A method for recalibration of laser selective melting process parameters

Abstract

The application discloses a kind of re-calibration methods of laser selective melting process parameters, it is related to the field of process re-calibration, the method acquires source domain data and target domain data, and pre-training model is constructed based on source domain data;Source domain process-geometric feature parameters and target domain process-geometric feature parameters are normalized, and one-hot coding is executed to category variable;Whether to carry out migration learning re-calibration is judged using the comprehensive judgment of calculation edge distribution difference and conditional distribution difference;When migration learning, TrAdaBoost-R algorithm and MO-GPR algorithm are iteratively trained submodel, and the submodel obtained each time is fused to form single callable prediction function;Pareto front is searched to prediction function, and the optimal process parameter solution is selected from Pareto front.The application can shorten new product introduction cycle, reduce cost, and provide key support for SLM technology industrial application.

Description

Technical Field

[0001] This invention relates to the field of process recalibration, and in particular to a method for recalibrating laser selective melting process parameters. Background Technology

[0002] In the additive manufacturing process of critical components for major equipment such as aero-engine blades, spacecraft modules, and heavy-duty gas turbine impellers, the original process window becomes invalid after powder batch replacement, laser maintenance, or equipment upgrades (referred to as "powder / machine replacement"), requiring recalibration of the laser selective melting process parameters. Currently, companies generally adopt "full factorial / full-level" re-experimentation, but this has the following drawbacks: (1) The experiment is large in scale, consumes a lot of metal powder, and takes a long time to operate; (2) The historical data of the old batch is completely discarded, resulting in waste; (3) It takes 1-2 weeks from changing powder to resuming production, which seriously conflicts with the demand for tight development nodes and fast iteration pace of major equipment; (4) The re-experiment is still judged by the single index of density ρ, ignoring the synchronous drift of size error δ and surface roughness Ra, which has high quality risk and directly affects the equipment delivery node.

[0003] In recent years, traditional transfer learning (TrAdaBoost, TCA) has been introduced into SLM to recalibrate laser selective melting process parameters in this industry. However, the following bottlenecks exist: (1) Simple sample weighting does not consider the conditional distribution shift of the process parameter-target mapping, resulting in prominent negative transfer phenomena and still high prediction errors; (2) 10-20 sets of target domain experiments are required, which fails to break through the small sample limit; (3) Only transferring ρ single target cannot meet the stringent requirements of major equipment for dimensional accuracy and surface quality.

[0004] To address the aforementioned shortcomings, new research is needed on techniques for recalibrating process parameters. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems by providing a recalibration method for laser selective melting (SLM) process parameters. Under similar domain conditions where the difference in laser absorptivity between the source and target domain materials is ≤30% and the difference in thermal diffusivity is ≤25%, a domain adaptation and recalibration coupling strategy is employed to reduce the number of target domain calibration experiments to no more than 5 sets. This achieves simultaneous migration of multiple targets, including density, dimensional error, and roughness, with a density prediction RMSE ≤4% and dimensional error and roughness prediction R² ≥0.90. This shortens the new product introduction cycle, reduces costs, and provides crucial support for the large-scale industrial application of SLM technology.

[0006] The technical solution adopted in this invention is as follows: A method for recalibrating laser selective melting process parameters, comprising the following steps: S1: Construction of source domain pre-trained model; Based on multiple samples produced from the source and target domains, the process parameters and feature data of the source domain production samples are obtained to construct the source domain data. ; Obtain the process parameters and feature data of the production samples in the target domain to construct the target domain data. ;data It consists of source domain process-geometric characteristic parameters and source domain performance response characteristic parameters, data The source domain model consists of target domain process-geometric feature parameters and target domain performance response feature parameters. The MO-GPR algorithm is used to train the source domain model, with the source domain process-geometric feature parameters as input and the source domain performance response feature parameters as output. Ensure the source domain model The goodness of fit R² ≥ 0.85; the specific content of step S1 is as follows.

