Glass fiber reinforced composite material milling process parameter optimization method

Abstract

The invention relates to a method for optimizing milling process parameters of a glass fiber reinforced plastic (GFRP), which comprises the following steps of: constructing a full-factor experimental scheme by taking the three-dimensional surface roughness, a layering factor and processing energy consumption of GFRP milling as optimization targets and taking the rotating speed of a main shaft and the feeding speed as optimization variables; according to experimental results, training to obtain a three-dimensional surface roughness model set, a hierarchical factor model set and a processing energy consumption model set, and verifying the fitting accuracy; objective weights corresponding to the three-dimensional surface roughness, the layering factor and the machining energy consumption are calculated through a variable coefficient method and an independence weight method, weights corresponding to a three-dimensional surface roughness model, a layering factor model and a machining energy consumption model are obtained through an additive synthesis method, and therefore a multi-target comprehensive evaluation model is obtained; and taking the multi-target comprehensive evaluation model as a fitness function of IHBA, and optimizing the process parameters by adopting an IHBA algorithm to obtain optimal milling process parameters, thereby realizing optimization of the GFRP milling process parameters.

Description

Technical Field

[0001] This invention relates to the field of manufacturing technology, and specifically to a method for optimizing milling process parameters of glass fiber reinforced composite materials. Background Technology

[0002] Glass fiber reinforced polymer (GFRP) is a new type of functional material made by composite processing of glass fiber and its products (glass cloth, tape, mat, yarn, etc.) as reinforcement and synthetic resin as the matrix. Combining the advantages of glass fiber and synthetic resin, it possesses advantages such as light weight, corrosion resistance, good insulation, and high design flexibility. Therefore, GFRP is widely used in infrastructure, roads and bridges, wind power, aerospace, and new energy vehicles.

[0003] To meet the engineering application requirements of GFRP product structures, secondary processing of GFRP is unavoidable. Milling is one of the main processing methods for GFRP. Due to its heterogeneity and poor heat dissipation, GFRP is prone to tool wear, delamination, and fiber tearing during cutting, which increases the difficulty of machining GFRP. Numerous studies have shown that milling process parameters are the main factors affecting GFRP workpieces. Inappropriate selection of milling parameters can exacerbate delamination damage and reduce the surface quality of GFRP workpieces. The energy efficiency of machine tools during machining is quite low, typically less than 30%, and unreasonable milling process parameters can lead to high energy consumption. In practice, optimizing cutting process parameters can achieve ideal surface quality and save costs. Therefore, research on optimizing GFRP milling process parameters is of practical significance for improving GFRP milling quality, reducing production costs, and lowering scrap rates.

[0004] Currently, milling process parameter optimization mainly relies on operator experience or consulting technical manuals. This cannot guarantee optimal cutting parameters and increases costs and manpower waste. A small number of studies use swarm intelligence algorithms for milling process parameter optimization; however, most of these use native swarm intelligence optimization algorithms. In practical applications, native swarm intelligence optimization algorithms often suffer from getting trapped in local optima, resulting in limited optimization effectiveness. Therefore, there is an urgent need to improve native swarm intelligence optimization algorithms to enhance their global optimization capabilities, thereby more effectively solving practical optimization problems and achieving ideal optimization results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing milling process parameters of glass fiber reinforced composite materials, so as to overcome the shortcomings of the prior art and achieve effective optimization of milling process parameters of glass fiber reinforced composite materials.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for optimizing milling process parameters of glass fiber reinforced composite materials, comprising the following steps:

[0007] Step S1: Using the three-dimensional surface roughness, delamination factor, and machining energy consumption of glass fiber reinforced composite milling as optimization objectives, and spindle speed and feed rate as optimization variables, construct a full-factor experimental scheme.

[0008] Step S2: Based on the experimental results, organize and obtain the three-dimensional surface roughness dataset D. S Stratified Factor Dataset D F and processing energy consumption dataset D E Dataset D is processed using leave-one-out cross-validation. S Divided into training set D t and test set D v ; through n training sets D t1 D t2 ,…,D tn The RBFNN models are trained separately and then combined to obtain a set of three-dimensional surface roughness models M. S The same method was used to obtain the hierarchical factor model set M. F And the set of processing energy consumption models M E ;

[0009] Step S3: Through test set D v Calculate and validate the set of three-dimensional surface roughness models M S The fitting accuracy was calculated; the same method was used to calculate and validate the hierarchical factor model set M. F And the set of processing energy consumption models M E The fitting accuracy;

[0010] Step S4: Based on the experimental results, calculate the objective weights corresponding to the three-dimensional surface roughness, layering factor, and processing energy consumption using the coefficient of variation method and the independence weighting method, respectively. Then, obtain the weights corresponding to the three-dimensional surface roughness model, layering factor model, and processing energy consumption model using the additive synthesis method, thereby obtaining the multi-objective comprehensive evaluation model.

[0011] Step S5: Using the multi-objective comprehensive evaluation model obtained in step S4 as the fitness function of IHBA, and using the IHBA algorithm to optimize the process parameters, the optimal milling process parameters are obtained, thereby optimizing the milling process parameters of glass fiber reinforced composite materials.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for optimizing GFRP milling process parameters based on the RBFNN-IHBA algorithm. This method uses GFRP milling process parameters (spindle speed and feed rate) and GFRP milling three-dimensional surface roughness, machining energy consumption and stratification factor as the training model dataset. It combines the Radial Basis Function Neural Network (RBFNN) with the Improved Honey Badger Algorithm (IHBA) to achieve effective optimization of GFRP milling process parameters. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the implementation of the method for optimizing milling process parameters of glass fiber reinforced composite materials provided in this embodiment of the invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0017] like Figure 1 As shown, this embodiment provides a method for optimizing milling process parameters of glass fiber reinforced composite materials, including the following steps:

[0018] Step S1: Using the three-dimensional surface roughness, layering factor, and machining energy consumption of GFRP milling as optimization objectives, and the spindle speed and feed rate as optimization variables, construct a full-factor experimental scheme.

