H13 steel surface laser hardening and polishing parameter optimization method based on small sample data

Abstract

The invention discloses a parameter optimization method for H13 steel surface laser hardening and polishing based on small sample data. Establishing small sample data of laser scanning process parameters in the H13 steel surface laser hardening and polishing process, including laser power P, scanning speed v and overlap rate delta, and H13 steel surface response targets, including through quenching depth H, peak-valley difference value delta H and surface roughness Ra, by adopting a Latin hypercube sampling method; based on the small sample data, an OOA-XGBOOST algorithm is adopted to construct a multi-target prediction model between the process parameters and the response targets; a DNSGA-II algorithm is adopted to carry out process parameter optimization solution to obtain a Pareto solution set, and a comprehensive evaluation sorting method is constructed for the obtained Pareto solution set through combination of VIKOR and CRITIC, so that an optimal process parameter combination is obtained; the optimal process parameter combination is verified, after verification, the hardening and polishing effects of the H13 steel mold are remarkable after laser scanning under the optimal parameters, multi-target parameter optimization can be achieved, and reference is provided for surface modification and process parameter optimization of the H13 steel mold.

Description

Technical Field

[0001] The present invention relates to the technical field of laser processing, and particularly relates to a parameter optimization method for laser hardening and polishing of the surface of H13 steel based on small sample data. Background Art

[0002] At present, it is difficult to achieve overall hardening treatment of large molds by using conventional heat treatment processes, and it is difficult to achieve surface hardening of specific areas of large molds and improvement of their surface topography. However, laser processing technology can flexibly process designated areas on the complex cavity surfaces of molds, which not only effectively improves the surface quality and service life of molds, but also reasonably reduces energy consumption, improves the production and processing efficiency of molds, and reduces production costs.

[0003] Currently, for laser surface modification, researchers pay more attention to single laser hardening effect or polishing effect, and less research is done on synchronous laser hardening and laser polishing. Although some studies have found that laser scanning can simultaneously improve surface hardness and roughness, most of them focus on optimizing surface roughness, and there is little further research on the depth and overlapping situation of the hardened layer, and these parameters significantly affect the hardening effect of the surface layer. Therefore, a parameter optimization for laser hardening and polishing of the surface of H13 steel driven by small sample data is proposed. By taking three process parameters, namely laser power, scanning speed and overlapping rate, as variables, the hardening depth of the specimen, the peak-valley difference between adjacent scanning passes and the surface roughness are optimized. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a parameter optimization method for laser hardening and polishing of the surface of H13 steel based on small sample data.

[0005] The technical solution is as follows: A parameter optimization method for laser hardening and polishing of the surface of H13 steel based on small sample data includes the following steps:

[0006] S1: Using the Latin hypercube sampling method to construct small sample data of laser scanning process parameters during laser hardening and polishing of the surface of H13 steel: laser power P, scanning speed ν, overlapping rate δ, and surface response targets of H13 steel: hardening depth H, peak-valley difference ΔH, and surface roughness Ra;

[0007] S2: Based on the small sample data, using the OOA-XGBOOST algorithm to construct a multi-objective prediction model between process parameters and response targets;

[0008] S3: Using the DNSGA-II algorithm to perform process parameter optimization and solution to obtain the Pareto solution set, and constructing a comprehensive evaluation and ranking method for the obtained Pareto solution set through VIKOR combined with CRITIC, so as to obtain the optimal process parameter combination;

[0009] S4: Perform response target measurement under the optimal process parameter combination conditions obtained in step S3, and verify the optimal process parameter combination based on the measurement value.

[0010] Preferably, the measurement method of the response target in step S1 is as follows:

[0011] The hardening depth H and the peak-valley difference ΔH are measured by a Leica DVM6 ultra-depth-of-field microscope for the hardening depth and the peak-valley difference after laser scanning. The measurement positions of the hardening depth H and the peak-valley difference ΔH are cross-sections perpendicular to the hardening path.

[0012] The surface roughness Ra of the material is measured by an SJ5760-PR surface roughness profiler for the surface roughness before and after scanning. The measurement direction of the surface roughness after polishing is perpendicular to the scanning path.

[0013] Preferably, in the OOA-XGBOOST algorithm in step S2, data is first imported. After import, the dataset is divided. After this dataset division step, the OOA algorithm is introduced. The XGBOOST model is trained by the OOA algorithm, the fitness value of the objective function is calculated, the predation position is determined, the best position is updated, and it is judged whether the final condition is reached. If not, return to training the XGBOOST model to continue execution. After reaching, the optimal hyperparameters are output and obtained, the optimal model is further trained, and the model error is calculated and output.

[0014] Preferably, the model error is calculated by the coefficient of determination R 2 , the Mean Absolute Percentage Error (MAPE), and the Root Mean Squard Error (RMSE) to further evaluate the prediction performance of the established OOA-XGBOOST algorithm model. The calculation formulas are as follows:

[0015]

[0016] where n is the total number of data points, y i is the true value, is the predicted value, is the average value of the true values.