[0007] The source domain process-geometric feature parameters and the target domain process-geometric feature parameters are both The seven-dimensional vector, where P is the laser power, v is the scanning speed, h is the scanning interval, t is the layer thickness, θ is the overhang angle, d is the wall thickness, and A is the local cross-sectional area; The source domain performance response characteristic parameters and the target domain performance response characteristic parameters are: A three-dimensional vector, where ρ is density, δ is dimensional error, and Ra is surface roughness.

[0008] The source domain is a combination of source domain equipment and source domain powder, which is equivalent to the original equipment and original powder; the target domain is a combination of target domain equipment and target domain powder, which is equivalent to the new equipment and new powder; in this step S1, the main steps S11-S12 are included.

[0009] S11: Construct source domain data and target domain data This includes steps S111-S112.

[0010] S111: Construct source domain process-geometric feature parameters and target domain process-geometric feature parameters; Based on a strategy combining the single-variable method and orthogonal experimental design, the following parameters were obtained from the processes of the source domain production samples and the target domain production samples, respectively: Laser power P, scanning speed v, scanning spacing h, layer thickness t; Based on the characteristics of the overhanging structure, the following parameters are obtained: Three types of geometric constraint variables: overhang angle θ, wall thickness d, and local cross-sectional area A; The parameters obtained above are integrated and constructed to form The source domain process-geometric feature parameters of this seven-dimensional vector, and The target domain process-geometric feature parameters of this seven-dimensional vector.

[0011] S112: Construct the source domain performance response characteristic parameters and the target domain performance response characteristic parameters; Perform the following operations on the samples generated from the source domain and the samples generated from the target domain respectively: The compactness of the samples was determined by Archimedes' displacement method and metallographic method, respectively, and the weighted average of the results of the two methods was taken as the compactness ρ. The dimensions of key features on the sample are measured using a coordinate measuring machine. If the error between the measured value and the theoretical value is less than 0.5%, it is considered valid data. Each feature is measured no less than 3 times, and the average value of the resulting error is taken as the dimensional error δ. The sample was measured using a white light interferometer or a stylus profilometer. The sample was measured three times at three directions (0°, 45°, and 90°) at typical surface locations. After removing outliers, the arithmetic mean was taken as the surface roughness Ra to eliminate the influence of anisotropy. The parameters obtained above are integrated and constructed to form The source domain performance response characteristic parameters of this three-dimensional vector, and This is a three-dimensional vector representing the target domain performance response characteristic parameters.

[0012] S12: Source Domain Model train; The MO-GPR algorithm is used to train the source domain model, with source domain process-geometric feature parameters as input and source domain performance response feature parameters as output. Ensure the source domain model The goodness of fit R² ≥ 0.85; where: The similarity between samples in the input space is measured using the squared exponential kernel function, whose length scale parameter... Automatic optimization using the logarithmic marginal likelihood function reflects the sensitivity of each process parameter to its impact on quality, including the signal variance parameter. To control the fluctuation range of the model; further, considering that there is often a coupling relationship between the three quality indicators of density ρ, dimensional error δ and surface roughness Ra, an intrinsic core regionalization model is adopted to realize multi-output joint modeling. By constructing a shared latent variable space, the synergistic influence mechanism of process parameters on multiple quality indicators is captured, and the correlation information between quality indicators is fully utilized; finally, a source domain model is constructed to realize the nonlinear mapping relationship between source domain process-geometric features and source domain performance response features, which significantly improves the prediction accuracy and generalization performance of the model.

[0013] S2: Inter-domain difference measure; Normalize the process-geometric feature parameters of both the source and target domains, and perform one-hot encoding on the categorical variables to ensure that the feature spaces of the source and target domains have the same dimension. Set a threshold ε and calculate the difference in edge distribution. and conditional distribution differences ,when Initiate transfer learning and proceed to steps S3-S4; when The source domain model trained in step S1 is used directly. The specific details of step S2 are as follows, including steps S21-S23.