[0019] Step S2: Based on the experimental results, organize and obtain the three-dimensional surface roughness dataset D. S Stratified Factor Dataset D F and processing energy consumption dataset D E Dataset D is processed using leave-one-out cross-validation. S Divided into training set Dt and test set D v Through n training sets D t1 D t2 ,…,D tn The RBFNN models are trained separately and then combined to obtain a set of three-dimensional surface roughness models M. S The same method was used to obtain the hierarchical factor model set M. F And the set of processing energy consumption models M E .

[0020] Step S3: Through test set D v Calculate and validate the set of three-dimensional surface roughness models M S The fitting accuracy was calculated. The same method was used to calculate and validate the hierarchical factor model set M. F And the set of processing energy consumption models M E The fitting accuracy.

[0021] Step S4: Based on the experimental results, calculate the objective weights corresponding to the three-dimensional surface roughness, layering factor, and processing energy consumption using the coefficient of variation method and the independence weighting method, respectively. Then, obtain the weights corresponding to the three-dimensional surface roughness model, layering factor model, and processing energy consumption model using the additive synthesis method, thereby obtaining the multi-objective comprehensive evaluation model.

[0022] Step S5: Using the multi-objective comprehensive evaluation model obtained in step S4 as the fitness function of IHBA, and using the IHBA algorithm to optimize the process parameters, the optimal milling process parameters are obtained, thereby optimizing the GFRP milling process parameters.

[0023] In this embodiment, step S1 includes:

[0024] Step S1.1: Due to the actual machining process, the axial cutting depth a p and radial depth of cut a e The spindle speed N is determined by the machining process. Therefore, in this invention, the spindle speed N is selected. C Feed rate V f As experimental factors, the levels of each factor were reasonably set, and a two-factor, four-level full factorial experiment was designed. Constraints on each factor were set based on the GFRP milling conditions.

[0025] Spindle speed: N min ≤N C ≤N max ;

[0026] Feed rate: V min ≤V f ≤V max .

[0027] Where, N min Nmax It is the spindle speed N C Minimum and maximum values; V min V max It is the feed rate V f The minimum and maximum values.

[0028] Step S1.2: Conduct a milling experiment to measure the surface quality, delamination damage, and processing energy consumption of the workpiece after milling.

[0029] In step S2 of this embodiment, a set of three-dimensional surface roughness models M is obtained. S The methods include:

[0030] Step S2.1: Based on the experimental results, organize and obtain the three-dimensional surface roughness dataset D. S ={(x1,y1),(x2,y2),…,(x i ,y i ),…,(x n ,y n The dataset is normalized, with the original independent variables being X = {x1, x2, ..., x}. i ,…,x n The normalization method is as follows: The output variable is Y = {y1, y2, ..., y} i ,…,y n The normalization method is as follows: The dataset obtained after normalization transformation is D S ={(x g1 ,y g1 ),(x g2 ,y g2 ),…,(x gi ,y gi ),…,(x gn ,y gn )};where, x i and x ig These represent the initial values ​​of the input variables and the normalized input variables, respectively; y i and y ig , are the initial value of the output variable and the normalized output variable, respectively; Max(x) and Min(x) are the maximum and minimum values ​​of the original input variables, respectively; Max(y) and Min(y) are the maximum and minimum values ​​of the original output variables, respectively.

[0031] Step S2.2: Divide the normalized dataset into training set D using leave-one-out cross-validation. t =[D t1 D t2 ,…,D ti,…,D tn ] and test set D v =[D v1 D v2 ,…,D vi ,…,D vn ], where n is the total number of data points in the dataset.

[0032] Step S2.3: Set the number of input layer neuron nodes p, the number of hidden layer neuron nodes m, the number of output layer neuron nodes l, the learning rates η1, η2, and η3, and the loss function threshold E.

[0033] Step S2.4: Use training set D ti Training an RBFNN model M for three-dimensional surface roughness Si .

[0034] Step S2.5: The input layer directly inputs the training set D ti Input vectors are mapped to hidden layers.

[0035] Step S2.6: Initialize the hidden layer radial basis function centers u from the training set D ti Randomly select m training data points as the centers u of the radial basis functions of the hidden layer, u = {u1, u2, ..., u...} i ,…,u m} Where m is the number of radial basis function centers in the hidden layer, which is the same as the number of neurons in the hidden layer.

[0036] Step S2.7: Initialize the center width of the radial basis functions in the hidden layer, and calculate the width b at the center of the hidden layer neurons:

[0037]

[0038] b = [b1, b2, ..., b m ] T

[0039] Among them, i=1,2,…,m, j=1,2,…,m, i≠j; u i u j and b i These represent the selected i-th and j-th centers, and the i-th center u, respectively. i The corresponding width.

[0040] Step S2.8: Randomly initialize the initial weight matrix connecting the hidden layer neurons to the output layer neurons:

[0041]

[0042] Step S2.9: Set the radial basis function of the hidden layer neuron activation function to a Gaussian function, and calculate the output of the hidden layer neuron:

[0043]

[0044] h j (X)=[h 1j (X1),h 2j (X2),…,h (n-1)j (X n-1 )] T

[0045] Where i = 1, 2, ..., n-1, j = 1, 2, ..., m; h ij (X i ) represents the output of the j-th hidden layer neuron for the i-th training sample; h j (X) represents the output of the j-th hidden layer neuron; X i u represents the i-th input training data sample in the training set; j and b j These represent the center point and width of the j-th hidden layer neuron, respectively.

[0046] Step S2.10: Calculate the output of the neural network:

[0047]

[0048] Out = [Out1, Out2, ..., Out l ] T

[0049] Among them, Out k 'Out' and 'Out' represent the output value of the k-th output node and the output of the neural network, respectively.