[0017] Preferably: in step S3, the DNSGA-II algorithm is an improved DNSGA-II algorithm, which improves the population initialization part and the population iteration part respectively; first, the population is initialized by the good point set strategy, and then on the basis of NSGA-II, the adaptive detection and adjustment of the environment are added. When the environmental change is detected, the mutation or initialization operation is randomly selected to replace some individuals to maintain the diversity of the population, and the fitness value is adjusted during the individual update process, so as to effectively solve the problem of falling into the local optimal solution. The formula used is:

[0018] F i =f i (t)+λ i

[0019] In the formula: F i is the adjusted fitness value, f i (t) is the fitness value of the i-th individual corresponding to the target, and λ i is the correction value increased based on the change.

[0020] Preferably: the Pareto solution set is obtained by the improved DNSGA-II algorithm, and a comprehensive evaluation and ranking method is constructed for the Pareto solution set by combining the VIKOR and CRITIC algorithms. The ranking method includes the following steps in sequence: establishing a decision matrix and normalizing processing, calculating the weights by the CRITIC method, calculating the comprehensive score, and determining the optimal solution.

[0021] Preferably: the establishment of the decision matrix and normalization processing forms the decision matrix T ij from the obtained Pareto solution set, and the positive and normalization processing are carried out on the response targets of different attributes in the decision matrix to obtain the standard matrix X ij ;

[0022] The calculation of the weights by the CRITIC method considers the relative strength of the samples and the potential conflicts between the response targets. The relative strength is characterized by the standard deviation σ, and the potential conflict between the targets is characterized by the C value. The final weight value W is calculated by comprehensively calculating σ and C. The calculation formula is as follows:

[0023]

[0024] In the formula: σ j is the error value included in the j-th response target, is the mean value of the j-th column; n is the number of solutions in the solution set; m is the number of indicators; r jk is the Pearson correlation coefficient between the j-th column and the k-th column of the response targets; C j is the information entropy included in the j-th response target, and W j is the finally calculated weight;

[0025] Calculate the comprehensive score and determine the optimal solution, using the group benefit S i , the individual regret R i , the decision maker's inclination direction v for S i and R i to calculate the comprehensive score value Q i , the smaller the Q i value, the better the solution. The obtained score value Q i is judged to obtain the optimal solution. The comprehensive score calculation formula and the score value judgment conditional formula are as follows respectively:

[0026]

[0027] In the formula: x j + and x j - represent the positive ideal solution and the negative ideal solution of each response objective in X ij respectively, which are the maximum value and the minimum value of the j-th column response objective; W j is the weight finally calculated for the j-th column; S i is the group benefit, S + and S - represent the maximum value and the minimum value of S i respectively; R i is the individual regret value, R + and R - represent the maximum value and the minimum value of R i respectively; v is the decision coefficient, and Q i represents the comprehensive evaluation index of the i-th group of solution sets.

[0028] Preferably: when the obtained comprehensive evaluation index Q i is in ascending order, only the top 2 solutions before the final sorting need to be judged to select the optimal solution;

[0029] If the C.1 in the score value judgment conditional formula is not satisfied, recursively judge backward until the i-th solution satisfies, then the first to i-th solutions are all optimal solutions;

[0030] If only the condition C.1 is satisfied, the first 2 solutions are both optimal solutions; if all the judgment conditions are satisfied, the solution ranked 1st is the optimal solution.

[0031] Preferably: conduct preliminary verification and in-depth verification on the obtained best parameter combination;

[0032] Among them, the preliminary verification is to measure the response targets of hardening depth H, peak-valley difference ΔH, and surface roughness Ra of the material under the conditions of the optimal combined process parameters. The surface hardening and polishing effect and the prediction ability of the multi-objective prediction model are evaluated according to the measured values of the response targets and the predicted values of the OOA-XGBOOST algorithm model;

[0033] The in-depth verification is to further observe and analyze the three-dimensional morphology and the hardness of the hardened layer before and after laser scanning of the H13 steel surface, and further evaluate the surface hardening and polishing effect.

[0034] Preferably: in the in-depth verification, the three-dimensional morphology is observed by a laser scanning confocal microscope KC-X1000, and the hardness of the hardened layer is measured by an HV-1000B Vickers microhardness tester.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: the optimization method provided by the present invention constructs the connection between process parameters and corresponding targets, realizes multi-objective parameter optimization through laser power, scanning speed, and overlapping rate, can reduce the workload, reduce the consumption of materials, etc., and can provide a reference for the surface modification of H13 steel molds and the optimization of their process parameters;

[0036] Among them, the established OOA-XGBOOST model has better prediction performance and generalization ability; the improved DNSGA-II algorithm shows better diversity and search ability in the process of optimizing the prediction model driven by small samples; the best parameters for laser scanning of the H13 steel surface are determined by combining the VIKOR and CRITC algorithms. Under the conditions of the best parameter combination, the relative error between the predicted value and the actual value is within 6%. After laser scanning, the surface roughness Ra of the H13 steel surface is reduced to 2.4335 μm, and the reduction rate reaches 53.8%. The maximum average hardness reaches 510.67 ± 19.11 HV 0.5 which is about 3.3 times that of the substrate, and the hardening and polishing effects are significant. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a simple flowchart of parameter optimization;

