An optimization method for laser cleaning parameters based on multi-objectives
Through the multi-objective parallel GA-BP neural network based on the optimal hidden layer and the improved multi-objective sparrow algorithm, the laser cleaning parameters are solved, and the multi-objective model modeling difficulties and local minimum value problems are achieved. Efficient laser cleaning effect meets the surface quality and corrosion resistance requirements of aviation aluminum alloy materials.
Patent Information
- Application Number
- CN202310727588.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-19
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2043-06-19
AI Technical Summary
The multi-objective model of existing laser cleaning process parameters is difficult to model, and it is easy to fall into the problem of local minimum values, resulting in poor cleaning results and difficult to meet the high requirements of aviation aluminum alloy materials for surface quality and corrosion resistance.
Using multi-objective parallel GA-BP neural network based on the optimal hidden layer and the improved multi-objective sparrow algorithm, the laser cleaning parameters are optimized, combined with the entropy weight-Topsis method Select the optimal solution to improve cleaning efficiency and quality.
It effectively solves the problem of local minimum value, improves the prediction accuracy and efficiency of laser cleaning, meets the requirements of aviation aluminum alloy materials for surface quality and corrosion resistance, and reduces cleaning costs.
Smart Images

Figure CN116956708B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of laser cleaning, and in particular to a laser cleaning parameter prediction method based on a multi-objective parallel GA-BP neural network with an optimal hidden layer and an improved multi-objective sparrow algorithm. Background Art
[0002] Although traditional cleaning technology in the industrial field still occupies the mainstream position, its disadvantages have become more and more obvious as people's environmental awareness increases. Laser cleaning is a cleaning method that removes the interaction between metal workpieces and their surface contaminants through laser high-energy beam processing. It has the advantages of high cleaning accuracy, green environmental protection, and non-contact. Aviation aluminum alloy materials are widely used as aerospace substrates. In order to meet its safety performance requirements, it is necessary to avoid damaging the aluminum alloy substrate as much as possible while removing the anodized film on its surface. This puts higher requirements on the cleaning rate of the anodized layer on the surface, the surface quality after cleaning, and the corrosion resistance. Therefore, it is necessary to coordinate and optimize the cleaning process parameters and select the appropriate cleaning parameter configuration to enable the substrate surface to achieve better performance requirements. Usually, the optimal solution of one target is accompanied by the inferior solution of another or more targets; at the same time, the laser cleaning process is affected by multiple parameter configurations, and there are obvious differences in the cleaning mechanism of removing oxide films between its energy parameters and motion parameters, which makes it difficult to model the multi-objective model of laser cleaning process parameters. Although the BP neural network has a good fitting function, it is easy to fall into local minima and other disadvantages. Summary of the invention
[0003] The present invention proposes an optimization method based on multi-objective laser cleaning parameters, which aims to solve the problems of existing parameter prediction methods, such as difficulty in modeling a multi-objective model of laser cleaning process parameters and easy falling into local minima.
[0004] Technical solution: A multi-objective laser cleaning parameter optimization method, the steps are:
[0005] Step 1, cleaning the oxide film on the surface of aluminum alloy and constructing a theoretical fitting model for laser cleaning;
[0006] Step 2, using BBD response surface design experiment and using different laser cleaning parameters designed by it to test the aluminum alloy oxide film, obtain the data of multi-objective response indicators after cleaning, and construct the RSM model through the data of laser cleaning parameters and response indicators;
[0007] Step 3, using the RANDI function in the matlab software to randomly generate N groups of laser cleaning parameters within the upper and lower limits of the laser cleaning parameters of the response surface test design, and using the RSM model in step 2 to predict the response index values corresponding to the N groups of cleaning parameters to obtain the data of the response index;
[0008] Step 4: Use the data of laser cleaning parameters and response indexes obtained in step 2 and step 3 as data sets for training the BP neural network, respectively construct multiple single-output layer BP neural networks of the same number of response indexes, and apply genetic algorithm to optimize them. At the same time, optimize the number of hidden layer nodes of the network, complete the construction of a multi-objective parallel GA-BP neural network based on the optimal hidden layer, so that it can predict the nonlinear fitting relationship between the laser cleaning parameters and the response index, and test the model prediction accuracy to obtain a qualified GA-BP neural network;
[0009] Step 5: Based on the actual needs of laser cleaning, the GA-BP neural network is used as a prediction model to construct the objective function and constraint conditions for laser cleaning parameter prediction optimization;
[0010] Step 6: Use the improved sparrow multi-objective algorithm to iteratively solve the objective function and constraints established in step 5, and finally obtain a set of excellent solutions for laser cleaning parameters. Then, the optimal solution in the excellent solution set for laser cleaning parameters is selected through entropy weight-Topsis as the prediction parameter for laser cleaning.
[0011] Furthermore, the theoretical fitting model of laser cleaning in step 1 is: Y(sa / O c / HV / CA / U / G)=F(P / f / v / v T / W / D).
[0012] Furthermore, the RSM model in step 2 is:
[0013] Sa=2.8804-9.9178×10 -1 *f+7.3633×10 -3 *P-1.4828×10 -4 *v+1.4933×10 -3 *P*f+6.1333×10 -5 *f*v-1.6978×10 -6 *P*v+7.6222×10 -2 *f 2 +1.68×10 -5 *P 2 -3.8889×10 -8 *v 2
[0014] O c =5.9241+18.2159*f-0.2919*P+4.2816×10 -3 *v+0.0298*P*f-1.7333×10 -4 *f*v+4.2667×10-6 *P*v-2.73185*f 2 +2.3882×10 -4 *P 2 -1.1185×10 -6 *v 2
[0015] CA=380.2944-113.950*f-1.3832*P+0.0455*v-0.1084*P*f-7.42220×10 -3 *f*v-8.44440×10 -5 *P*v+17.8222*f 2 +4.96×10 -3 *P 2 +1.5111×10 -6 *v 2 .
[0016] Furthermore, the input data of the GA-BP neural network in step 4 is:
[0017] The output data is:
[0018] Fitness as:
[0019] The weights and thresholds are:
[0020] Furthermore, the objective function and constraints of the GA-BP neural network in step 5 are:
[0021] F(f,p,v)=(max{CA},min{sa},{Oc})
[0022]
[0023] Furthermore, the steps of the improved sparrow multi-objective algorithm in step 6 are as follows:
[0024] a. Initialize the population through the good point set to obtain the initial population;
[0025] b. Perform non-dominated sorting on the fitness of individual sparrows in the initial population and introduce the judgment condition when the guard updates the position And update the judgment conditions of the current global optimal position Enable the algorithm to handle multi-objective problems;
[0026] c. After the non-dominated sorting, for individuals with the same non-dominated level, the screening comparison crowding operator is introduced to sort the crowding of the individuals with the same non-dominated level to obtain the initial sorting solution set;
[0027] d. Some sparrows are introduced as discoverers with adaptive safety threshold and normal distribution weight factor and perform discoverer position update; e. Follower position is updated according to the classic sparrow algorithm;
[0028] f. Randomly select some sparrows as guards and introduce improved step control parameters to implement the guard update method;
[0029] g. Basis Determine whether the updated position is better than the current position. If so, update the position, otherwise keep the current position;
[0030] h. Determine whether the iteration end condition is reached. If so, end the algorithm and output the optimal solution set. Otherwise, return to step b and output the excellent solution set after the iteration ends.
