse the features, obtaining a sparse solution, that is, features with zero model coefficients, thereby selecting features relevant to the target. The objective function consists of two parts: the loss function and the regularization penalty term: the squared error of the linear regression model is the loss function; the L1 regularization penalty term introduced to the feature weight vector; The objective function is defined as follows: Where b is the intercept; the first term of the objective function L(w,b) is the sum of squares of the prediction error; w j is the feature weight coefficient matrix; p represents the number of features, j = 1, 2, ..., p; λ is the regularization penalty parameter used to control the L1 regularization penalty strength; Step S3.3: Use the alternating direction multiplier method (ADMM) for optimization to decompose the composite problem into multiple subproblems to solve the objective function with L1 regularization. Introduce the z variable to represent w so that each subproblem can be optimized independently, which facilitates solving the sparsity problem of L1 regularization. The objective function is expressed as the minimization form shown in the following formula: Among them, the auxiliary variable z=w represents the sparse output of w; Step S3.4: In the solution process, introduce the Lagrange multiplier u to construct the Lagrange function, as shown in the following formula: Among them, ρ is the control parameter of ADMM, which is used to adjust the weight of the Lagrange multiplier and is usually set to 1; Step S3.5: Decompose the original optimization problem into subproblems, use Lagrange multipliers to calculate the solutions of the subproblems, and finally make the function converge to the minimum value of the objective function; the update step of ADMM is divided into alternating optimization of subproblems; fix z and u, and update w as shown in the following formula: Fix w and u and update z as shown below: Update the Lagrange multiplier u as shown below: u (k+1) =u (k) +z (k+1) -w (k+1) ; Step S3.6: During the iterative calculation of the objective function, the update formula for the variable z is converted to a direct soft threshold operation to achieve sparsity, compressing some feature weight coefficients to zero to achieve feature selection. The function of the soft threshold operation is derived from the optimal solution of L1 norm regularization. The soft threshold operation expression and threshold operation function are shown as follows: Among them, λ / ρ represents the threshold value after iterative adjustment for the optimal solution; Step S3.

7. After ADMM iteration, the objective function reaches its minimum and L1 norm regularization is achieved; the final eigenvector is its sparse solution, and the penalty term compresses some feature weight coefficients to zero, indicating that the feature is irrelevant to the target; for the remaining non-zero solutions screened out, the feature weight coefficient is calculated according to the soft threshold operation to determine the weight coefficient of the feature. Under L1 regularization, the absolute value of the non-zero feature weight coefficient can reflect the strength of the relationship between the feature and the target variable to a certain extent, and its absolute value reflects the degree of influence of different features on the three-dimensional surface roughness under the convergence state of ADMM.

6. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: Step S4 includes the following steps: Step S4.1: The features screened in step S3 and the milling 3D surface roughness are incorporated into the training data set D of the LSSVR model. L1-g and divide the data set into training set D according to the required ratio L1-train and the test set D L1-text ; In step S4.2, the kernel function of LSSVR uses a multi-kernel function that combines the radial basis kernel function and the polynomial kernel function. The linear combination strategy is used for weighted summation. The calculation formula for constructing the multi-kernel function is shown in the following formula: in, is a multi-core function; x i and are two input feature vectors; γ1>0 is the kernel parameter of the radial basis kernel function, which controls the width of the kernel function; represents the Euclidean distance between two eigenvectors; α p >0, c p and d are the kernel parameters of the polynomial kernel function, representing the scaling factor, bias, and order of the polynomial, respectively; Step S4.3: The hyperparameter selection of the LSSVR algorithm will affect the accuracy of the established model. Therefore, a combination of hyperparameters is selected, namely, the regularization penalty parameter λ, the kernel parameter γ1 of the radial basis kernel function, and the scaling factor α of the polynomial kernel function. p To enhance the performance of the algorithm, the improved osprey optimization algorithm based on the Tent chaotic mapping and convex lens imaging reverse learning strategy is used for hyperparameter optimization. Specifically, the osprey population N is set, the number of iterations is set to T, and the parameter constraints to be optimized are set; the osprey population is initialized to {X i ,i=1,2,...,N}, N represents the size of the osprey population, and the individual osprey is X i ={x i,1 ,x i,2 ,x i,3 ,....,x i,j }, j represents the dimension of the osprey position vector; Step S4.4, the initialization step of the improved population of Tent chaotic map is as follows: x i,j =lb j +(ub j -lb j )×Z k+1 Among them, Z k+1 represents the chaotic sequence of Tent mapping, Z k is a pseudo-random number between [0,1]; a z is the setting coefficient, the value is 0.4; x i,j represents the value of the i-th osprey in the j-th dimension; lb j is the jth dimension to search for the lower bound, ub j is the jth dimension of the search upper bound; Step S4.5: In the first stage, the osprey optimization algorithm searches for other osprey positions with higher fitness values ​​in the solution space and hunts at these positions. The position update formula is as follows: Among them, P represents the individual set of underwater fish; represents the fitness value of the kth osprey; represents the fitness value of the jth osprey; X k represents the position of the kth osprey with better fitness; X best It is the location of the best individual osprey; is the new position of the i-th osprey in this stage The value in the jth dimension; r i,j Represents a random number between the interval [0,1]; SP i is the position of the fish randomly selected by the i-th osprey from its underwater fish school P, set to a random number between 0 and 1; I i，j is any number in the set {1,2}; X i represents the individual position of the i-th osprey, Indicates the individual osprey at this stage; F ei represents the individual fitness value of the i-th osprey, is the fitness value of the i-th osprey in this stage; Step S4.6: In the second stage of the Osprey optimization algorithm, after the Osprey catches a fish, it moves it to a suitable location for feeding. The reverse learning strategy based on the convex lens imaging principle is introduced to improve the Osprey optimization algorithm. Its mathematical expression is shown as follows: Where r is a random number in the interval [0,1]; t is the current iteration number; is the new position of the i-th Osprey in the second phase The value in the jth dimension; represents the inverse solution generated by the convex lens imaging principle; The generated reverse solution is added to the position update judgment condition of the second stage of the Osprey optimization algorithm to find the optimal individual. The position update formula is as follows: in, Reverse individuals for the new Osprey individuals of the second phase; Represents the fitness value of the reverse individual; Step S4.7, set the osprey population N, the number of iterations T and the hyperparameter constraints; initialize the osprey population, and the individual osprey position vector is composed of the regularization penalty parameter γ, the kernel parameter γ1 of the radial basis kernel function, and the scaling factor α of the polynomial kernel function. p The hyperparameters are updated and the LSSVR model is trained. The root mean square error (RMSE) of the test set is fed into the Osprey optimization algorithm to calculate the fitness of the individual Ospreys. After completing the iteration of the Osprey optimization algorithm, the optimal hyperparameter combination is determined within the hyperparameter constraints.

7. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: The modeling method is used for simulating milling processing scenarios with a cutting speed range of 150 to 300 m / min and a dry milling environment.

8. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 7, characterized in that: In the milling processing scenario, a polycrystalline diamond tool is used as the cutting tool to mill the carbon fiber reinforced polyetheretherketone CF / PEEK sheet.

9. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 7, characterized in that: In step S1, after the experiment is completed, the three-dimensional surface roughness of each milling experiment is obtained, and the three-dimensional surface roughness S is calculated using the surface information of the specimen measured by the three-dimensional profilometer. q ; The calculation formula is as follows: Where A is the total area of ​​the measurement surface; Z(x,y) is the deviation height of the measurement surface.

10. The CF / PEEK milling surface roughness modeling method based on the L1-LSSVR algorithm according to claim 1, characterized in that: Model performance of the CF / PEEK milling surface roughness model using mean absolute percentage error (MAPE) and coefficient of determination (R) 2 The model fitting accuracy was evaluated.

Images