[0038] Figure 2 is a diagram of the distribution of process parameters;

[0039] Figure 3(a) is a fitting effect diagram of the prediction model for H;

[0040] Figure 3(b) is a fitting effect diagram of the prediction model for ΔH;

[0041] Figure 3(c) is a fitting effect diagram of the prediction model for Ra;

[0042] Figure 4(a) is a Pareto solution set diagram obtained by the improved DNSGA-II algorithm;

[0043] Figure 4(b) is the Pareto solution set diagram obtained by the MOGWO algorithm;

[0044] Figure 4(c) is the Pareto solution set diagram obtained by the MOPSO algorithm;

[0045] Figure 4(d) is the Pareto solution set diagram obtained by the DNSGA-II algorithm;

[0046] Figure 5 is the cross-sectional view after laser scanning;

[0047] Figure 6 is the surface profile diagram before and after polishing;

[0048] Figure 7(a) is the surface topography diagram before laser scanning;

[0049] Figure 7(b) is the surface topography diagram after laser scanning;

[0050] Figure 8 is the schematic diagram of the hardness measurement method;

[0051] Figure 9 is the diagram of the hardness change law at different depths;

[0052] Figure 10 is the cross-sectional microstructure diagram of the hardened layer;

[0053] Figure 11 is the surface XRD diagram before and after laser scanning. Specific implementation manner

[0054] The present invention will be further described below in conjunction with the embodiments and the drawings.

[0055] As Figure 1 shown, the parameter optimization method for laser hardening and polishing of the H13 steel surface based on small sample data includes the following steps:

[0056] S1: Use the Latin hypercube sampling method to construct small sample data of the laser scanning process parameters in the laser hardening and polishing process of the H13 steel surface: laser power P, scanning speed ν, overlap rate δ, and the surface response targets of the H13 steel: hardening depth H, peak-to-valley difference ΔH, and surface roughness Ra;

[0057] S2: Based on the small sample data, use the OOA-XGBOOST algorithm to construct a multi-objective prediction model between the process parameters and the response targets;

[0058] S3: The DNSGA-II algorithm is used to optimize and solve the process parameters to obtain the Pareto solution set, and a comprehensive evaluation and ranking method is constructed for the obtained Pareto solution set by combining VIKOR with CRITIC, so as to obtain the optimal process parameter combination;

[0059] S4: Under the condition of the optimal process parameter combination obtained in the step S3, the response target is measured, and the optimal process parameter combination is verified according to the measured value.

[0060] Example

[0061] A mold made of H13 steel is selected, and its surface is scanned by a laser scanning system. The laser scanning system mainly consists of the following parts: RFL-C3000 fiber-optic output semiconductor laser, FANUC six-axis robot, TS-6A laser head, TFLW-3000 water-cooling device, etc. The laser beam is a Gaussian beam with a spot diameter of 2 mm. To prevent surface oxidation during the laser processing, coaxial gas blowing is carried out during the test, and the auxiliary gas is argon. The size of the surface of the H13 steel mold is 22 mm×170 mm×10 mm, and its chemical composition is shown in Table 1 below:

[0062] Table 1 Main chemical composition table of H13 steel (mass ratio, %)

[0063] C Si Mn Cr Mo V P S Fe 0.32-0.45 0.80-1.20 0.20-0.50 4.75-5.50 1.10-1.75 0.80-1.20 ≤0.03 ≤0.03 Bal

[0064] To obtain a good initial surface, the surface is rough-ground, milled, and rust-removed before the test, and then the surface is cleaned and dried with ethanol and ultrasonic waves. The size of the laser scanning area is 15×15 mm, and the scanning path is in the shape of a "bow".

[0065] During the laser scanning process, the power P and scanning speed v of the laser significantly affect the thermal change process of the material surface, while the overlap ratio δ dominates the overlap effect of adjacent passes. To better evaluate the hardening and polishing effects after laser scanning, the hardening depth H and the peak-valley difference ΔH are selected as the indicators to evaluate the quality of the hardened layer, and the surface roughness Ra of the material surface is selected as the indicator to evaluate the surface topography quality. The three process parameters of P, v, and δ are selected as the test factors, and the corresponding H, ΔH, and Ra are measured as the response targets;

[0066] The surface roughness before and after scanning is measured by an SJ5760-PR surface roughness profiler. The measurement is based on ISO 4288:1996 and ISO 4287:2010. The sampling length is 2.5 mm, the evaluation length is 12.5 mm, and the measurement direction of the surface roughness after polishing is perpendicular to the scanning path;

[0067] The surface hardness was measured using an HV-1000B Vickers microhardness tester. When measuring the hardness of the hardened layer, the loading force was 500 g, the holding time was 10 s, and the average value was taken after randomly measuring 3 points.