[0014] S21: Data processing; First, based on the source domain process-geometric feature parameters and target domain process-geometric feature parameters obtained in step S1, In order to eliminate the dimensional differences between parameters, Perform max-min normalization to map each parameter to the [0,1] interval; Subsequently, one-hot encoding is performed on powder grades (such as 316L, GH3625, TC4, etc.) and equipment models (such as EOS M290, BLT-S310, etc.); that is, assuming there are a types of powder and b types of equipment, the extended feature dimension is 7+a+b. Finally, the data is checked to ensure that the feature vector dimensions of the source domain and the target domain are consistent, so that the pre-trained MO-GPR can directly parse the target domain input without rebuilding the basic model architecture.

[0015] S22: Calculation and ; The edge distribution difference The Euclidean distance between the source and target domains, calculated using the mean of their feature spaces, is given by the following formula: ; in: The mean value of the source domain process-geometric characteristic parameters; The mean value of the process-geometric characteristic parameters of the target domain; The conditional distribution difference The prediction bias of the source domain model in the target domain is calculated using the following formula: ; in: Let sample x conform to the target domain data The expectation operator; For target domain data The number of samples in the sample; The measured performance of sample x in the target domain; This is the prediction output of the source domain model for the j-th sample in the target domain; Let be the true performance value of the j-th sample in the target domain.

[0016] S23: Determine whether transfer learning is necessary; Set the threshold ε = 0.15, if If the differences between the decision domains are negligible, the source domain model can be directly adopted. Further optimization will be carried out; if Initiate the S3 transfer learning process to recalibrate the model.

[0017] In step S23, the threshold ε is dynamically adjusted according to the material type and equipment model, with a default value of 0.15.

[0018] S3: Transfer learning recalibration; The TrAdaBoost-R algorithm is used as the input source domain data. and target domain data After initializing the sample weights, the MO-GPR algorithm is used to iteratively train the sub-models, with one sub-model obtained in each iteration. The source domain sample weights are dynamically updated based on the target domain prediction error, and then all sub-models are fused together by weighting by the inverse of the prediction error to form a single callable prediction function. The specific steps are as follows: S30-S34.

[0019] S30: Sample weight initialization; Source domain sample weights are The target domain sample weights are , This represents the number of iterations; that is, during the first training of the sub-model (without iterations), the sample weights are initialized as follows: the source domain sample weights are... The target domain sample weights are: ;in: For source domain data The number of samples in For source domain data The sample number; For target domain data The number of samples in For target domain data The sample number.

[0020] S31: Sub-model training; Source domain data The source domain process-geometric feature parameters are used as inputs, and the source domain data are used as inputs. The performance response feature parameters are used as outputs. The MO-GPR algorithm is applied to train the sub-model according to the updated sample weights to obtain the sub-model. Where t is the number of training iterations; the initialized sample weights (source domain sample weights are...) The target domain sample weights are: This only applies to the initial training of the sub-model; S32: Error analysis and calculation of the credibility weights of sub-models; With target domain data The target domain process-geometric feature parameters are taken as input, and the wheel model obtained in step S31 is used. Predictions are made on performance response characteristic parameters to obtain predicted data, which is then compared with the target domain data. The target domain performance response characteristic parameters are compared to obtain the parameter error value. and average relative error ;in: ; in, To predict the value of a certain feature parameter in the data; For this feature parameter in the target domain data The value in; ; in, The number of feature parameters in the predicted data is the number of feature parameters of the target domain performance response. Based on average relative error Calculate the credibility weight of the wheel model. .

[0021] S33: Source domain sample weight update; like or If so, only the weights of the source domain samples are updated; the method for updating the sample weights is: The inspection caused or The sample in the source domain data is called the event sample; the weight of the event sample is updated as follows: The weights of the remaining samples in the source domain data are updated to... ;in, For the number of iterations, " " is the number of the event sample, " "This represents the number of event samples;" "This is the number for the other samples," " is the number of other samples, there are .

[0022] S34: Iterative training to obtain a new sub-model; Repeat steps S31-S33 with the updated source domain sample weights to obtain the sub-model set. Simultaneously, the model for each wheel is obtained based on the calculations in step S32. Credibility weight To obtain the weight set .