[0050] Step S2.11: Calculate the loss function (Loss) of the neural network:

[0051]

[0052] Among them, Out ik and y ik These represent the output value and label value of the k-th output node for the i-th training data, respectively.

[0053] Step S2.12: Calculate the difference factor:

[0054] δ ik =Out ik -y ik

[0055] Step S2.13: Update the weights, center, and width between the hidden layer and the output layer:

[0056]

[0057] in, This represents the updated connection weight between the j-th hidden layer neuron and the k-th output layer neuron; and η1, η2, and η3 represent the center and width of the updated j-th hidden layer neuron, respectively; η1, η2, and η3 represent the learning rate, where η1, η2, and η3 > 0.

[0058] Step S2.14: Repeat steps S2.9-S2.13 until the change in the training error value Loss over 100 generations is less than the error threshold E, finally obtaining the RBFNN model M. Si .

[0059] Step S2.15: Based on the leave-one-out cross-validation principle, repeat steps S2.1-S2.14 n times for n training sets until all n training sets have undergone one round of training, thus obtaining the RBFNN model set for three-dimensional surface roughness, i.e., the three-dimensional surface roughness model set M. S ={M S1 M S2 ,…,M Si ,…,M Sn}

[0060] Based on the experimental results, the hierarchical factor dataset D was obtained. F and processing energy consumption dataset D E Then, the hierarchical factor model set M is obtained using the same method as in steps S2.1-S2.15. F And the set of processing energy consumption models M E .

[0061] In this embodiment, step S3 includes:

[0062] Step S3.1: Test set D v Calculate and validate the set of three-dimensional surface roughness models M S Fitting accuracy A c .

[0063] For the case where the output dimension l=1 of the RBFNN model in this invention, the formula for calculating the fitting accuracy is as follows:

[0064]

[0065] Among them, Out j y represents the output value predicted by the j-th RBFNN model for the test set. j This represents the label value of the j-th RBFNN model in the test set; This represents the average value of the label values ​​in the test set of the RBFNN model set.

[0066] Step S3.2: Using the same method as in step S3.1, calculate and verify the stratified factor model set M. F And the set of processing energy consumption models M E The fitting accuracy.

[0067] In this embodiment, step S4 includes:

[0068] Step S4.1: Take the three-dimensional surface roughness model set M respectively. S Stratified factor model set M F And the set of processing energy consumption models M E The RBFNN models with the smallest mean square error are identified and denoted as the three-dimensional surface roughness model RBFNN1, the layering factor model RBFNN2, and the processing energy consumption model RBFNN3, respectively.

[0069] Step S4.2: Calculate the weights using the coefficient of variation method:

[0070] The three-dimensional surface roughness y obtained from the experiment S Normalized to a positive indicator:

[0071]

[0072] Among them, y S ={y S1 ,…,y Si ,…,y Sn}, y Si For the three-dimensional surface roughness y S The i-th value; max(y S ) and min(y S ) represent the three-dimensional surface roughness y S The maximum and minimum values.

[0073] Calculate the coefficient of variation (CV) of three-dimensional surface roughness. S :

[0074]

[0075] The coefficient of variation (CV) of the stratification factor was calculated using the same method. F and the coefficient of variation (CV) of processing energy consumption E .

[0076] The weights ω corresponding to the three-dimensional surface roughness are calculated using the coefficient of variation method. S1 :

[0077]

[0078] The weights ω corresponding to the stratification factors were calculated using the same method. F1The weight ω corresponding to processing energy consumption E1 .

[0079] Step S4.3: Calculate the weights using the independence weighting method:

[0080] Establish three-dimensional surface roughness y S Regarding the stratification factor y F Multiple linear regression model for processing energy consumption:

[0081] y S =α S0 +α S1 y F +α S2 y E +ε S

[0082] make

[0083] α S =(A T A) -1 A T y S

[0084] Where, ε S ε F and ε E This is the error term.

[0085] Calculate the multiple correlation coefficient R of three-dimensional surface roughness S :

[0086]

[0087] in, For y S The estimate, for The i-th value in the vector.

[0088] Repeat the above steps to calculate the multiple correlation coefficient R between the stratification factor and processing energy consumption. F and R E .

[0089] The weights ω corresponding to the three-dimensional surface roughness are calculated using the independence weighting method. S2 :

[0090]

[0091] The weights ω corresponding to the stratification factors were calculated using the same method. F2 The weight ω corresponding to energy consumption E2 .

[0092] Step S4.4: Calculate the weights ω corresponding to the three-dimensional surface roughness model RBFNN1, the hierarchical factor model RBFNN2, and the processing energy consumption model RBFNN3 using the additive synthesis method. S ω F and ω E :

[0093] ω S =aω S1 +(1-a)ω S2

[0094] ω F =aω F1 +(1-a)ω F2

[0095] ω E =aω E1 +(1-a)ω E2

[0096] Where a is a constant, and in this embodiment a = 0.5.

[0097] Step S4.5: Establish a multi-objective comprehensive evaluation model RBFNN m :

[0098] RBFNN m =ω S RBFNN1+ω F RBFNN2+ω E RBFNN3.

[0099] In this embodiment, step S5 includes:

[0100] Step S5.1: Using the multi-objective comprehensive evaluation model established in step S4 as the fitness function of the IHBA algorithm, the IHBA algorithm is used to optimize the three-dimensional surface roughness, layering factor and machining energy consumption of the GFRP after milling, so as to obtain the optimal milling process parameters corresponding to the model when the optimal three-dimensional surface roughness, layering factor and machining energy consumption are achieved, thereby realizing the optimization of milling process parameters.