[0068] The quenching depth and the peak-valley difference after scanning were measured using a Leica DVM6 ultra-depth-of-field microscope. The measurement positions of the quenching depth H and the peak-valley difference ΔH were cross-sections perpendicular to the hardening path. Considering that the thickness of the remelting zone and the heat-affected zone is small and the differentiation is low, measuring a single area alone is likely to cause large measurement errors. Therefore, the thickness of the "remelting zone + heat-affected zone" was selected as the quenching depth H. To achieve a good hardening effect, adjacent scanning passes should be fully overlapped. The adjacent peak-valley difference ΔH of the hardened area was used as the evaluation index, and the average value was taken after measuring the values of 3 hardened areas as the final measurement result. Before measuring H and ΔH, the cross-section was ground, polished, then etched with a 4% (volume fraction) nitric acid alcohol solution and cleaned with anhydrous ethanol.

[0069] The Latin Hypercube Sampling method (LHS) is a method for sampling from a multi-variable parameter distribution and belongs to the stratified sampling method. It has achieved remarkable results and has been widely applied in numerical analysis, physical and computational experimental design, and neural network learning optimization. The value ranges of each process parameter were determined as follows: P ∈ [520 - 920] W, v ∈ [5 - 15] mm / s, δ ∈ [10 - 70]%. The LHS was used for experimental design, and 25 groups of process parameter combinations were selected. The spatial distribution of the experimental points designed by the LHS is shown in Figure 3, and the experimental scheme and measurement results are shown in Table 2:

[0070] Table 2 Experimental Scheme and Measurement Results

[0071]

[0072] The OOA-XGBOOST algorithm was established to construct a multi-objective prediction model for laser hardening and polishing. 20 samples were randomly selected from the above-mentioned 25 groups of samples as the training set, and the remaining samples formed the test set. The ratio of the training set to the test set was 4:1. The value ranges of the XGBOOST hyperparameters were as follows: n_estimators = 0 - 100, max_features = 1 - 5, max_depth = 1 - 20, learning_rate = 0.1 - 0.5, subsample = 0.6 - 1.0, colsample_bytree = 0.6 - 1.0. In the OOA optimization algorithm, the number of ospreys was set to 20, and the number of iterations was 100. The mean-square error (MSE) of the test set was used as the fitness value. The optimal XGBOOST hyperparameters were obtained through optimization, and the coefficient of determination R2 The mean absolute percentage error (MAPE), root mean squared error (RMSE) are used to evaluate the prediction ability of its prediction model. The calculation formulas are as follows:

[0073]

[0074] where n is the total number of data points, y i is the true value, is the predicted value, is the average value of the true values.

[0075] Meanwhile, in order to verify the superiority of the OOA-XGBOOST model in terms of prediction performance, other prediction models are set for comparison, including the backpropagation neural network prediction model (BPNN), decision tree prediction model (DT), random forest prediction model (RF), and extreme gradient boosting algorithm prediction model (XGBOOST);

[0076] BPNN is modeled through the forward information transmission and backward error transmission of the input layer, hidden layer, and output layer; DT is a simple and applicable classification and regression method that completes the division of data features and regression prediction through the internal nodes, leaf nodes, branches, etc. of the tree structure; the RF model is an ensemble learning method that improves the accuracy and generalization ability of the model by constructing multiple decision trees and combining their prediction results; XGBOOST is an ensemble learning algorithm based on decision trees. It forms a strong learner by combining multiple learners. First, start from a simple model and gradually add new models to reduce the prediction error, then determine the direction and degree of the new model to be added by calculating the gradient of the loss function, and add a regularization term to the loss function to improve the generalization ability of the model.

[0077] On the premise that the training set and test set data remain unchanged, the other 4 models are modeled. The parameters in the models are all default values. The fitting effects of the prediction models on H, ΔH, and Ra are shown in Figure 3(a), Figure 3(b), Figure 3(c), and Table 3:

[0078] Table 3 Comparison table of model accuracies

[0079]

[0080] In summary, the prediction accuracy and fitting degree of the OOA-XGBOOST model are higher than those of BPNN, DT, RF, and XGBOOST. The OOA-XGBOOST model has good prediction and generalization abilities, which can provide reliable prediction model support for the subsequent optimization of the laser hardening and polishing process parameters of H13 steel surface.

[0081] To achieve good surface hardening-polishing effects, desired directions are proposed for each response target. Among them, H should be as large as possible to obtain a hardened layer with sufficient thickness; ΔH should be as small as possible to ensure the thickness of the effective hardened area; Ra should be as small as possible to improve the original rough profile and achieve a smoother surface morphology, and the optimization objective function is set. By optimizing the three response targets, good surface hardening-polishing effects can be achieved. Based on the preliminary experiment and comprehensively considering the convergence effect of the optimization process, constraint conditions are set. The constraint conditions and the expression of the optimized objective function are as follows:

[0082]

[0083] In the improved DNSGA-II algorithm, the initial population size is set to 200, the number of iterations is 200, the crossover ratio is 0.9, the mutation probability is 0.1, the change detection ratio is 0.1, the diversity maintenance ratio is 0.3, and the environmental change detection threshold is 0.01. The process is as follows: First, the population is initialized by the good point set strategy, and then an adaptive detection and adjustment of the environment are added on the basis of DNSGA-II. When the environment change is detected, random mutation or initialization operations are selected to replace some individuals to maintain the diversity of the population, and the fitness value is adjusted during the individual update process, so as to effectively solve the problem of falling into the local optimal solution.