[0023] S35: Sub-model integration; Using sub-model sets and weight set Construct a weighted ensemble model to form a single, callable prediction function. .

[0024] S4: Parameter optimization; The NSGA-II algorithm is used to predict the function in step S3. Search the Pareto front and use the TOPSIS method to select the optimal process parameter solution from the Pareto front to complete the recalibration of the laser selective melting process parameters; detailed steps for selecting the optimal process parameter solution are shown in steps S41-S43.

[0025] S41: Based on prediction function A multi-objective optimization model is established, with P, v, h, t as process constraints, θ, d, A as geometric constraints, and ρ, δ, Ra as quality constraints. .

[0026] S42: Use the NSGA-II genetic algorithm for real-number encoding of the multi-objective optimization model in step S41. Pareto frontier search is performed to determine the population size and number of iterations. Real number encoding is set, and binary crossover and polynomial mutation are simulated to obtain the non-dominated solution set between process parameters and quality indicators.

[0027] S43: The TOPSIS approximation ideal solution sorting method is used to select the process parameter combination with the best comprehensive performance from the non-dominated solution set as the optimal process parameter solution.

[0028] Furthermore, a process card is generated based on the obtained optimal process parameters and then pushed to the equipment's control system.

[0029] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The method disclosed in this invention couples the TrAdaBoost-R algorithm with the MO-GPR algorithm in the SLM field. By automatically suppressing source domain data that conflicts with the target domain through weighted samples, it significantly reduces the risk of negative migration, and the density prediction RMSE is ≤4%. 2. The method disclosed in this invention proposes two indicators: edge distribution difference Dm and conditional distribution difference Dc, to automatically determine the necessity of migration after toner / machine replacement; when The source domain model is directly used, avoiding unnecessary computation and saving ≥90% of computing time. 3. The method disclosed in this invention is compatible with the simultaneous migration of three targets: density ρ, size error δ, and surface roughness Ra. It does not require separate modeling for each target, and can be recalibrated with ≤5 experimental samples. Attached Figure Description

[0030] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a flowchart illustrating step S1 in the present invention; Figure 3 This is a flowchart illustrating step S2 in this invention; Figure 4 This is a flowchart illustrating step S3 in this invention. Detailed Implementation

[0031] It should be noted that, in this specification, the recalibration method for the parameters of this laser selective melting process may result in... If the differences between the decision domains are negligible, the source domain model can be directly adopted. Subsequent optimization is then performed; that is, no transfer learning recalibration is required. Therefore, no adjustment of process parameters is necessary, and consequently, this type of implementation will not be described in detail in this specification. This specification focuses on... This requires an implementation method for transfer learning. Details are as follows.

[0032] like Figures 1-4 As shown, a method for recalibrating laser selective melting process parameters is illustrated, taking the migration of 316L stainless steel additive manufacturing process from EOS M290 (source domain) to BLT-S310 (target domain) as an example. The difference in laser absorptivity between the source domain and the target domain is ≤30%, and the difference in thermal diffusivity is ≤25%. The source domain is a combination of EOS M290 (laser wavelength 1070nm, spot diameter 100μm) and EOS 316L powder (particle size 15-45μm, oxygen content ≤300ppm); the target domain is BLT-S310 (laser wavelength 1064nm, spot diameter 80μm) and EOS 316L powder. The method includes steps S1-S4.

[0033] S1: Source domain pre-trained model construction; such as Figure 2 As shown, it includes steps S11-S12.

[0034] S11: Construct source domain data and target domain data ; Among them, source domain data It has 25 groups, which is equivalent to having 25 samples; source domain data See Table 1.

[0035] Table 1: Source domain data (25 groups).

[0036]

[0037] Among them, target domain data It has 5 groups, which is equivalent to having 5 samples; source domain data See Table 2.

[0038] Table 2: Target domain data (5 groups).