[0101] Step S5.2: Optimize process parameters using the IHBA algorithm. There are N honey badger individuals, and the number of iterations is set to T. Assume the honey badger population consists of X = {x} 1 ,x 2 Let the first dimension be the spindle speed N. C The size, the second dimension is the feed rate V f Size. Define the upper and lower bounds ub and lb for dimension one and dimension two.

[0102]

[0103] lb=[lb1,lb2], ub=[ub1,ub2];

[0104]

[0105] in, and Let ub1 and lb1 represent the values ​​of the i-th honey badger individual in the first and second dimensions, respectively; ub1 and lb1 represent the upper and lower bounds of the honey badger individual in the first dimension, respectively; and ub2 and lb2 represent the upper and lower bounds of the honey badger individual in the second dimension, respectively.

[0106] Step S5.3: Initialize the population. The calculation process for randomly initializing the population individuals is as follows:

[0107] x i =[lb1+r 01 (ub1-lb1),lb2+r 02 (ub2-lb2)]

[0108] Where, x i For the i-th honey badger individual, r 01 and r 02 It is a random number in the range [0,1].

[0109] Step S5.4: Substitute the initialized honey badger individuals into the multi-objective comprehensive evaluation model to obtain the population individual fitness values ​​Y = {y1, y2, ..., y N}

[0110] Step S5.5: Set the reward baseline for the Q-learning agent:

[0111]

[0112] Set the initial state s1 of the individual to 0.

[0113] Step S5.6: Initialization of the Q-learning agent:

[0114] Initialize the Q table:

[0115]

[0116] Where s∈{0,1}, s represents the state of fitness value; when s is 0, it means that the fitness corresponding to the current state has been improved; when s is 1, it means that the fitness corresponding to the current state has not been improved; a∈{0,1,2}, a represents action; when a is 0, it means that the action is mining behavior; when a is 1, it means that the action is honey gathering behavior; when a is 2, it means that the action is adaptive Gaussian mutation behavior.

[0117] Initialize the reward table:

[0118]

[0119] Step S5.7: Define odor intensity. The higher the odor intensity, the faster the search speed. The formula for calculating odor intensity I is as follows:

[0120]

[0121] S=(x i -x i+1 ) 2

[0122] d i =x prey -x i

[0123] Among them, I i The odor intensity represents the i-th honey badger individual at its current location; r1 is a random number between [0,1]; x i Indicates the position of the i-th honey badger, x prey Indicates the position of the best-fit individual in the current honey badger population; d i This represents the distance between the prey and the i-th individual honey badger.

[0124] Step S5.8: Calculate and update the density factor.

[0125] The density factor controls the random factors that change over time. The formula for calculating the density factor α is as follows:

[0126]

[0127] Where C represents a constant, and in this embodiment, C = 2; t represents the current iteration number of the honey badger population, and T represents the maximum iteration number.

[0128] Step S5.9: Train the agent based on Q-learning to achieve adaptive behavior policy selection:

[0129] Assumption Current exploration rate ε t The calculation formula is as follows:

[0130] ε t =ε0×(1-δ t )

[0131] Where ε0 represents the initial exploration rate, and we take ε0 = 0.25.

[0132] Set the prior probability distribution:

[0133] make These correspond to the preset prior probabilities of actions 0, 1, and 2 in the early, middle, and late stages, respectively.

[0134] Fitting prior probabilities using linear interpolation:

[0135]

[0136] p(a|δ t = [p0, p1, p2]

[0137] in, p0, p1, and p2 represent the probabilities of choosing action 0, action 1, and action 2, respectively.

[0138] Select an action:

[0139]

[0140] in, This indicates that based on the current probability p(a|δ) t Choose action 0, action 1, or action 2; This indicates that action A takes a value that follows The probability distribution; max a [Q(s t ,a)+λ·p(a|δ t ] indicates that in a given state s t and δ t Below, for all possible actions a∈{0,1,2}, the corresponding Q(s) t ,a)+λ·p(a|δ t The maximum value in ) ; argmax a [Q(s t ,a)+λ·p(a|δ t )] represents Q(s t ,a)+λ·p(a|δ t The action corresponding to the maximum value; λ is a constant, and in this embodiment, λ = 0.2; r represents a random number between [0, 1].

[0141] Step S5.10: According to a t Choose a behavioral strategy, if a t If a is 0, proceed with the mining action in step S5.11. t If a is 1, then proceed with the honey-gathering action in step S5.12. t If the result is 2, then the adaptive Gaussian mutation behavior in step S5.13 will be performed.

[0142] And determine the direction F of the digging and honey-gathering behaviors based on the random number r between [0,1].

[0143]

[0144] Step S5.11: Mining behavior:

[0145] In digging mode, honey badgers primarily rely on the scent of honey to locate it. Using both digging and their sense of smell, they move in a heart-shaped pattern around the honey, digging in suitable spots. The formula for updating an individual honey badger's location in digging mode is as follows:

[0146] x inew =x prey +F·β·x prey +F·r2·α·d i ·|cos(2πr3)·(1-cos(2πr4))|

[0147] Where β is a constant greater than or equal to 1, with a default empirical value of 6; r2, r3, and r4 are all random numbers between [0,1], and r2, r3, and r4 are all distinct. Then proceed to step S5.14.

[0148] Step S5.12: Honey gathering behavior:

[0149] In the honeyguide bird foraging mode, the honey badger follows the honeyguide bird to find the honey; the formula for updating the individual location of the honey badger in the honeyguide bird foraging mode is as follows:

[0150] x inew =x prey +F·r5·α·d i

[0151] Where r5 is a random number between [0,1]. Then proceed to step S5.14.

[0152] Step S5.13: Adaptive Gaussian mutation behavior:

[0153] Adaptive Gaussian mutation is performed on the locations of individual honey badgers. The standard deviation of the adaptive Gaussian mutation and the locations of individual honey badgers are then updated and calculated.