[0084] To compare and analyze the superiority of the improved DNSGA-II algorithm, the multi-objective particle swarm optimization (MOPSO), multi-objective grey wolf optimization (MOGWO) algorithms, and the non-dominated sorting genetic algorithm-II (DNSGA-II) are added for comparison. The optimization performance of the algorithms is compared and analyzed through two indicators, Hypervolume (HV) and Inverted Generational Distance (IGD);

[0085] Among them, HV considers the volume occupying the objective space. It calculates the volume of the region enclosed by the Pareto solution set and a reference point. The larger this value, the better the diversity and more reasonable the distribution of the Pareto solution set. IGD considers the distance between the Pareto solution set and the ideal solution. The smaller this value, the closer the obtained Pareto solution set is to the ideal situation. The calculation formulas for HV consideration and IGD consideration are as follows:

[0086]

[0087] where: η is the Lebesgue measure; S is the number of Pareto solution sets; v i is the hypervolume formed by the reference point and the i-th solution set point; P1 is the obtained solution set; P1 * is a set of reference solution sets; is the reference set P1 * in the points and the Euclidean distance between the point k1 in the solution set P1.

[0088] The settings of the MOGWO algorithm are: population size 200, number of iterations 200, and the default linear convergence factor is used as the default linear convergence factor; the settings of the MOPSO algorithm are: population size 200, number of iterations 200, inertia weight 0.5, learning factor 1.5, and the adaptive grid equal component 2;

[0089] The distribution of the first 70 groups of Pareto solution sets finally obtained is shown in Figure 7. The corresponding HV and IGD of the 4 optimization algorithms are shown in Table 4.

[0090] Table 4 Comparison table of algorithm performance

[0091] Evaluation index MOPSO MOGWO DNSGA-Ⅱ Improved DNSGA-Ⅱ HV <![CDATA[2.53×10 5 > <![CDATA[2.16×10 5 > <![CDATA[2.92×10 5 > <![CDATA[3.03×10 5 > IGD 15.9 32.1 6.89 6.19

[0092] As shown in Table 4 and Figures 4(a), 4(b), 4(c), and 4(d) above, the HV value of the improved DNSGA-II is greater than the other 3 algorithms, indicating that this algorithm is more adept at exploring the diversity of the solution set during the optimization process. At the same time, it can be seen that the IGD of the improved DNSGA-II is less than the other 3 algorithms, which indicates that this algorithm has stronger search ability. It shows that the improved DNSGA-II algorithm demonstrates more excellent comprehensive performance during the optimization process, so the solution set obtained by this algorithm is selected as the final optimized solution set.

[0093] The Pareto solution set obtained by the above improved DNSGA-II is re-ranked through VIKOR combined with CRITIC, and the steps are as follows:

[0094] 1. Establish the decision matrix and perform normalization processing. The Pareto solution set obtained by the multi-objective optimization algorithm forms the decision matrix T ij and the positive and normalization processing of each response target with different attributes in the decision matrix is performed to obtain the standard matrix X ij ;

[0095] 2. Calculate the weights by the CRITIC method. CRITIC is an objective weighting method that comprehensively considers the relative strength of the samples and the potential conflicts between the response targets. The relative strength is characterized by the standard deviation σ, the potential conflict between the targets is characterized by the C value, and the final weight value W is calculated by comprehensively considering σ and C. The relevant calculation formula is:

[0096]

[0097] Where: σ j is the error value included in the j-th response target, is the mean value of the j-th column; n is the number of solutions in the solution set; m is the number of indicators; r jk is the Pearson correlation coefficient between the j-th column and the k-th column response targets; C j is the information entropy included in the j-th response target, W j is the finally calculated weight.

[0098] 3. Calculate the comprehensive score and determine the optimal solution. Combine VIKOR and CRITIC through the sorting calculation formula:

[0099]

[0100] where x j + and x j - respectively represent the positive ideal solution and the negative ideal solution of each response target in X ij , which are the maximum value and the minimum value of the j-th column response target respectively; W j is the finally calculated weight of the j-th column; S i is the group benefit, S + and S - respectively represent the maximum value and the minimum value of S i ; R i is the individual regret value, R + and R - respectively represent the maximum value and the minimum value of R i ; v is the decision coefficient, Q i represents the comprehensive evaluation index of the i-th group of solution sets.

[0101] Sort the solution set. Among them, the group benefit S i reflects the overall closeness of this group of solutions to the ideal solution, reflecting the absolute performance of the solutions; the individual regret R i represents the performance of a solution set relative to the worst solution under each criterion, which is a measure reflecting the relative superiority and inferiority of the solutions. v depends on the decision maker's tendency direction for S i and R i . Here, take v = 0.5 to make a compromise for the solutions. Finally, calculate the comprehensive score Q i , and the smaller the Q i value, the better the solution. The judgment condition formula is:

[0102]

[0103] When the obtained comprehensive evaluation index Q i is in ascending order, only the top 2 solutions before the final ranking need to be judged to select the optimal solution;

[0104] If C.1 in the scoring value judgment condition formula is not satisfied, then recursively judge backward until the i-th solution satisfies, and then the first to i-th solutions are all optimal solutions;

[0105] If only condition C.1 is satisfied, then the top 2 solutions are both optimal solutions; if all judgment conditions are satisfied, then the solution ranked 1st is the optimal solution.