[0031] i. Apply entropy weight-Topsis to select the best solution from the set of optimal solutions.
[0032] Furthermore, the congestion calculation formula in step c is: Furthermore, the iterative formula for position update in step d is: Among them, AST stands for adaptive security threshold, and its expression is as follows: NDW represents the weight of the normal distribution operator optimization, and its expression is as follows:
[0033]
[0034] Furthermore, the iterative formula for updating the guard in step f is:
[0035]
[0036] Furthermore, the excellent solution set output in step h is: The corresponding response indicator set is: Beneficial effects:
[0037] Compared with the traditional laser cleaning optimization method, the present invention uses a multi-objective parallel GA-BP neural network based on the optimal hidden layer to predict the laser cleaning results, which improves the problems of the BP neural network with many iterations, inaccurate predictions, and easy to fall into the local optimal solution. At the same time, the improved multi-objective sparrow algorithm is used to optimize the multiple objectives of laser cleaning, which solves the problems of the original algorithm's own poor initial population quality, easy to fall into the local optimal solution, small population diversity in the later iteration, and low local development ability. The introduction of a multi-objective search mechanism solves the cleaning problem that is difficult to optimize multiple objectives, and improves the efficiency of optimization. In order to meet the minimum oxygen content and surface roughness of the material surface after cleaning, and the maximum static contact angle, the improved multi-objective sparrow algorithm is used to perform multi-objective optimization on the above model to obtain an excellent solution set of laser cleaning parameters, and the most suitable cleaning parameters are selected from the solution set through entropy weight-Topsis as the optimal solution of laser cleaning oxide film parameters. It helps to reduce cleaning costs while improving the cleaning quality of laser cleaning. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is the principle flow chart of the present invention;
[0039] Figure 2 It is the schematic diagram of the test device structure;
[0040] Figure 3 It is a multi-objective parallel GA-BP neural network training flow chart based on the optimal hidden layer;
[0041] Figure 4 It is the fitness value iteration diagram of the GA-BP network under different numbers of hidden layer nodes;
[0042] Figure 5 It is the relationship diagram between the prediction and actual value of the multi-objective parallel GA-BP neural network based on the optimal hidden layer;
[0043] Figure 6 is a randomly generated initial population graph;
[0044] Figure 7 It is the principle diagram of the improved multi-objective sparrow algorithm based on good point set and adaptive normal distribution weight;
[0045] Figure 8 The initial population graph generated for applying the good point set optimization;
[0046] Fig. 9 Schematic diagram of crowding operator for crowding comparison and screening;
[0047] Fig.10 It is the iteration curve graph of AST and NDW;
[0048] Fig.11This is the excellent solution diagram of laser cleaning parameters;
[0049] Fig.12 It is the entropy weight-Topsis principle diagram;
[0050] Fig.13 The morphology, element content, surface roughness, and hydrophobicity of the samples before and after optimization. DETAILED DESCRIPTION
[0051] The examples of the present invention are described in detail below with reference to the accompanying drawings.
[0052] In this field, an excellent solution to one objective is usually accompanied by a poor solution to another or multiple objectives. In view of this situation, a multi-objective sparrow algorithm based on a good point set and an adaptive normal distribution weight is selected to optimize the parameters of laser cleaning. At the same time, in order to ensure the accuracy of the laser cleaning model, a multi-objective parallel GA-BP neural network based on the optimal hidden layer is selected to fit the cleaning model. The BP neural network has a good fitting function, but it is easy to fall into local minima and other shortcomings. The GA-BP model obtained by optimizing the weights and thresholds of the BP neural network using a genetic algorithm is more accurate than the BP model in predicting the laser cleaning model. At the same time, the multi-objective sparrow algorithm based on a good point set and an adaptive normal distribution weight can be used to seek an excellent solution set for multiple objectives for the cleaning parameters, and the entropy weight-Topsis is applied in the solution set to select the most suitable cleaning parameters, which will improve the cleaning efficiency, reduce the cleaning cost, and better guide the application of laser cleaning in the aviation industry.
[0053] The present invention proposes a multi-objective parallel GA-BP neural network based on the optimal hidden layer and an improved multi-objective sparrow algorithm laser cleaning prediction model and parameter optimization method. The principle flow chart is as follows: Figure 1 As shown, the steps are:
[0054] Step 1, cleaning the oxide film on the surface of the aluminum alloy, and determining the optimization target of the laser cleaning process parameters and the multi-target response index to be optimized;
[0055] It is determined to use a pulsed fiber laser to conduct a laser cleaning test on the oxide film on the surface of the aluminum alloy, to construct a theoretical fitting model for laser cleaning, and to determine the optimization targets of the laser cleaning process parameters and the multi-objective response indicators to be optimized. The laser cleaning theoretical fitting model is a mapping relationship between the laser cleaning parameter combination and the response indicator. The laser cleaning parameter combination may include but is not limited to: laser power, pulse frequency, scanning speed, travel speed, pulse width, defocus amount, etc.; the response indicator is a combination of surface roughness, oxygen content, hardness, static contact angle, corrosion rate, and wear rate after cleaning.
[0056] The theoretical model can be specifically expressed as: Y(sa / Oc / HV / CA / U / G)=F(P / f / v / v T / W / D)
[0057] In the formula, sa represents the surface roughness, O c represents oxygen content, HV represents hardness, CA represents static contact angle, U represents corrosion rate, G represents wear rate, P represents laser power, f represents pulse frequency, v represents scanning speed, v T represents the travel speed, W represents the pulse width, and D represents the defocus amount.
[0058] Step 2, using BBD response surface design experiment and using different laser cleaning parameters designed by it to test the aluminum alloy oxide film, obtain the data of multi-objective response indicators after cleaning, and construct the RSM model based on this data;
[0059] The BBD test design factor level and the I group test scheme in the response surface test were selected, and multiple response indicators were tested using analytical testing instruments to obtain the test values of each response indicator. The cleaning parameters were set as pulse frequency, laser power, and scanning speed for the laser cleaning problem of the present invention. At the same time, a quadratic polynomial function was used to mathematically fit the surface roughness, oxygen content, and static contact angle. The model can be defined as follows:
[0060]
[0061] Where n is the number of cleaning parameters; x i represents the i-th cleaning parameter; β0, β i , β ii , β ij They represent the constant term coefficient, the linear term coefficient, the quadratic term coefficient, and the interaction term coefficient respectively.