[0154]

[0155] x inew =x i +x i Gaussion(σ)

[0156] Where σ represents the adaptive Gaussian variability standard deviation; σ max This represents the standard deviation of the maximum Gaussian variation; in this embodiment, σ is taken as the standard deviation. max =0.5; k1 and k2 are constants, and in this embodiment, k1 = 30 and k2 = 8; x i Let x represent the position of the i-th individual honey badger.inew This represents the updated position of the i-th honey badger individual; Gaussian(σ) represents a randomly generated random number that follows a Gaussian distribution. Then proceed to step S5.14.

[0157] Step S5.14: Calculate the immediate reward and update the reward baseline:

[0158]

[0159] b t+1 =(1-α) b )·b t +α b ·|y i |

[0160] Where, α b This represents the moving average coefficient, which is α in this embodiment. b =0.1; |y i | represents the absolute value of the fitness of the i-th individual honey badger.

[0161] Step S5.15: Update the reward table:

[0162] R t+1 (s,a)=(1-α R )·R t (s,a)+α R ·r t

[0163] Where, α R In this embodiment, α is taken as the learning rate coefficient. R =0.2.

[0164] Step S5.16: Update the Q-table based on the Bellman equation.

[0165] Determine the new state s t ′:

[0166] If y prey -y inew If the value is less than 0, it indicates a state improvement, and the new state s is... t ' is recorded as 0, otherwise the new state s is... t ′ is denoted as 1.

[0167] Calculate the Q-value and learning rate α t :

[0168]

[0169] Update the Q value based on the Bellman equation:

[0170] Q(s t ,a t)=Q(s t ,a t )+α t ·[r t +γ·max a Q(s t ′,a)-Q(s t ,a t )]

[0171] Where, α t α0 and α0 represent the learning rate for the current Q value and the initial learning rate, respectively. In this embodiment, α0 = 0.4; r t Indicates the immediate reward for the current state and action; y inew and y prey Let Q(s) represent the updated fitness value of the i-th honey badger individual and the globally optimal fitness value, respectively; γ represents the discount factor, which is 0.8 in this embodiment; Q(s) t ,a t ) indicates that the state is s t Action is a t Q value under; max a Q(s t ′,a) represents the new state as s t The maximum value of Q under any action at time ′.

[0172] State update: Updates the current state to a new state for the next individual.

[0173] s t =s t ′

[0174] Step S5.17: When the number of iterations t > T, output the optimal honey badger position x. prey and its fitness value y prey This yields the solution with the optimal fitness value.

[0175] In this embodiment, a 10mm four-flute diamond-coated carbide end mill was selected as the milling tool. The GFRP plate with a diameter of l×d×h = 150×100×10mm was milled using a VMC vertical machining center at Shenyang Machine Tool Plant. In actual machining, the axial and radial depths of cut are determined by the machining process. Therefore, in this experiment, the axial and radial depths of cut were set to constant values. Specifically, the axial depth of cut was 10mm, and the radial depth of cut was 1mm. Four levels of spindle speed and feed rate were used, and the full factors are shown in Tables 1 and 2. The tool parameters used in this invention are shown in Table 3.

[0176] Table 1. Factors and Levels in the Full Factorial Experiment Design

[0177]

[0178] Table 2. Full Factorial Experiment Design

[0179]

[0180] Table 3 Carbide End Mill Tool Parameters

[0181]

[0182] According to the experimental arrangement in Table 2, groove milling was performed on GFRP sheets. The methods for measuring three-dimensional surface roughness, delamination factor, and machining energy consumption were as follows: After milling based on a set of process parameters, the three-dimensional morphological height of the milled stable region was measured using a Micromesure 2STIL white light confocal three-dimensional profilometer, with a sampling area of ​​3.3mm × 3.3mm, to obtain the three-dimensional surface roughness Sa. Surface morphology images of the milled grooves on the GFRP workpiece were acquired using a high-resolution industrial camera under illumination conditions. Image calibration was performed using ImageJ software combined with a microscopic scale, and finally, the maximum damage width was used as the quantitative characterization index of the delamination factor. Machining energy consumption was collected by an LCDG-DG3-71760TH intelligent data acquisition unit paired with a Zigbee gateway master station. Table 4 presents the data from the GFRP milling experiment.

[0183] Table 4 GFRP Milling Experiment Data

[0184] RBFNN model parameter settings

[0185] Through repeated experiments, parameters that maximized the fitting accuracy of the three-dimensional surface roughness, layering factor, and processing energy consumption model sets were selected for model training. The final fitting accuracies of each model reached 90.1%, 89.3%, and 95.2%, respectively. The RBFNN model parameter settings are shown in Table 5. In this invention, the RBFNN model is a three-layer neural network model with two neurons in the input layer and one neuron in the output layer.

[0186] Table 5 RBFNN Model Parameter Table

[0187]

[0188]

[0189] Comparison of algorithm optimization results before and after improvement

[0190] With the relevant parameter settings of the HBA algorithm before and after the improvement being the same, the optimal fitness values ​​and optimal process parameters obtained by running the RBFNN-HBA algorithm 100 times before and after the improvement, using the established multi-objective comprehensive evaluation model as the fitness function, are shown in Tables 6 and 7. The experimental results of three-dimensional surface roughness, layering factor, and machining energy consumption obtained from milling verification experiments based on the optimal process parameters are shown in Tables 6 and 7.

[0191] Table 6 shows the optimization results and verification experiment results of the unimproved RBFNN-HBA algorithm after 100 runs.

[0192]

[0193] Table 7 shows the optimization results and verification experiment results of the improved RBFNN-IHBA algorithm after 100 runs.

[0194]

[0195] Optimization Result Analysis:

[0196] As shown in Tables 6 and 7, the improved RBFNN-IHBA algorithm significantly outperforms the unimproved RBFNN-HBA algorithm. Furthermore, the experimental results obtained based on the optimal milling process parameters are all superior compared to those in Table 4, demonstrating the effectiveness of the proposed RBFNN-IHBA algorithm in the multi-objective process parameter optimization problem of GFRP milling.