[0106] According to the Pareto solution set obtained by improving DNSGA-II, re-rank it through VIKOR combined with CRITIC. The top 5 results after the final ranking are shown in Table 5. Through the above judgment conditions for the optimal solution judgment, it is found that Q1 only satisfies judgment condition C.1. Therefore, the solution sets corresponding to Q1 and Q2 are both optimal solutions.

[0107] Table 5 The top 5 groups of solutions for comprehensive evaluation ranking

[0108] NO. P / W <![CDATA[v / (mm·s -1 )]]> δ / % Ra / μm H / μm ΔH / μm <![CDATA[S i > <![CDATA[R i > <![CDATA[Q i > 1 793.0291 9.2820 41.8260 2.5715 982.6173 144.3267 0.0570 0.0570 0.0009 2 787.7155 9.4735 41.9124 2.5759 978.6181 145.5351 0.0645 0.0578 0.0070 3 774.4715 9.0337 43.5136 2.5765 969.5222 143.4933 0.0726 0.0564 0.0101 4 774.4715 9.0337 41.0374 2.5760 975.5882 152.6799 0.0725 0.0627 0.0197 5 792.5727 10.3157 41.8353 2.6071 901.2695 131.3647 0.1516 0.0819 0.1003

[0109] Since both Q1 and Q2 are optimal solutions, but the response targets of these two solutions are extremely close (the relative difference of the response targets is less than 1%), so Q1 is taken as the optimal solution. After rounding, a process experiment is carried out for verification, that is, P = 793W, ν = 9.3mm·s -1 , δ = 41.8%. The surface roughness before and after scanning is measured by an SJ5760-PR surface roughness profiler, and the hardened depth and peak-valley difference after scanning are measured by a Leica DVM6 super-depth-of-field microscope. The relative error of H after scanning is 994.9038μm compared with the predicted value is 1.23%, the relative error of ΔH is 138.7315μm compared with the predicted value is 3.88%, and the surface roughness Ra is 2.4335μm with a relative error of 5.4% compared with the predicted value. This shows that the prediction accuracy of the model is good; its cross-section perpendicular to the laser movement direction is as Figure 5 shown. The hardened depth H obtained is 994.9038μm, and the peak-valley difference between adjacent passes is ΔH of 138.7315μm. It can be seen that the laser overlaps from right to left in turn, and the overlapping effect of the hardened layers between adjacent passes is good;

[0110] As Figure 6As shown, the surface roughnesses of the three scanned paths are Ra = 2.3258 μm, Ra = 2.2480 μm, and Ra = 2.7267 μm respectively. The calculated average roughness is Ra = 2.4335 μm. Compared with the surface roughness Ra of 5.2674 μm before scanning, the roughness reduction rate reaches 53.8%. It can be seen that laser scanning effectively reduces the original surface roughness of H13 steel.

[0111] To further verify the surface hardening and polishing effects, a laser scanning confocal microscope KC-X1000 and an HV-1000B Vickers microhardness tester were used to observe and analyze the surface topography before and after laser scanning and measure and analyze the surface hardness.

[0112] As shown in Figures 7(a) and 7(b), after laser scanning, the original topography has been reconstructed, and the surface shows a regular pass distribution. Under the action of the Gaussian laser, the heated area melts and mainly flows under the action of capillary force and thermocapillary force, forming an "M"-type surface structure with a slightly higher middle and slightly lower sides.

[0113] As Figure 8 shown, the hardness test of the hardened layer was carried out in the following two ways;

[0114] 1. At a depth of 100 μm from the surface in three adjacent passes, 3 points were tested in each pass, for a total of 9 points (named NO.1 - NO.9 in sequence);

[0115] 2. In a single pass, 9 points were measured at equal intervals of 100 μm downward, and 3 groups were measured, for a total of 27 points (named D1 - D9 in sequence);

[0116] The hardness test results show that the average hardness at different passes and the same depth (100 μm from the surface) reached 510.67 ± 19.11 HV 0.5 , about 3.3 times the matrix hardness of 156.7 ± 5.9 HV 0.5 . The hardening effect is significant. At the same time, the hardness value fluctuates less, the hardness uniformity is better, and the overlapping effect of adjacent passes is good. This is because the H13 matrix structure is composed of pearlite and ferrite, as Figure 10 shown in a1, and its hardness is relatively low; while under the action of the laser, the surface of H13 steel undergoes a phase transformation, as Figure 10 shown in b1, and the structure transforms into martensite, resulting in surface hardening and an increase in the hardness value.