[0062] Applying the above BBD test design data combined with the above definition model, the RSM model is constructed to obtain the relationship between each cleaning parameter and each response index:
[0063] Sa=2.8804-9.9178×10 -1 *f+7.3633×10 -3 *P-1.4828×10 -4 *v+1.4933×10 -3 *P*f+6.1333×10 -5 *f*v-1.6978×10 -6 *P*v+7.6222×10 -2 *f 2 +1.68×10 -5 *P 2 -3.8889×10-3 *v 2
[0064] O c =5.9241+18.2159*f-0.2919*P+4.2816×10 -3 *v+0.0298*P*f-1.7333×10 -4 *f*v+4.2667×10 -6 *P*v-2.73185*f 2 +2.3882×10 -4 *P 2 -1.1185×10 -6 *v 2
[0065] CA=380.2944-113.950*f-1.3832*P+0.0455*v-0.1084*P*f-7.42220×10 -3 -f*v-8.44440×10 -5 *P*v+17.8222*f 2 +4.96×10 -3 *P 2 +1.5111×10 -6 *v 2
[0066] Where Sa represents the surface roughness, O c represents the oxygen content, CA represents the static contact angle, f represents the pulse frequency, P represents the laser power, and v represents the scanning speed.
[0067] Step 3, using the RANDI function in the matlab software to randomly generate N groups of laser cleaning parameters within the upper and lower limits of the laser cleaning parameters of the response surface test design, and using the RSM model in step 2 to predict the response index values corresponding to the N groups of cleaning parameters to obtain the data of the response index;
[0068] Input matrix for the highest and lowest levels of the laser cleaning parameters:
[0069]
[0070] The RANDI function in the matlab software is used to generate a 3×N matrix of random N groups of cleaning parameter combinations within the highest and lowest level intervals of the laser cleaning parameters in the BBD test in (2):
[0071]
[0072] Where P min, P max ;f min , f max ;v min , v max They represent the lowest and highest levels of laser power, pulse frequency, and scanning speed within the BBD test range, respectively.
[0073] Substitute the above expression Y into the RSM model in step 2 to obtain the N groups of response indicators corresponding to the N groups of cleaning parameter combinations:
[0074]
[0075] Step 4: Use the data of laser cleaning parameters and response indexes obtained in step 2 and step 3 as data sets for training the BP neural network, respectively construct multiple single-output layer BP neural networks of the same number of response indexes, and apply genetic algorithm to optimize them, while optimizing the number of hidden layer nodes of the network, completing the construction of a multi-objective parallel GA-BP neural network based on the optimal hidden layer so that it can predict the nonlinear fitting relationship between the laser cleaning parameters and the response indexes, and testing the model prediction accuracy to obtain a qualified GA-BP neural network;
[0076] The data of N groups of cleaning parameters and response indicators in step 3 and the data of I group of cleaning parameters and response indicators obtained by the response surface test in step 2 are combined and used together as the data set for training the BP neural network.
[0077] The input dataset is:
[0078]
[0079] The output dataset is:
[0080] At the same time, the data set is preprocessed by standard normalization to standardize the data. The formula is:
[0081]
[0082] Where, X i and X represent the data before and after standard normalization preprocessing, respectively. max and X min Represent the maximum and minimum values in the data set, respectively. Represents the mean of the data set X.
[0083] The uniformly normalized input data obtained is:
[0084] Output data:
[0085] The BP neural network is further improved, and three single-output neural networks based on surface roughness, oxygen content, and static contact angle are established respectively, so that they can solve the prediction function in parallel and realize the parallel operation of three neural networks. Genetic algorithm is used for encoding, in which the calculation of individual encoding length is: the number of input layer nodes is m, the number of hidden layer nodes is X, and the number of output layer nodes is 1. There are m*X+X weights and X+1 thresholds. The encoding length of the genetic algorithm is m*X+X+X+1.
[0086] The genetic algorithm is used to optimize the initial weights and thresholds of each neural network by iteratively optimizing the fitness (searching for the optimal weights and thresholds) to achieve the purpose of accurate neural network prediction. The present invention provides a method based on the accuracy of each iteration of the neural network using fitness as an evaluation index. The formula is:
[0087]
[0088] In the formula, Fit is the fitness value. The smaller the value, the smaller the relative error between the actual value and the predicted value of the current iteration, and the better the fitness; x i , Respectively represent the actual and predicted values of surface roughness; y i , Represents the actual and predicted value of oxygen content; z i , represents the actual and predicted values of the contact angle; a is the number of experimental data sets; T is the number of output parameters.
[0089] Further, the training is looped through the for loop statement (m represents the number of input layer nodes, and the step length is 1) The number of hidden layer nodes U in the range, at this time the encoding length of the genetic algorithm is calculated as: (The step size is m+2).
[0090] The fitness values of the GA-BP neural network under different numbers of hidden layer nodes are compared, and the number of hidden layer nodes, weights and thresholds corresponding to the minimum fitness value after iteration of three parallel single-output layer neural networks are selected and substituted into the neural network to realize the construction of a multi-objective parallel GA-BP neural network based on the optimal hidden layer for subsequent prediction work. The weights and thresholds after iteration are as follows:
[0091]
[0092] In the formula, W1 represents the weight from the input layer to the hidden layer of the neural network; B1 represents the threshold of the neurons in the hidden layer of the neural network; W2 represents the weight from the hidden layer to the output layer of the neural network; B2 represents the threshold of the neurons in the output layer of the neural network; w m×URepresents the weight from the input layer to the hidden layer of the m×U neural network; β U Represents the threshold value w of the Uth neural network hidden layer neuron U represents the weight from the hidden layer to the output layer of the Uth neural network; b e represents the threshold of the neuron in the output layer of the e-th neural network; m represents the number of nodes in the input layer; U represents the number of nodes in the hidden layer; and e represents the number of nodes in the output layer.
[0093] Furthermore, the accuracy of the above neural network was verified by applying the determination coefficient R 2 Judgement, R 2 The closer it is to 1, the higher the prediction accuracy is. Its expression is as follows:
[0094]
[0095] In the formula, y i represents the actual value, represents the model prediction value, Represents the mean of the actual value. Analyze and judge whether the model can well reflect the fitting relationship between the cleaning parameters and the output response.
[0096] Furthermore, the generalization ability of the above neural network is verified. F comparison cleaning parameter groups are randomly generated within the BBD test parameter range in (2) to conduct cleaning tests, and the results are tested. At the same time, each comparison parameter group is fitted with the above multi-objective parallel GA-BP neural network model based on the optimal hidden layer to verify the generalization effect of the model. Step 5: Based on the actual needs of laser cleaning, the GA-BP neural network is used as a prediction model to construct the objective function and constraint conditions for laser cleaning parameter prediction optimization;
[0097] The values of laser cleaning parameters are usually limited by the actual conditions of the selected cleaning equipment, such as laser power, pulse frequency, scanning speed, etc. They cannot be less than the equipment limit value, nor can they be higher than the equipment limit value. Furthermore, considering the equipment limit value constraints, the multi-objective optimization model based on laser cleaning can be expressed as:
[0098] F(f,p,v)=(max{CA},min{sa},{Oc})
[0099]
[0100] In the formula, K max is the maximum permissible energy density of the equipment, f min is the minimum pulse frequency of the laser head, f max P is the maximum pulse frequency allowed by the laser head; min Enter the minimum power P for the laser device maxInput the maximum power for the laser device; v min is the minimum scanning speed of the laser device, v max is the peak scanning speed of the laser device.