[0039]

[0040] S12: Source Domain Model train; The MO-GPR algorithm is used to train the source domain model, taking the source domain process-geometric feature parameters from 25 sets of source domain data as input and the source domain performance response feature parameters as output. Ensure the source domain model The goodness of fit R² ≥ 0.85; where: after log-marginal likelihood optimization, the length scale parameter Signal variance parameter .

[0041] S2: Inter-domain difference measure; such as Figure 3 As shown, it includes steps S21-S23.

[0042] S21: Data processing; First of all, Perform max-min normalization to map each parameter to the [0,1] interval; Subsequently, one-hot encoding was performed on the powder EOS 316L, the EOS M290 device, and the BLT-S310 device, and the expanded feature dimension was 10. Finally, ensure that the feature vector dimensions of the source domain and the target domain are consistent.

[0043] S22: Calculation and ; in: The source domain model obtained in step S12 Directly applied to target domain data In (5 groups), the average error of predicted density was 0.56%, the average error of dimensional error was 31%, and the average error of roughness was 21%. The calculated values ​​are... .

[0044] S23: Determine whether transfer learning is necessary; Set the threshold ε=0.15, based on It was determined that the S3 transfer learning process needed to be initiated to recalibrate the model.

[0045] S3: Transfer learning recalibration; such as Figure 4 As shown, it includes steps S30-S34.

[0046] S30: Sample weight initialization; Based on source domain data With 25 samples, target domain data There are 5 samples; therefore, the sample weights are initialized as follows: the source domain sample weights are... The target domain sample weights are: , For source domain data The sample size is 25. For source domain data The sample number; For target domain data The number of samples in the sample is equal to 5. For target domain data The sample number.

[0047] S31: Sub-model training; Sub-model Training: Based on the source domain sample weights The target domain sample weights are: Source domain data The source domain process-geometric feature parameters are used as inputs, and the source domain data are used as inputs. The performance response feature parameters are used as outputs. The MO-GPR algorithm is applied to train the sub-model to obtain the sub-model. ; Other wheel model training: using source domain data The source domain process-geometric feature parameters are used as inputs, and the source domain data are used as inputs. The performance response feature parameters are used as outputs. The MO-GPR algorithm is applied to train the sub-model according to the updated sample weights to obtain the sub-model. , where t is the number of training iterations.

[0048] S32: Error analysis and calculation of the credibility weights of sub-models; With target domain data The target domain process-geometric feature parameters are taken as input, and the wheel model obtained in step S31 is used. Predictions are made on performance response characteristic parameters to obtain predicted data, which is then compared with the target domain data. The target domain performance response characteristic parameters are compared to obtain the parameter error value. and average relative error Based on the average relative error Calculate the credibility weight of the wheel model. .in, ; .

[0049] S33: Source domain sample weight update; like or If so, only the weights of the source domain samples are updated; the method for updating the sample weights is: The inspection caused or The sample in the source domain data is called the event sample; the weight of the event sample is updated as follows: The weights of the remaining samples in the source domain data are updated to... ;in, For the number of iterations, " " is the number of the event sample, " "This represents the number of event samples;" "This is the number for the other samples," " is the number of other samples, there are .

[0050] have If so, only the weights of the source domain samples are updated; the method for updating the sample weights is: The inspection caused The sample from the source domain data was found to be the first Group and No. The third and seventh groups of samples caused a large prediction bias (deviation from the measured target domain > 30%), and these samples are referred to as event samples; the weights for these event samples are updated as follows: The weights of the remaining samples in the source domain data are updated to... .

[0051] S34: Iterative training to obtain a new sub-model; Repeat steps S31-S33 for 10 iterations using the updated source domain sample weights to obtain a set of 10 sub-models. Simultaneously, the model for each wheel is obtained based on the calculations in step S32. Credibility weight To obtain the weight set .

[0052] Such as sub-model Training: Source domain sample weights updated according to step S33 ( , ), with source domain data The source domain process-geometric feature parameters are used as inputs, and the source domain data are used as inputs. The performance response feature parameters are used as outputs. The MO-GPR algorithm is applied to train the sub-model to obtain the sub-model. ; And so on, until the sub-model is obtained. and .