[0117] As Figure 9For the hardness measurement results at different depths shown, within a depth of 700 μm, the hardness value changes little and the hardness value fluctuates slightly. As the depth increases, at a distance of 900 μm from the surface, the average hardness is 294.03 ± 102.41 HV 0.5 , indicating that the hardening effect decreases at this depth, and at the same time, the hardness value fluctuates greatly. This is because the heat is lower in the deeper area. Under the energy characteristics of the Gaussian laser, the outermost part of a single pass absorbs less energy, resulting in a relatively low heating temperature in this area, incomplete austenitization, and a situation similar to subcritical quenching appears.

[0118] As Figure 10 shown in c and c1, in the adjacent pass junction area at a deeper depth, pearlite + ferrite + martensite tissues are simultaneously distributed, and the hardness measurement results show that there is an obvious area with a lower hardness at the junction of adjacent passes.

[0119] In summary, after laser scanning, the surface hardening effect of H13 steel is significant, the overall hardness distribution of the hardened layer is relatively close, and the overlapping effect of adjacent passes is good.

[0120] The surface of H13 steel before and after laser treatment was scanned using an X-ray diffractometer in the diffraction angle range of 30° to 110°, with a step size of 0.02° and a holding time of 0.5 s at each diffraction angle. The results are as Figure 11 shown. By comparing with the PDF card, it can be found that the H13 steel matrix is mainly composed of α-Fe, (Fe,Cr)7C3, and FeC3. After scanning the surface, it is mainly composed of α-Fe, γ-Fe, (Fe,Cr)7C3, and FeC3. Among them, α-Fe in the matrix mainly exists in the form of ferrite and pearlite tissues, and α-Fe on the scanned surface mainly exists in the form of martensite tissues. From the change of the diffraction peaks, the corresponding explanation can be obtained. Compared with the initial surface, the second diffraction peak slightly increases and the third diffraction peak slightly decreases after laser scanning, indicating that more carbon elements are dissolved into martensite, thus effectively hardening the surface layer. Due to the large cooling rate during laser scanning, the stability of supercooled austenite is enhanced, resulting in retained austenite on the surface after hardening. The EDS line scans were performed on the hardened and non-hardened areas of the cross-section perpendicular to the scanning path using a TESCAN VEGA scanning electron microscope and energy spectrometer to observe the changes in the main composition elements of H13 steel respectively. The measurement results show that the surface element contents of the two areas are similar, and the element changes on the hardened surface are not obvious.

[0121] Under laser scanning with appropriate process parameters, the surface of the H13 steel mold can achieve good hardening effect while effectively improving the surface topography. Due to the complex relationship between process parameters and the surface hardening and polishing objectives of the material, this study proposes an optimization method for surface laser hardening and polishing process parameters based on small-sample data. The established OOA-XGBOOST model shows better prediction performance and generalization ability. The improved DNSGA-II algorithm shows better diversity and search ability in the process of optimizing the small-sample-driven prediction model. A comprehensive evaluation and ranking method is constructed by combining VIKOR with CRITC to rank the Pareto solution set, and the optimal process parameter combination is obtained. Experiments are carried out under this parameter combination, and the relative error between the predicted value and the actual value is within 6%. The prediction accuracy of the model is good. After laser scanning, the surface roughness Ra of the H13 steel surface is reduced to 2.4335 μm, and the reduction rate reaches 53.8%. The maximum average hardness reaches 510.67 ± 19.11 HV 0.5 , about 3.3 times that of the matrix, and the hardening and polishing effects are remarkable. This optimization method can provide a reasonable reference for the surface modification of die steel and the optimization of its process parameters.

[0122] Finally, it should be noted that the above description is only the preferred embodiment of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can make various similar representations without departing from the purpose and claims of the present invention. Such transformations all fall within the protection scope of the present invention.

Claims

1. A method for optimizing the parameters of laser hardening and polishing on the surface of H13 steel based on small sample data, characterized in that: It includes the following steps: S1: Use the Latin hypercube sampling method to construct small sample data of the laser scanning process parameters during the laser hardening and polishing process of the H13 steel surface: laser power P, scanning speed ν, overlapping rate δ, and the surface response targets of the H13 steel surface: hardening depth H, peak-to-valley difference ΔH, and surface roughness Ra; S2: Based on the small sample data, use the OOA-XGBOOST algorithm to construct a multi-objective prediction model between the process parameters and the response targets; S3: Use the DNSGA-II algorithm to optimize and solve the process parameters to obtain the Pareto solution set, and construct a comprehensive evaluation and ranking method for the obtained Pareto solution set through VIKOR combined with CRITIC, so as to obtain the optimal process parameter combination; S4: Measure the response targets under the conditions of the optimal process parameter combination obtained in step S3, and verify the optimal process parameter combination according to the measured values; 2. The parameter optimization method according to claim 1, wherein: The measurement method of the response target in step S1 is as follows: The hardening depth H and the peak-to-valley difference ΔH are measured by a Leica DVM6 super-depth-of-field microscope for the hardening depth and the peak-to-valley difference after laser scanning. The measurement positions of the hardening depth H and the peak-to-valley difference ΔH are cross-sections perpendicular to the hardening path; The surface roughness Ra of the material surface is measured by an SJ5760-PR surface roughness profiler for the surface roughness before and after scanning. The measurement direction of the surface roughness after polishing is perpendicular to the scanning path; 3. The parameter optimization method according to claim 1, characterized in that: In the OOA-XGBOOST algorithm in step S2, first import the data. After the import, divide the data set. After this data set division step, introduce the OOA algorithm. Train the XGBOOST model through the OOA algorithm, calculate the fitness value of the objective function, determine the predation position, update the best position, judge whether the final condition is reached. If not, return to train the XGBOOST model to continue execution. After reaching, output and obtain the optimal hyperparameters, further train the optimal model, calculate the model error and output; 4. The parameter optimization method according to claim 3, characterized in that: The calculation model error is determined by the coefficient of determination R 2 , the mean absolute percentage error (MAPE), and the root mean squared error (RMSE) to further evaluate the prediction performance of the established OOA-XGBOOST algorithm model. The calculation formulas are as follows: where n is the total number of data points, y i is the true value, is the predicted value, is the average value of the true values.