[0101] Step 6: Use the improved sparrow multi-objective algorithm to iteratively solve the objective function and constraints established in step 5, and finally obtain a set of excellent solutions for laser cleaning parameters. Then, the optimal solution in the excellent solution set for laser cleaning parameters is selected through entropy weight-Topsis as the prediction parameter for laser cleaning.
[0102] The steps of the improved sparrow multi-objective algorithm are:
[0103] a. Initialize the population through the good point set to generate the initialized population within the range of the following formula. In view of the fact that the multi-objective search algorithm can only seek the minimum value, the parameter of the required maximum target is set to the opposite of itself, and define related parameters such as the number of iterations M, the number of individuals in the population L, the number of decision variables J, the number of discoverers Pd, etc.
[0104] [f min , f max ]
[0105] [P min , P max ]
[0106] [v min , v max ]
[0107] The initial population obtained is:
[0108]
[0109] b. Each column in the initial population matrix represents the position of a sparrow in space, i.e., a sparrow individual. The fitness of the sparrow individuals is sorted in a non-dominated manner. The judgment condition f is introduced when the vigilant updates the position. g <f i (f i Dominate f g );and update the judgment condition of the current global optimal position if f g <f i (f i Dominate f g ), which enables the algorithm to handle multi-objective problems.
[0110] In the initial population, a part (10% - 20%) of the sparrows are regarded as sentinels, a part (10% - 20%) of the sparrows are discoverers, and the rest of the sparrows are followers; Sentinels: A certain proportion (10% - 20%) of the sparrow individuals in the initial population are selected for reconnaissance and early warning. When encountering danger (i.e., the judgment condition when the sentinels update their positions), they will quickly move to the safe area (the current global optimal position) to obtain a better position, and the sparrows located in the middle of the population will move randomly to get closer to other sparrows. The sparrows performing this task are called sentinels. Discoverers: Responsible for finding food (optimal position) for the entire population. Followers: Follow the discoverers to obtain food (optimal position).
[0111] c. After non - dominated sorting, for individuals with the same non - dominated rank, introduce a screening and comparison crowding degree operator to sort the crowding degree of the population individuals with the same non - dominated rank, and calculate the crowding degree with the following formula:
[0112]
[0113] In the formula, Li represents the crowding degree of point i, represents the j - th objective solution of the i + l - th point, represents the j - th objective solution of the i - 1 - th point.
[0114] Further execute the screening and comparison crowding degree operator, calculate the crowding degree values of the current population individuals with the same non - dominated rank according to the above formula and calculate the mean of their mean and median, screen out the solutions lower than this value and delete them, retain the remaining population individuals as the initial population to perform the next operation, and obtain the initial sorted solution set, find the optimal individual X best and the worst individual Xworst. d. A certain proportion of the sparrow individuals in the initial population are selected as discoverers. The discoverers are responsible for finding food (optimal position) for the entire population. An adaptive safety threshold and a normal distribution weight factor are introduced for the discoverers and the position update of the discoverers is performed. Its iteration formula is as follows:
[0115]
[0116] In the formula, t represents the current iteration number; iter max represents the maximum number of iterations; X i,j represents the position information of the i - th sparrow in the j - th dimension; random number α; early warning value R2 and adaptive safety threshold AST; random number Q follows a normal distribution; L is a 1×t matrix with all elements being 1. When R2 < AST, it means that the population is not threatened by predators, and the discoverers can search a wide area; when R2 ≥ AST, the discoverers perceive danger and give an alarm, and all sparrows quickly flee from the current position.
[0117] Among them, AST represents the adaptive safety threshold, and its expression is as follows:
[0118]
[0119] NDW represents the weight of the normal distribution operator optimization, and its expression is as follows:
[0120]
[0121] AST is improved based on the characteristics of the exponential curve. In the early stage of the iteration, AST rises slowly from 0.5, and R2 has a high probability of being greater than AST, so that the discoverer has sufficient time to search in other areas, improving the algorithm's search ability; in the later stage of the iteration, the difference between the screened population is small, and AST rises quickly to 1, accelerating the convergence of the algorithm. In the early stage of the NDW iteration, the normal distribution operator makes the weight NDW slowly decrease from 1, and maintains a large weight as much as possible, so that the discoverer has sufficient time to explore the area near the initial population in the early stage of the iteration to prevent the loss of excellent solutions. In the later stage of the iteration, NDW decreases slowly and the value is small. The positions before and after the update are far apart, which makes it easy to maintain the diversity of the sparrow population, enhance the local development ability of the algorithm, and accelerate the convergence of the algorithm to a certain extent.
[0122] e. In the initial population, all sparrow individuals except the discoverer and the sentinel are called followers. Followers follow the discoverer to obtain food (optimal position) and update the follower position according to the classic sparrow algorithm;
[0123] f. Randomly select some sparrows as sentinels and introduce improved step length control parameters to implement sentinel update method. The iterative formula is as follows:
[0124]
[0125] Where X best represents the current global best position (i.e. the best individual); X worst represents the current global worst position (i.e. the worst individual); the step size control parameter β controls the distribution of the updated sparrows around the best position; the random number K is between -1 and 1, controlling the moving direction of the individual sparrows; f i is the fitness value of sparrow individual i; f g and f w are the global maximum and minimum fitness values after t iterations respectively; the minimum value ε prevents the denominator from being 0. i >f g It means that the sparrow individual is at the edge of the population, senses the danger, and moves to the best position individual; f i =f gIt means that the sparrow is in the best position. When it senses danger, it will fly to the edge of the population and escape from its current position. A is a random number that follows the normal distribution Α~N(0, S), is the improved step length control parameter, and the standard deviation S formula related to the fitness of the sparrow individual is as follows:
[0126]
[0127] At this time, when f i Dominate f g When the sparrow's current individual fitness is f i With f g The larger the gap, the smaller the standard deviation S of A, and the greater the probability of being close to the current global optimal individual after updating, which helps to improve the overall convergence speed of the algorithm.
[0128] g. According to the formula if (f i Dominate f g ) Determine whether the updated position is better than the current position. If so, update the position, otherwise keep the current position.
[0129] h. Determine whether the iteration end condition is reached. If so, the algorithm ends and outputs the optimal solution set. Otherwise, return to step [2]. The excellent solution set output after the iteration is:
[0130]
[0131] The corresponding response indicator set is:
[0132]
[0133] In the formula, f good1 Represents the first excellent pulse frequency of the output, f goodL Represents the Lth best pulse frequency of the output; P good1 Represents the first excellent laser power output, P goodL represents the Lth best laser power output; v good1 Represents the first excellent scanning speed of the output, v goodL Represents the Lth best scan speed of the output.
[0134] i. Apply entropy weight-Topsis to select the best solution from the set of optimal solutions.
[0135] The optimal cleaning parameter combination obtained by the method of (6) is used in combination with the test method of (2) to perform test analysis and compare the test analysis results of random samples of the BBD experiment in (2) to verify the effectiveness of the present invention.