[0197] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0198] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0199] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0200] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method of optimizing glass fiber reinforced composite material milling process parameters, characterized in that, Comprise the following steps: Step S1: the three-dimensional surface roughness of the glass fiber reinforced composite material milling, layering factor, processing energy consumption as the optimization goal, with the spindle speed and feed speed as the optimization variable, build full factor experiment scheme; Step S2: According to the experimental results, the three-dimensional surface roughness dataset D S , the layering factor dataset D F and the machining energy consumption dataset D E is obtained ; the dataset D S is divided into a training set D t and a test set D v according to the leave-one-out cross-validation method; RBFNN models are trained respectively through n training sets D t1 ,D t2 ,…,D tn , and a three-dimensional surface roughness model set M S is obtained by combination; the same method is used to obtain a layering factor model set M F and a machining energy consumption model set M E ; Step S3: Verify the fitting accuracy of the set of three-dimensional surface roughness models M by testing set D v Compute and verify the fitting accuracy of the set of three-dimensional surface roughness models M S ; use the same method to compute and verify the fitting accuracy of the set of hierarchical factor models M F and the set of machining energy consumption models M E ; Step S4: according to the experimental results, the objective weight corresponding to the three-dimensional surface roughness, layering factor and processing energy consumption is calculated by the coefficient of variation method and the independence weight method respectively, and the weight corresponding to the three-dimensional surface roughness model, layering factor model and processing energy consumption model is obtained by the additive synthesis method, so as to obtain the multi-objective comprehensive evaluation model; Step S5: the multi-objective comprehensive evaluation model obtained in step S4 is used as the fitness function of IHBA, and the IHBA algorithm is used to optimize the process parameters, so as to obtain the optimal milling process parameters, and realize the optimization of the milling process parameters of the glass fiber reinforced composite material.

2. The glass fiber reinforced composite material milling process parameter optimization method according to claim 1, characterized in that, Step S1 includes: Step S1.1: Selecting spindle speed N C , feed speed V f The experimental factors are designed as two factors and four levels. The milling conditions of the glass fiber reinforced composite material are set as the constraints of each factor: Spindle speed: N min ≤ N C ≤ N max ; Feed speed: V min ≤ V f ≤ V max ; where N min , N max are the minimum and maximum values of the spindle speed N C ; V min , V max are the minimum and maximum values of the feed speed V f ; Step S1.2: carry out milling experiment, measure the surface quality of the workpiece after milling, layering damage and processing energy consumption.

3. The method of claim 1, wherein the glass fiber reinforced composite milling process parameter optimization method is characterized by, In step S2, a set M of three-dimensional surface roughness models is obtained S The method comprises: Step S2.1: According to the experimental results, the three-dimensional surface roughness dataset D is obtained S = {(x1, y1), (x2, y2), …, (x i ,y i ), …, (x n ,y n )}; The original independent variable is X = {x1, x2, …, x i ,…,x n}, and the normalization method is: The output variable is Y = {y1, y2, …, y i ,…,y n}, and the normalization method is: The normalized dataset is D S = {(x g1 ,y g1 ), (x g2 ,y g2 ), …, (x gi ,y gi ), …, (x gn ,y gn )}; Where x i and x ig are the initial value of the input variable and the input variable after normalization, respectively; y i and y ig are the initial value of the output variable and the output variable after normalization, respectively; Max(x) and Min(x) are the maximum value of the original input variable and the minimum value of the original input variable, respectively; Max(y) and Min(y) are the maximum value of the original output variable and the minimum value of the original output variable, respectively; Step S2.2: Dividing the normalized dataset into training sets D t = [D t1 , D t2 ,..., D ti ,..., D tn ] and testing sets D v = [D v1 , D v2 ,..., D vi ,..., D vn ], n is the total number of data in the dataset; Step S2.3: set the number of input layer neurons p, the number of hidden layer neurons m, the number of output layer neurons l, the learning rate η1, η2, η3, and the loss function threshold E; Step S2.4: Using the training set D ti , train the RBFNN model M of the three-dimensional surface roughness Si ; Step S2.5: The input layer directly maps the training set D ti The input vector is mapped to the hidden layer; Step S2.6: initialize the hidden layer radial basis function centers u from the training set D ti m training data are randomly selected as the hidden layer radial basis function centers u, u = {u1, u2, …, u i ,…,u m}; wherein m is the number of hidden layer radial basis function centers, which is the same as the number of hidden layer neurons. Step S2.7: initialize the width of the hidden layer radial basis function center, calculate the width b of the hidden layer neuron center: b = [b1, b2,..., b m ] T wherein, i = 1, 2, …, m, j = 1, 2, …, m, i≠j; u i , u j and b i respectively represent the selected i-th and j-th centers and the i-th center u i corresponding width; Step S2.8: randomly initialize the initial weight matrix of the hidden layer neuron connecting the output layer neuron: Step S2.9: set the hidden layer neuron activation function radial basis function as Gaussian function, calculate the output of the hidden layer neuron: h j (X) = [h 1j (X1),h 2j (X2),…,h (n-1)j (X n-1 )] T where i = 1, 2,..., n-1, j = 1, 2,..., m; h ij (X i ) represents the output of the jth hidden layer neuron for the ith training sample; h j (X) represents the output of the jth hidden layer neuron; X i represents the ith input training data sample of the training set; u j and b j represent the center point and width of the jth hidden layer neuron, respectively; Step S2.10: calculate the output of the neural network: Out = [Out1, Out2,..., Out l ] T wherein Out k and Out respectively denote the kth output node output value and the output of the neural network; Step S2.11: calculate the loss function Loss of the neural network: wherein Out ik and y ik respectively represent the output value and the label value of the kth output node for the ith training data. Step S2.12: calculate the difference factor: delta ik = Out ik minus ik Step S2.13: update the weight, center and width between the hidden layer and the output layer: wherein, denotes the updated connection weight between the jth hidden layer neuron and the kth output layer neuron; and denote the updated center and width of the jth hidden layer neuron, respectively; η1, η2, η3denote the learning rate, η1, η2, η3>