[0053] S35: Sub-model integration; Using sub-model sets and weight set Construct a weighted ensemble model to form a single, callable prediction function. .

[0054] S4: Parameter optimization; including steps S41-S43.

[0055] S41: Based on prediction function A multi-objective optimization model is established, with P, v, h, t as process constraints, θ, d, A as geometric constraints, and ρ, δ, Ra as quality constraints. ;in: , , .

[0056] S42: Use the NSGA-II genetic algorithm for real-number encoding of the multi-objective optimization model in step S41. Pareto front search was performed, with a population size of 50 individuals and 30 iterations. Real number encoding was set, and after simulating binary crossover (probability 0.9) and polynomial mutation (probability 1 / 7), the non-dominated solution set between process parameters and quality indicators was obtained, i.e., 12 Pareto front solution sets were obtained, with the quality indicator distribution as follows: ρ (99.52%-99.71%), δ (0.06%-0.09%), Ra (4.2-5.8 μm).

[0057] S43: The TOPSIS approximation ideal solution ranking method is adopted, that is, a normalized decision matrix is ​​constructed, the positive ideal solution [ρ=99.71%, δ=0.06%, Ra=4.2μm] is determined, the closeness of each solution is calculated, and the three optimal process parameter combinations with comprehensive performance are selected from the non-dominated solution set as the optimal process parameter solutions. The three process parameter combinations are shown in Table 3.

[0058] Table 3: Combination of process parameters.

[0059]

[0060] Based on the obtained optimal process parameters, a process card is generated and pushed to the equipment's control system; that is, combinations 1-3 are pushed to the BLT-S310 PLC system via the OPC UA protocol. Combination 1 was selected for trial production on-site, and the measured density was 99.66%, dimensional error was 0.075%, and roughness was 4.6μm, with a deviation of <3% from the predicted values, verifying the effectiveness of the migration. After processing, the quality data is fed back to the system to update the inter-domain difference database.

[0061] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for recalibrating laser selective melting process parameters, characterized in that: Includes the following steps: S1: Construction of source domain pre-trained model; Based on multiple samples produced from the source and target domains, the process parameters and feature data of the source domain production samples are obtained to construct the source domain data. ; To construct target domain data, process parameters and feature data of production samples from the target domain are obtained. Source domain data The target domain data consists of source domain process-geometric characteristic parameters and source domain performance response characteristic parameters. The source domain model consists of target domain process-geometric feature parameters and target domain performance response feature parameters. The MO-GPR algorithm is used to train the source domain model, with the source domain process-geometric feature parameters as input and the source domain performance response feature parameters as output. Ensure the source domain model The goodness of fit R² ≥ 0.85; S2: Inter-domain difference measure; The process-geometric feature parameters of the source and target domains are normalized, and one-hot encoding is performed on the categorical variables to ensure that the feature spaces of the source and target domains have the same dimension. A threshold ε is set, and the difference in edge distribution is calculated. and conditional distribution differences ,when Initiate transfer learning and proceed to steps S3-S4; when The source domain model trained in step S1 is used directly. ; S3: Transfer learning recalibration; The TrAdaBoost-R algorithm is used as the input source domain data. and target domain data After initializing the sample weights, the MO-GPR algorithm is used to iteratively train the sub-models, with one sub-model obtained in each iteration. The source domain sample weights are dynamically updated based on the target domain prediction error, and then all sub-models are fused together by weighting by the inverse of the prediction error to form a single callable prediction function. ; S4: Parameter optimization; The NSGA-II algorithm is used to predict the function in step S3. By searching the Pareto front and selecting the optimal process parameter solution from the Pareto front using the TOPSIS method, the laser selective melting process parameters are recalibrated.

2. The recalibration method according to claim 1, characterized in that: In step S2, the edge distribution difference The Euclidean distance between the source and target domains, calculated using the mean of their feature spaces, is given by the following formula: ； in: The mean value of the source domain process-geometric characteristic parameters; The mean of the process-geometric characteristic parameters of the target domain.