[0136] Embodiment 1:
[0137] Step 1, Example 7075 aluminum alloy oxide film is selected as the research object. First, the surface of 7075 aluminum alloy is anodized. The sample is hung on the positive electrode of the anodizing reaction tank and the surface of the workpiece and the two side walls of the reaction tank are kept parallel. The two side walls of the reaction tank are the negative electrode. The constant reaction voltage is 16V. The concentration of sulfuric acid solution in the reaction tank is 18%. After the anodizing reaction for 30 minutes, the workpiece is taken out and dried. At the same time, it is determined to use IPG YLPN-100 pulse fiber laser to perform laser cleaning test on the surface oxide film of 7075 aluminum alloy. The schematic diagram of the test device is shown in the figure. Figure 2 As shown, it is composed of a laser, a control console, a galvanometer, etc., and a theoretical fitting model of laser cleaning is constructed.
[0138] The laser cleaning theoretical fitting model is the mapping relationship between the laser cleaning parameter combination and the response index. When laser cleaning the aluminum alloy anodized film, laser power, pulse frequency, and scanning speed are important cleaning parameters. Their reasonable matching will seriously affect the cleaning rate, corrosion resistance, and wear resistance of the material surface after the aluminum alloy surface oxide film is cleaned. Therefore, this embodiment selects the above three parameters for cleaning test research.
[0139] At the same time, oxygen content, surface roughness and static contact angle were selected as response indicators of the cleaning test.
[0140] Therefore, the theoretical model of the embodiment can be expressed as:
[0141] Y(sa,O c , CA) = F(P, f, v)
[0142] In the formula, sa represents the surface roughness, O c represents the oxygen content, CA represents the static contact angle, P represents the laser power, f represents the pulse frequency, and v represents the scanning speed.
[0143] Step 2: Apply the BBD experiment in the response surface design to design 15 groups of three factors, three levels, and three centers.
[0144] Repeated point laser cleaning of 7075 aluminum alloy oxide film test. The factor level table is shown in Table 1.
[0145] Table 1BBD test factor level table
[0146]
[0147] The oxygen content, surface roughness and static contact angle were measured by energy dispersive spectrometer, ultra-depth microscope and static contact angle measuring instrument. The experimental response results are shown in Table 2.
[0148] Table 2 BBD test results
[0149]
[0150]
[0151] The RSM model was constructed by fitting the data in Table 2 to obtain the relationship between the cleaning parameters and each response index:
[0152] Sa=2.8804-9.9178×10 -1 *f+7.3633×10 -3 *P-1.4828×10 -4 *v+1.4933×10 -3 *P*f+6.1333×10 -5 *f*v-1.6978×10 -6 *P*v+7.6222×10 -2 *f 2 +1.68×10 -5 *P 2 -3.8889×10 -8 *v 2
[0153] 0 c =5.9241+18.2159*f-0.2919*P+4.2816×10 -3 *v+0.0298*P*f-1.7333×10 -4 *f*v+4.2667×10 -6 *P*v-2.73185*f 2 +2.3882×10 -4 *P 2 -1.1185×10 -6 *v 2
[0154] CA=380.2944-113.950*f-1.3832*P+0.0455*v-0.1084-P*f-7.42220×10 -3 *f*v-8.44440×10 -5 *P*v+17.8222*f 2 +4.96×10 -3 *P 2 +1.5111×10 -6 *v 2
[0155] In the formula, sa represents the surface roughness, O crepresents the oxygen content, CA represents the static contact angle, f represents the pulse frequency, P represents the laser power, and v represents the scanning speed.
[0156] Step 3, input the input matrix of the highest and lowest level areas of the laser cleaning parameters:
[0157]
[0158] The RANDI function in the matlab software generates a 3X100 matrix of 100 random cleaning parameter combinations within the highest and lowest level ranges of the laser cleaning parameters of the BBD test:
[0159]
[0160] The above expression Y is introduced into the RSM model in step 2 to obtain 100 groups of response indicators (surface roughness, oxygen content, static contact angle) corresponding to 100 groups of cleaning parameter combinations:
[0161]
[0162] Step 4: Combine the 15 sets of input and output data of the BBD experiment and the Y and X data sets in step 3 to serve as the input and output data sets for training the BP neural network.
[0163] The input dataset is:
[0164]
[0165] The output dataset is:
[0166]
[0167] At the same time, the data set is normalized and preprocessed according to the following formula to unify and standardize the data.
[0168]
[0169] In the formula, X i and X represent the data before and after standard normalization preprocessing, respectively. max and X min Represent the maximum and minimum values in the data set, respectively. Represents the mean of the data set X.
[0170] The uniformly normalized input data obtained is:
[0171]
[0172] Output data:
[0173]
[0174] The BP neural network is further improved. The BP neural network improvement scheme based on this embodiment is as follows: Figure 3 As shown. First, three single-output neural networks based on surface roughness, oxygen content, and static contact angle are established respectively, so that they can solve the prediction function in parallel and realize the parallel operation of three neural networks. Genetic algorithm is used for encoding, in which the individual encoding length is calculated as follows: the number of input layer nodes is m=3, and the number of hidden layer nodes is X. There are m*X+X=4X weights and X+1 thresholds. The encoding length of the genetic algorithm is m*X+X+X+1=5X+1.
[0175] The genetic algorithm optimizes the initial weights and thresholds of each neural network by iteratively optimizing the fitness (searching for the optimal weights and thresholds) to achieve the purpose of accurate neural network prediction. The fitness value of this iteration is calculated using the following formula as an evaluation index to calculate the accuracy of the neural network for each iteration.
[0176]
[0177] In the formula, Fit is the fitness value. The smaller the value, the smaller the relative error between the actual value and the predicted value of the current iteration, and the better the fitness; x i , Respectively represent the actual and predicted values of surface roughness; y i , Represents the actual and predicted value of oxygen content; z i , represents the actual and predicted values of the contact angle; a is the number of experimental data sets; T is the number of output parameters.
[0178] Furthermore, according to the empirical value range formula of the number of hidden layer nodes (m represents the number of input layer nodes, n represents the number of output layer nodes, and the step length is 1) The three groups of ranges are trained by the built-in for loop statement in Matlab. The GA-BP neural network with a hidden layer node number of 3 to 12 is used. At this time, the encoding length of the genetic algorithm is determined by (step length is m+2) and calculated: 16~61 (step length is 5). Finally, the fitness values of three multi-objective single-output layer GA-BP networks of surface roughness, static contact angle, and oxygen content under different hidden layer node numbers are compared. The results are as follows Figure 4 Furthermore, the number of hidden layer nodes when the neural network based on surface roughness, oxygen content, and static contact angle prediction reaches the minimum fitness value is selected as 10, 8, and 7.
[0179] The weights and thresholds after iterative convergence are as follows, and their coding lengths are 48, 40, and 36 respectively. The weights and thresholds after iterative convergence are substituted into the neural network to realize the construction of a multi-objective parallel GA-BP neural network based on the optimal hidden layer for subsequent prediction work.