0. Step S2.14: Repeat steps S2.9-S2.13 until the change in the training error value Loss 100 over the last 100 generations is less than the error threshold E, resulting in the final RBFNN model M Si ; Step S2.15: According to the principle of leave-one-out cross-validation, steps S2.1-S2.14 are repeated n times for n training sets until n training sets are trained for one round, obtaining the RBFNN model set of three-dimensional surface roughness, i.e. the three-dimensional surface roughness model set M S = {M S1 , M S2 , …,M Si , …,M Sn} According to the experimental results, the hierarchical factor dataset D is obtained F and the processing energy consumption dataset D E Then, the same method as steps S2.1-S2.15 is used to obtain the hierarchical factor model set M F and the processing energy consumption model set M E .

4. The method of claim 1, wherein, Step S3 includes: Step S3.1: Compute the set M of three-dimensional surface roughness models by testing set D v Compute and verify the fitting accuracy A of the set M of three-dimensional surface roughness models S c ;​ For the case that the output dimension of RBFNN model is l=1, the calculation formula of fitting accuracy is as follows: where Out j represents the output value predicted by the jth RBFNN model for the test set, y j represents the label value of the test set for the jth RBFNN model; represents the average value of the label values of the test set for the RBFNN model set; Step S3.2: Calculate and verify the fitting accuracy of the set of hierarchical factor models M using the same method as in step S3.1 F and the set of machining energy consumption models M E .

5. The method of claim 1, wherein, Step S4 includes: Step S4.1: Take the RBFNN model with the minimum error in the three-dimensional surface roughness model set M S , the layered factor model set M F and the machining energy consumption model set M E respectively, and mark them as three-dimensional surface roughness model RBFNN1, layered factor model RBFNN2 and machining energy consumption model RBFNN3 respectively; Step S4.2: calculate the weight by using the coefficient of variation method: The three-dimensional surface roughness y obtained from the experiment is normalized as S Normalized to positive indicators: where y S = {y S1 ,…,y Si ,…,y Sn}, y Si is the i-th value of the three-dimensional surface roughness y S ; max(y S ) and min(y S ) are the maximum and minimum values of the three-dimensional surface roughness y S , respectively. Computing the coefficient of variation CV of a three-dimensional surface roughness S : The variation coefficient CV of the delamination factor and the variation coefficient CV of the machining energy consumption are calculated respectively by using the same method F and the variation coefficient CV of the machining energy consumption E ; The weight ω corresponding to the three-dimensional surface roughness is calculated by using the coefficient of variation method S1 : The weight ω corresponding to the delamination factor is calculated by the same method F1 The weight ω corresponding to the machining energy consumption is calculated by the same method E1 ; Step S4.3: calculate the weight by using the independence weight method: Establishing a three-dimensional surface roughness y S Regarding the layering factor y F And a multiple linear regression model of machining energy consumption: y S = a S0 + a S1 y F + a S2 y E + ε S Let a S = (A T A) -1 A T y S where ε S , ε F , and ε E are error terms; Computing the complex correlation coefficient R of a three-dimensional surface roughness S : wherein, for y S the estimate of is the i-th value in the vector; The above steps are repeated to calculate the hierarchical factor and the complex correlation coefficient R of the processing energy consumption F and R E ; The weight ω corresponding to the three-dimensional surface roughness is calculated by using the independence weight method S2 : The weight ω corresponding to the delamination factor is calculated by the same method F2 and the weight ω corresponding to the energy consumption E2 ; Step S4.4: Calculate the weights ω of the three-dimensional surface roughness model RBFNN1, the layering factor model RBFNN2 and the machining energy consumption model RBFNN3 by using the additive synthesis method S , ω F and ω E : ω S = aω S1 + (1 - a)ω S2 ω F = aω F1 + (1 - a)ω F2 ω E = a ω E1 + (1 - a) ω E2 Wherein, a is a constant; Step S4.5: Establishing multi-objective comprehensive evaluation model RBFNN m : RBFNN m = ω S RBFNN1+ ω F RBFNN2+ ω E RBFNN3.