3. The recalibration method according to claim 1, characterized in that: In step S2, the conditional distribution difference The prediction bias of the source domain model in the target domain is calculated using the following formula: ； in: For the sample x Obey the target domain data The expectation operator; For samples in the target domain x The corresponding measured performance.

4. The recalibration method according to claim 1, characterized in that: The default value of the threshold ε mentioned in step S2 is 0.

15.

5. The recalibration method according to claim 1, characterized in that: The threshold ε mentioned in step S2 is dynamically adjusted according to the material type and equipment model.

6. The recalibration method according to claim 1, characterized in that: After initializing the sample weights in step S3, the source domain sample weights are: The target domain sample weights are: ; in: For source domain data The number of samples in For source domain data The sample number; For target domain data The number of samples in For target domain data The sample number.

7. The recalibration method according to claim 6, characterized in that: In step S3, the steps for obtaining the prediction function are as follows: S31: Using source domain data The source domain process-geometric feature parameters are used as inputs, and the source domain data are used as inputs. The performance response feature parameters are used as outputs. The MO-GPR algorithm is applied to train the sub-model according to the updated sample weights to obtain the sub-model. , where t is the number of training iterations; Initializing sample weights is only applied during the initial training of the sub-model; S32: Target domain data The target domain process-geometric feature parameters are taken as input, and the wheel model obtained in step S31 is used. Predictions are made on performance response characteristic parameters to obtain predicted data, which is then compared with the target domain data. The target domain performance response characteristic parameters are compared to obtain the parameter error value. and average relative error ;in: ； in, To predict the value of a certain feature parameter in the data; For this feature parameter in the target domain data The value in; ； in, The number of feature parameters in the predicted data; Based on average relative error Calculate the credibility weight of the wheel model. ; S33: If or If so, only the weights of the source domain samples are updated; the method for updating the sample weights is: The inspection caused or The sample in the source domain data is called the event sample; the weight of the event sample is updated as follows: The weights of the remaining samples in the source domain data are updated to... ;in, " is the number of iterations, "This is the event sample number." "This represents the number of event samples;" "This is the number for the other samples." " is the number of other samples, there are ; S34: Repeat steps S31-S33 with the updated source domain sample weights to obtain the sub-model set. Simultaneously, the model for each wheel is obtained based on the calculations in step S32. Credibility weight To obtain the weight set ; S35: Utilizing the set of sub-models and weight set Construct a weighted ensemble model to form a single, callable prediction function. .

8. The recalibration method according to claim 1, characterized in that: The source domain process-geometric feature parameters and the target domain process-geometric feature parameters are both The seven-dimensional vector, where P is the laser power, v is the scanning speed, h is the scanning interval, t is the layer thickness, θ is the overhang angle, d is the wall thickness, and A is the local cross-sectional area; The source domain performance response characteristic parameters and the target domain performance response characteristic parameters are: A three-dimensional vector, where ρ is density, δ is dimensional error, and Ra is surface roughness.

9. The recalibration method according to claim 8, characterized in that: The steps to select the optimal process parameter solution are as follows: S41: Based on prediction function A multi-objective optimization model is established with P, v, h, t as process constraints, θ, d, A as geometric constraints, and ρ, δ, Ra as quality constraints. ; S42: Use the NSGA-II genetic algorithm with real-number encoding to optimize the multi-objective model in step S41. Pareto frontier search is performed to determine the population size and number of iterations. Real number encoding is set, and binary crossover and polynomial mutation are simulated to obtain the non-dominated solution set between process parameters and quality indicators. S43: The TOPSIS approximation ideal solution sorting method is used to select the process parameter combination with the best comprehensive performance from the non-dominated solution set as the optimal process parameter solution.

10. The recalibration method according to any one of claims 1-9, characterized in that: Based on the obtained optimal process parameters, a process card is generated and pushed to the equipment's control system.

Images