[0180]
[0181]
[0182]
[0183] Furthermore, the accuracy of the above neural network was verified by applying the determination coefficient R 2 Judgement, R 2 The closer it is to 1, the higher the prediction accuracy is. Its expression is as follows:
[0184]
[0185] In the formula, y i represents the actual value, represents the model prediction value, Represents the mean of the actual values. The relationship between the prediction and the actual value of the BBD test results of this embodiment by the multi-objective parallel GA-BP neural network model based on the optimal hidden layer is as follows: Figure 5 As shown, it can be seen that in the prediction of surface roughness, oxygen content, and static contact angle, the predicted values of the model are well distributed around the diagonal line, and can well reflect the fitting relationship between the cleaning parameters and the output response.
[0186] Further, the generalization ability of the above neural network is verified. F=4 comparison cleaning parameter groups are randomly generated within the BBD test parameter range in step 2 to perform a cleaning test, and the results are tested. At the same time, each comparison parameter group is fitted with the above multi-objective parallel GA-BP neural network model based on the optimal hidden layer. The predicted values and actual values are shown in Table 3. It can be seen that the generalization ability of the model is also strong, and it is suitable for the model prediction work of the problem in this embodiment.
[0187] Table 3 Prediction and actual values of each response index of multi-objective parallel GA-BP neural network based on optimal hidden layer
[0188]
[0189] Step 5: Analyze and select the value range of the laser cleaning parameters for this embodiment, as follows:
[0190] The change in the surface roughness of the workpiece after cleaning will significantly affect the subsequent processing of aluminum alloys. A larger surface roughness may seriously affect the wear resistance of the sample surface. Therefore, it is necessary to ensure that the surface roughness of the workpiece after cleaning is as small as possible; the cleaning rate can measure the cleaning ability of the laser, and the oxygen content on the substrate surface is an important basis for measuring the cleaning rate. The lower the oxygen content, the higher the cleaning rate; the stronger the hydrophobicity of the material, the better its resistance to corrosive liquid erosion. The static contact angle can judge the quality of hydrophobicity. The larger the angle, the stronger the hydrophobicity of the material and the stronger its ability to resist corrosive liquid erosion. In summary, the multi-objective optimization problem in this paper can be described as: while pursuing a larger static contact angle, the surface roughness and oxygen content must be as small as possible. At the same time, the energy density will significantly affect the effect of laser cleaning, and its expression is as follows:
[0191]
[0192] Considering the equipment limit constraints, the energy density cannot be greater than 4.44J / cm 2 .
[0193] Furthermore, the equipment limit value range of the laser power, pulse frequency, scanning speed and other parameters of the cleaning equipment in this embodiment is limited according to the maximum and minimum values allowed by the equipment, namely: 2.25kHz≤f≤4.4kHz, 158W≤P≤275W, 1500mm / s≤v≤3000mm / s. Therefore, the multi-objective optimization model of laser cleaning in this embodiment can be expressed as:
[0194] F(f,p,v)=(max{CA},min{sa},{Oc})
[0195]
[0196] In the formula, f min is the minimum pulse frequency of the laser head, f max P is the maximum pulse frequency allowed by the laser head; min Enter the minimum power P for the laser device max Input the maximum power for the laser device; v min is the minimum scanning speed of the laser device, v max is the peak scanning speed of the laser device.
[0197] Step 6: The sparrow search algorithm is a new type of swarm intelligence optimization algorithm, which features high robustness, fast convergence speed, and strong global optimization ability. However, it has problems such as poor quality of the initial population, being prone to falling into local optimal solutions, relatively small population diversity in the later stage of iteration, and low local development ability, which may lead to the solution result not being the optimal cleaning plan. Therefore, this embodiment proposes a multi-objective sparrow algorithm improved based on good point sets and adaptive normal distribution weights, combined with the entropy weight-Topsis method to solve the multi-objective parameter optimization problem of laser cleaning.
[0198] A) Traditional sparrow search algorithm
[0199] 1) Generate a randomly initialized population as shown in Figure 6 and define relevant parameters such as the number of iterations M, the number of decision variables J, and the number of discoverers Pd.
[0200] 2) Compare and sort the fitness of sparrow individuals to obtain an initial sorted set, and find the optimal and worst individuals (the individuals with the largest and smallest fitness sorting results).
[0201] 3) Some sparrows act as discoverers and play a leading role in the entire population to perform the position update of the discoverers. Its iterative formula is as follows:
[0202] In the formula, t represents the current iteration number; iter max represents the maximum number of iterations; X i,j represents the position information of the i-th sparrow in the j-th dimension; the random number α; the warning value R2 and the safety threshold ST; the random number Q follows a normal distribution; L is a 1×t matrix with all elements being 1. When R2 < ST, it means that the population is not threatened by predators, and the discoverers can search a wide area; when R2 ≥ ST, the discoverers perceive danger and give an alarm, and all sparrows quickly flee from the current position.
[0203] 4) Some sparrows act as followers and execute the update method of the vigilant. Its iterative formula is as follows:
[0204]
[0205] where X p is the optimal position currently occupied by the discoverer, and X worst represents the current worst position. A is a 1×d matrix with elements only being 1 or -1. When i > n / 2, it means that the i-th follower with lower fitness has not obtained food and is in a very hungry state, and needs to fly to other places to forage to obtain higher energy; when i < N / 2, the followers will follow the discoverers and compete for food to increase their own energy.
[0206] 5) Randomly select some sparrows as sentinels and execute the sentinel update method. The iteration formula is as follows:
[0207]
[0208] Where X best represents the current global best position; X worst represents the current global worst position; the step size control parameter β controls the distribution of the updated sparrows around the best position; the random number K is between -1 and 1, controlling the moving direction of the individual sparrows; f i is the fitness value of sparrow individual i; f g and f w are the global maximum and minimum fitness values after t iterations respectively; the minimum value ε prevents the denominator from being 0. i >f g It means that the sparrow individual is at the edge of the population, senses the danger, and moves to the best position individual; f i =f g It means that the sparrow is in the best position. When it senses danger, it will fly to the edge of the population and escape from its current position.
[0209] 6) Determine whether the updated position is better than the current position (compare the individual fitness before and after the update). If so, update the position; otherwise, keep the current position until the end of the iteration.
[0210] B) Application schematic diagram Figure 7 The multi-objective sparrow algorithm based on the best point set and the adaptive normal distribution weight is shown to optimize the multi-objective laser cleaning problem of this embodiment.
[0211] a. Initialize the population through the good point set, and set the maximum target cleaning parameter (ie, static contact angle) obtained in this embodiment to its opposite number according to the method that the multi-objective search algorithm can only seek the minimum value, and generate the following cleaning parameter range: Figure 8 The initial population of good point sets shown in FIG. 1 is more evenly distributed than the randomly generated initial population, and the initial population of good point sets is distributed throughout the interval. Related parameters such as the number of iterations M = 200, the number of individuals in the population n = 100, the number of decision variables J = 3, and the number of discoverers Pd = 30 are defined.