6. The glass fiber reinforced composite material milling process parameter optimization method according to claim 1, characterized in that, Step S5 includes: Step S5.1: the multi-objective comprehensive evaluation model established in step S4 is used as the fitness function of IHBA algorithm, and IHBA algorithm is used to optimize the three-dimensional surface roughness, layering factor and processing energy consumption of the glass fiber reinforced composite material after milling, so as to obtain the best milling process parameters corresponding to the model under the best three-dimensional surface roughness, layering factor and processing energy consumption, so as to realize the optimization of the milling process parameters; Step S5.2: process parameter optimization is performed by using the IHBA algorithm; there are N honey badger individuals, and the number of iterations is set to T; it is assumed that the honey badger population individuals are X={x 1 ,x 2}, the first dimension is the size of the main shaft speed N C , and the second dimension is the size of the feed speed V f ; the upper and lower bounds ub, lb of the dimensions one and two are defined; lb=[lb1,lb2],ub=[ub1,ub2]; wherein, and respectively represent the value of the i-th individual meerkat on the first and second dimensions; ub1 and lb1 respectively represent the upper and lower bounds of the first dimension for the meerkat individual; ub2 and lb2 respectively represent the upper and lower bounds of the second dimension for the meerkat individual; Step S5.3: initialize the population; the calculation process of randomly initializing the population is as follows: x i = [lb1 + r 01 (ub1 - lb1), lb2 + r 02 (ub2 - lb2)] where x i is the ith individual meerkat, r 01 and r 02 are random numbers in [0, 1]; Step S5.4: Substitute the initialized meerkat individual into the multi-objective comprehensive evaluation model to obtain the population individual fitness value Y = {y1, y2, …, yN}. N} Step S5.5: set the Q-learning agent reward benchmark: Set the initial state s1 of the individual to 0; Step S5.6: Q-learning agent initialization: Initialize Q table: Wherein, s∈{0,1}, s represents the state of fitness value; when s is 0, it indicates that the fitness corresponding to the current state is improved; when s is 1, it indicates that the fitness corresponding to the current state is not improved; a∈{0,1,2}, a represents action; when a is 0, it indicates that the action is digging behavior; when a is 1, it indicates that the action is honey gathering behavior; when a is 2, it indicates that the action is adaptive Gaussian variation behavior; Initialize the reward table: Step S5.7: define the smell intensity; the higher the smell intensity, the faster the search speed, and the calculation formula of the smell intensity I is as follows: S = (x i - x i+1 ) 2 d i = x prey - x i wherein I i represents the smell intensity corresponding to the current position of the i-th individual; r1 is a random number between [0, 1]; x i represents the position of the i-th individual meerkat, x prey represents the position of the best fitness individual in the current meerkat population; d i represents the distance between the prey and the i-th meerkat individual; Step S5.8: calculate and update the density factor; The density factor controls the random factor changing with time, and the calculation formula of the density factor α is as follows: Wherein, C represents a constant; t represents the current number of iterations of the meerkat population, and T represents the maximum number of iterations; Step S5.9: train the agent to realize adaptive behavior strategy selection based on Q-learning: Assume Current exploration rate ε t The calculation formula is as follows: ε t = ε0x (1 - δ t ) Wherein, ε0 represents the initial exploration rate; Set the prior probability distribution: Let P (0), P (1), and P (2) are preset prior probabilities corresponding to the early, middle, and late actions 0, 1, and 2, respectively; Fit the prior probability by linear interpolation: p(a | δ t ) = [p0, p1, p2] wherein, p0, p1, and p2 represent the probabilities of selecting action 0, action 1, and action 2, respectively; Perform action selection: wherein, represents selecting action 0, action 1 or action 2 according to the current probability p(a|δ t ); represents that action A takes a value subject to a probability distribution; max a [Q(s t ,a)+λ·p(a|δ t )] represents the maximum value among Q(s t ,a)+λ·p(a|δ t ) corresponding to all possible actions a∈{0,1,2} given state s t and δ t ; argmax a [Q(s t ,a)+λ·p(a|δ t )] represents the action corresponding to the maximum value of Q(s t ,a)+λ·p(a|δ t ); λ is a constant; r represents a random number between [0,1]. Step S5.10: According to a t selecting a behavior policy, if a t = 0, excavating behavior of step S5.11 is performed, if a t = 1, honey collecting behavior of step S5.12 is performed, if a t = 2, adaptive Gaussian variation behavior of step S5.13 is performed; And determine the direction F of digging behavior and honey gathering behavior according to the random number r between [0,1]: Step S5.11: digging behavior: In the digging mode, the formula for updating the individual position of the meerkat is as follows: x inew = x prey + F · β · x prey + F · r2· α · d i · |cos(2πr3) · (1 - cos(2πr4))| Wherein, β is a constant greater than or equal to 1; r2, r3 and r4 are random numbers between [0,1], and r2, r3 and r4 are different from each other; then enter step S5.14; Step S5.12: honey gathering behavior: In the honey guide bird honey gathering mode, the formula for updating the individual position of the meerkat is as follows: x inew = x prey + F · r5 · a · d i Wherein, r5 is a random number between [0,1]; then enter step S5.14; Step S5.13: adaptive Gaussian variation behavior: Adaptive Gaussian variation is performed on the individual position of the meerkat, and the adaptive Gaussian variation standard deviation and the individual position of the meerkat are updated and calculated: x inew = x i + x i · Gaussion (σ) wherein σ denotes an adaptive Gaussian variation standard deviation; σ max denotes a maximum Gaussian variation standard deviation; k1, k2 are constants; x i denotes a position of an i-th individual meerkat, x inew denotes an updated position of the i-th meerkat individual; Gaussion(σ) denotes a random number generated randomly and satisfying a Gaussian distribution; and then entering step S5.14; Step S5.14: calculate the immediate reward and update the reward benchmark: b t+1 = (1 - a b ) · b t + a b · |y i | wherein a b represents a moving average coefficient; |y i | represents the absolute value of the fitness of the ith meerkat individual; Step S5.15: update the reward table: R t+1 (s,a) = (1 - a R ) · R t (s,a) + a R · r t wherein a R is a learning rate coefficient; Step S5.16: update the Q table based on the Bellman equation; determining a new state s t ′: If y prey - y inew < 0, then the state is improved, and the new state s t ′ is set to 0, otherwise the new state s t ′ is set to 1. Computing Q-value learning rate a t : Update the Q value based on the Bellman equation: Q(s t ,a t ) = Q(s t ,a t ) + a t · [r t + g · max a Q(s t ', a) - Q(s t , a t )] wherein a t and a0represent the learning rate and the initial learning rate of the current Q value, respectively; r t represents the immediate reward of the current state and action; y inew and y prey represent the fitness value of the i-th meerkat individual after updating and the global optimal fitness value, respectively; g represents the discount factor; Q(s t ,a t ) represents the Q value under the state s t and the action a t ; max a Q(s t ',a) represents the maximum value of the Q value under any action when the new state is s t '. State update, update the current state to the new state for the next individual: s t = s t ' Step S5.17: output the best aardvark position x when the iteration number t > T prey and its fitness value y prey , the solution with the best fitness value.

Images