5. The parameter optimization method according to claim 1, wherein: The DNSGA-II algorithm in step S3 is an improved DNSGA-II algorithm, which improves the population initialization part and the population iteration part respectively; first initialize the population through the good point set strategy, and then add adaptive detection and adjustment of the environment on the basis of NSGA-II. When the environmental change is detected, randomly select mutation or initialization operations to replace some individuals to maintain the diversity of the population, and adjust the fitness value during the individual update process, so as to effectively solve the problem of falling into the local optimal solution. The formula used is: F i = f i (t) + λ i where: F i is the adjusted fitness value, f i (t) is the fitness value of the target corresponding to the i-th individual, and λ i is the correction value increased based on the change.

6. The parameter optimization method according to claim 5, characterized in that: Obtain the Pareto solution set through the improved DNSGA-II algorithm, and construct a comprehensive evaluation and ranking method for the Pareto solution set by using the VIKOR combined with the CRITIC algorithm. The ranking method includes: establishing a decision matrix and normalization processing, calculating the weights by the CRITIC method, calculating the comprehensive score, and determining the optimal solution in turn; 7. The parameter optimization method according to claim 6, characterized in that: The establishment of the decision matrix and the normalization process will form a decision matrix T with the obtained Pareto solution set ij , and perform positive and normalization processing on each response target with different attributes in the decision matrix to obtain a standard matrix X ij ; The weight calculation of the CRITIC method takes into account the potential conflict between the relative strength of the samples and the response objectives. The relative strength is characterized by the standard deviation σ, and the potential conflict between the objectives is characterized by the C value. The final weight value W is calculated by comprehensively considering σ and C, and its calculation formula is as follows: Where: σ j is the error value included in the j-th response target, is the mean value of the j-th column; n is the number of solutions in the solution set; m is the number of indicators; r jk is the Pearson correlation coefficient between the j-th column and the response target of the k-th column; C j is the information entropy included in the j-th response target, W j is the finally calculated weight; Calculate the comprehensive score and determine the optimal solution, using the group benefit S i , the individual regret R i , the decision maker's tendency direction v for S i and R i to calculate the comprehensive score value Q i , Q i The smaller the value, the better the solution. The obtained score value Q i is judged to obtain the optimal solution. The comprehensive score calculation formula and the score value judgment conditional formula are as follows: where: x j + and x j - represent the positive ideal solution and the negative ideal solution of each response target in X ij respectively, which are the maximum value and the minimum value of the j-th column response target; W j is the weight finally calculated for the j-th column; S i is the group benefit, S + and S - represent the maximum and minimum values of S i respectively; R i is the individual regret value, R + and R - represent the maximum and minimum values of R i respectively; v is the decision coefficient, Q i represents the comprehensive evaluation index of the i-th group of solution sets.

8. The parameter optimization method according to claim 7, characterized in that: When the obtained comprehensive evaluation index Q i is sorted in ascending order, it is only necessary to judge the top 2 solutions before the final sorting to select the optimal solution; If C.1 in the scoring value judgment condition is not satisfied, recursively judge backward until the i-th solution satisfies, then the first to i-th solutions are all optimal solutions; If only condition C.1 is satisfied, the first two solutions are both optimal solutions; if all judgment conditions are satisfied, the solution ranked first is the optimal solution.

9. The parameter optimization method according to claim 1 or 8, characterized in that: Conduct preliminary verification and in-depth verification on the obtained optimal parameter combination; Among them, the preliminary verification is to measure the response objectives of the hardening depth H, the peak-valley difference ΔH, and the surface roughness Ra of the material under the conditions of the optimal combined process parameters, and evaluate the surface hardening and polishing effect and the prediction ability of the multi-objective prediction model according to the measured values of the response objectives and the predicted values of the OOA-XGBOOST algorithm model; The in-depth verification is to further observe and analyze the three-dimensional morphology and the hardness of the hardened layer before and after laser scanning on the surface of H13 steel, and further evaluate the surface hardening and polishing effect.

10. The parameter optimization method according to claim 9, wherein: In the in-depth verification, the three-dimensional morphology is observed by a laser scanning confocal microscope KC-X1000, and the hardness of the hardened layer is measured by an HV-1000B Vickers microhardness tester.

Images