[0212] [f min , f max ] = [2.25, 4.4]
[0213] [P min , P max ]=[158.275]
[0214] [v min , v max] = [1500, 3000]
[0215] Get the initial population:
[0216]
[0217] b. Perform non-dominated sorting on the fitness of individual sparrows and introduce the judgment condition f when the guard updates the position g <f i (f i Dominate f g );and update the judgment condition of the current global optimal position if f g <f i (f i Dominate f g ), so that the algorithm has the ability to handle multi-objective problems. c. After the non-dominated sorting, for individuals with the same non-dominated level, the screening comparison crowding operator is introduced to sort the crowding of individuals with the same non-dominated level. The principle is as follows Fig. 9 As shown. At the same time, the congestion degree is calculated using the following formula:
[0218]
[0219] In the formula, Li represents the congestion degree of point i, represents the jth target solution of point i+1, Represents the jth target solution of point i-1. Further perform the screening and comparison crowding operator, calculate the crowding value of the population individuals with the same current non-dominated level according to the above formula and calculate the mean of their average and median, screen out the solutions below this value and delete them, retain the remaining population individuals as the initial population to perform the next operation, and obtain the initial sorting solution set to find the best and worst individuals.
[0220] d. Some sparrows in the population are used as discoverers to introduce adaptive safety thresholds and normal distribution weight factors and perform discoverer position updates. The iterative formula is as follows:
[0221]
[0222] In the formula, t represents the current iteration number; iter max Indicates the maximum number of iterations; X i,j represents the position information of the i-th sparrow in the j-th dimension; random number α; warning value R2 and adaptive safety threshold AST; random number Q follows normal distribution; L is a 1×t matrix with all elements 1. When R2<AST, it means that the population is not threatened by predators and the discoverer can search a wide area; when R2≥AST, the discoverer detects the danger and warns, and all sparrows quickly flee from the current location. Among them, AST represents the adaptive safety threshold, and its expression is as follows:
[0223]
[0224] NDW represents the weight of the normal distribution operator optimization, and its expression is as follows:
[0225]
[0226] The AST and NDW iteration curves are as follows: Fig.10 As shown. AST is improved based on the characteristics of the exponential curve. Before the 80th iteration, AST slowly rises from 0.5, and R2 has a high probability of being greater than AST, so that the discoverer has sufficient time to search in other areas, improving the algorithm's search ability; after the 160th iteration, the difference between the screened population is small, and AST quickly rises to 1, accelerating the convergence of the algorithm. Before the 80th iteration of NDW, the normal distribution operator makes the weight NDW slowly decrease from 1, and maintains a large weight as much as possible, so that the discoverer has sufficient time to explore the area near the initial population at the beginning of the iteration to prevent the loss of excellent solutions. After the 160th iteration, NDW decreases slowly and the value is small. The positions before and after the update are far apart, which makes it easy to maintain the diversity of the sparrow population, enhance the local development ability of the algorithm, and accelerate the convergence of the algorithm to a certain extent.
[0227] e. Update the follower position according to the classic sparrow algorithm.
[0228] f. Randomly select some sparrows as sentinels and introduce improved step length control parameters to implement sentinel update method. The iterative formula is as follows:
[0229]
[0230] Where X best represents the current global best position; X worst represents the current global worst position; the step size control parameter β controls the distribution of the updated sparrows around the best position; the random number K is between -1 and 1, controlling the moving direction of the individual sparrows; f i is the fitness value of sparrow individual i; f g and f w are the global maximum and minimum fitness values after t iterations respectively; the minimum value ε prevents the denominator from being 0. i >f g It means that the sparrow individual is at the edge of the population, senses the danger, and moves to the best position individual; f i =f gIt means that the sparrow is in the best position. When it senses danger, it will fly to the edge of the population and escape from its current position. A is a random number that follows the normal distribution Α~N(0, S), is the improved step length control parameter, and the standard deviation S formula related to the fitness of the sparrow individual is as follows:
[0231]
[0232] At this time, when f i Dominate f g When the sparrow's current individual fitness is f i With f g The larger the gap, the smaller the standard deviation S of A, and the greater the probability of being close to the current global optimal individual after updating, which helps to improve the overall convergence speed of the algorithm.
[0233] g. Determine whether the updated position is better than the current position. If so, update the position; otherwise, keep the current position.
[0234] h. Determine whether the iteration end condition of 200 generations is reached. If so, the algorithm ends and outputs the optimal solution set. Otherwise, return to step b. After 200 generations of iteration, the output of the cleaning parameter combination excellent solution set is as follows Fig.11 As shown,
[0235]
[0236] The corresponding response indicator set is:
[0237]
[0238] i. Application schematic diagram Fig.12 The entropy weight-Topsis shown selects the optimal solution from the excellent solution set. Some analysis results are shown in Table 4.
[0239] Table 4 Relative closeness of different cleaning parameter combinations
[0240]
[0241] As can be seen from the table above, the highest relative closeness calculated based on the entropy weight-Topsis method is 0.9013, and the optimal cleaning parameters are: laser power 2.6kHz, laser power 245W, scanning speed 2900mm / s. Selecting this cleaning parameter for laser cleaning can achieve a relatively satisfactory optimization effect.
[0242] In order to explore the quality of the optimization results, the optimized 7075 aluminum alloy sample tested based on the optimal cleaning parameter combination obtained in step 6 combined with the test method in step 2 and the random 7075 aluminum alloy sample (NO.3) of the BBD experiment in step 2 were analyzed for morphology, element content, surface roughness, and hydrophobicity. Further, the morphology, element content, surface roughness, and hydrophobicity of the random 7075 aluminum alloy sample (NO.3) based on the BBD experiment and the 7075 aluminum alloy sample optimized based on the method mentioned in the present invention are analyzed. Fig.13 The results are shown in Table 5.
[0243] Table 5 Optimization results of the present invention
[0244]
[0245] Compared with the random 7075 aluminum alloy sample (NO.3) of the BBD experiment, the optimized sample micromorphology has no obvious traces of residual oxide film, the oxygen content is reduced by 96.33%, the surface roughness is reduced by 48.29%, and the static contact angle is increased by 54.57%.
Claims
1. A method for optimizing laser cleaning parameters based on multiple objectives, characterized in that: The steps are: Step 1, cleaning the oxide film on the surface of aluminum alloy and constructing a theoretical fitting model for laser cleaning; Step 2, using BBD response surface design experiment and using different laser cleaning parameters designed by it to test the aluminum alloy oxide film, obtain the data of multi-objective response indicators after cleaning, and construct the RSM model through the data of laser cleaning parameters and response indicators; Step 3, using the RANDI function in the matlab software to randomly generate N groups of laser cleaning parameters within the upper and lower limits of the laser cleaning parameters of the response surface test design, and using the RSM model in step 2 to predict the response index values corresponding to the N groups of cleaning parameters to obtain the data of the response index; Step 4: Use the data of laser cleaning parameters and response indexes obtained in step 2 and step 3 as data sets for training the BP neural network, respectively construct multiple single-output layer BP neural networks of the same number of response indexes, and apply genetic algorithm to optimize them. At the same time, optimize the number of hidden layer nodes of the network, complete the construction of a multi-objective parallel GA-BP neural network based on the optimal hidden layer, so that it can predict the nonlinear fitting relationship between the laser cleaning parameters and the response index, and test the model prediction accuracy to obtain a qualified GA-BP neural network; Step 5: Based on the actual needs of laser cleaning, the GA-BP neural network is used as a prediction model to construct the objective function and constraint conditions for laser cleaning parameter prediction optimization; Step 6: Use the improved sparrow multi-objective algorithm to iteratively solve the objective function and constraint conditions established in step 5, and finally obtain a set of excellent solutions of laser cleaning parameters, and select the optimal solution in the excellent solution set of laser cleaning parameters through entropy weight-Topsis as the prediction parameter of laser cleaning; The steps of the improved sparrow multi-objective algorithm in step 6 are as follows: a. Initialize the population through the good point set to obtain the initial population; b. Perform non-dominated sorting on the fitness of individual sparrows in the initial population and introduce the judgment condition f when the guard updates the position g <f i ; and update the judgment condition f of the current global optimal position g <f i , so that the algorithm has the ability to handle multi-objective problems; c. After the non-dominated sorting, for individuals with the same non-dominated level, the screening comparison crowding operator is introduced to sort the crowding of the individuals with the same non-dominated level to obtain the initial sorting solution set; d. Some sparrows are introduced as discoverers with adaptive safety threshold and normal distribution weight factor and perform discoverer position update; e. Update the follower position according to the classic sparrow algorithm; f. Randomly select some sparrows as guards and introduce improved step control parameters to implement the guard update method; g.Based on f g <f i Determine whether the updated position is better than the current position. If so, update the position, otherwise keep the current position; h. Determine whether the iteration end condition is reached. If so, end the algorithm and output the optimal solution set. Otherwise, return to step b and output the excellent solution set after the iteration ends. i. Apply entropy weight-Topsis to select the best solution from the set of optimal solutions; fi is the fitness value of sparrow individual i; fg is the global maximum fitness value after t iterations.
2. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The theoretical fitting model of laser cleaning in step 1 is: Y(Sa / Oc / HV / CA / U / G)=F(P / f / v / v T / W / D) Where Sa represents the surface roughness, O c represents oxygen content, HV represents hardness, CA represents static contact angle, U represents corrosion rate, G represents wear rate, P represents laser power, f represents pulse frequency, v represents scanning speed, v T represents the travel speed, W represents the pulse width, and D represents the defocus amount.
3. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The RSM model in step 2 is: Sa=2.8804-9.9178×10 -1 *f+7.3633×10 -3 *P-1.4828×10 -4 *v+1.4933×10 -3 *P*f+6.1333×10 -5 *f*v-1.6978×10 -6 *P*v+7.6222×10 -2 *f 2 +1.68×10 -5 *P 2 -3.8889×10 -8 *in 2 O c =5.9241+18.2159*f-0.2919*P+4.2816×10 -3 *v+0.0298*P*f-1.7333×10 -4 *f*v+4.2667×10 -6 *P*v-2.73185*f 2 +2.3882×10 -4 *P 2 -1.1185×10 -6 *v 2 CA=380.2944-113.950*f-1.3832*P+0.0455*v-0.1084*P*f-7.42220×10 -3 *f*v-8.44440×10 -5 *P*v+17.8222*f 2 +4.96×10 -3 *P 2 +1.5111×10 -6 *v 2 Where Sa represents the surface roughness, O c represents the oxygen content, CA represents the static contact angle, f represents the pulse frequency, P represents the laser power, and v represents the scanning speed.
4. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The input data of the GA-BP neural network in step 4 is: The output data is: Fitness as: In the formula, Fit is the fitness value. The smaller the value, the smaller the relative error between the actual value and the predicted value of the current iteration, and the better the fitness; x i , Respectively represent the actual and predicted values of surface roughness; y i , Represents the actual and predicted value of oxygen content; z i , represents the actual and predicted values of the contact angle; a is the number of experimental data sets; T is the number of output parameters; The weights and thresholds are: In the formula, W1 represents the weight from the input layer to the hidden layer of the neural network; B1 represents the threshold of the neurons in the hidden layer of the neural network; W2 represents the weight from the hidden layer to the output layer of the neural network; B2 represents the threshold of the neurons in the output layer of the neural network; Represents the weight from the input layer to the hidden layer of the m×Uth neural network; represents the threshold value of the neuron in the hidden layer of the U-th neural network; wU represents the weight from the hidden layer to the output layer of the U-th neural network; b e represents the threshold of the neuron in the output layer of the e-th neural network; m represents the number of nodes in the input layer; U represents the number of nodes in the hidden layer; and e represents the number of nodes in the output layer.
5. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The objective function and constraints of the GA-BP neural network in step 5 are: F(f,p,v)=(max{CA},min{sa},{Oc}) In the formula, K max is the maximum permissible energy density of the equipment, f min is the minimum pulse frequency of the laser head, f max P is the maximum pulse frequency allowed by the laser head; min Enter the minimum power P for the laser device max Input the maximum power for the laser device; v min is the minimum scanning speed of the laser device, v max is the peak scanning speed of the laser device.
6. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The calculation formula for congestion in step c is: In the formula, Li represents the congestion degree of point i, represents the jth target solution of point i+1, Represents the j-th target solution at point i-1.
7. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The iterative formula for position update in step d is: In the formula, t represents the current iteration number; iter max Indicates the maximum number of iterations; X i,j Represents the position information of the i-th sparrow in the j-th dimension; random number α; warning value R2 and adaptive safety threshold AST; The random number Q follows a normal distribution; L is a 1×t matrix with all elements set to 1; Among them, AST stands for adaptive security threshold, and its expression is as follows: NDW represents the weight of the normal distribution operator optimization, and its expression is as follows: In the formula, exp represents the exponential function with base e; t represents the current iteration number; iter max Indicates the maximum number of iterations.
8. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The iterative formula for the update method of the vigilant in step f is: In the formula, X best represents the current global best position; X worst represents the current global worst position; the step size control parameter β controls the distribution of the updated sparrows around the best position; the random number K is between -1 and 1, controlling the moving direction of the individual sparrows; f i is the fitness value of sparrow individual i; f g and f w are the global maximum and minimum fitness values after t iterations; the minimum value ε prevents the denominator from being 0; f i >f g It means that the sparrow individual is at the edge of the population, senses the danger, and moves to the best position individual; f i =f g It indicates that the sparrow individual is in the best position. When it senses danger, it will fly to the edge of the population and escape from its current position. A is a random number that obeys the normal distribution A~N(0,S), which is the improved step length control parameter.
9. The optimization method based on multi-objective laser cleaning parameters according to claim 1, characterized in that: The excellent solution set output in step h is: The corresponding response indicator set is: In the formula, f good1 Represents the first excellent pulse frequency of the output, f goodL Represents the Lth best pulse frequency of the output; P good1 Represents the first excellent laser power output, P goodL represents the Lth best laser power output; v good1 Represents the first excellent scanning speed of the output, v goodL Represents the Lth best scan speed of the output.
