FDM printing hole size precision prediction and process parameter optimization method

Abstract

The present application relates to the technical field of FDM printing, and particularly relates to a kind of FDM printing hole size precision prediction and process parameter optimization method, comprising: constructing FDM hole size reference data set;Train forward weighted process mechanism model, expand and enhance to obtain enhanced mechanism characteristics;Train weighted hierarchical target decomposition model, obtain overall trend term and process fluctuation residual label;Train four tree model expert branch;Weighted construction final multi-expert prediction model, joint construction forward joint prediction model;Adopt reverse two-stage grid search method, reverse solution obtains optimal process parameter combination;Prediction method can make full use of hole forming process mechanism, overcome the problem that prediction is sensitive to sample change and segmented is not smooth, output more reliable FDM printing hole size precision prediction result;Optimization method can carry out high-resolution search optimization under the premise of real printing, model training is always carried out in training fold, to avoid overfitting.

Description

Technical Field

[0001] This invention relates to the field of FDM printing technology, and more specifically to a method for predicting the accuracy of FDM printing hole size and optimizing process parameters. Background Technology

[0002] Fused deposition modeling (FDM) printing is widely used for the low-cost, rapid manufacturing of customized parts and functional prototypes. Hole structures, as the most common local features in mechanical parts, directly serve positioning, assembly, and connection functions. However, due to the high thermal shrinkage of fused thermoplastic filaments (PLA) during extrusion deposition, the discrepancy in the circular path approximation caused by interlayer stacking, and the strong nonlinear coupling interference of process parameters such as layer height, nozzle temperature, printing speed, and infill rate, the actual hole diameter of the printed parts is generally smaller than the designed hole diameter, exhibiting a consistent inward shrinkage characteristic. This deviation in hole size directly worsens the fit clearance of assembly features, leading to connection failure or assembly difficulties.

[0003] Currently, in order to reduce hole size deviation in FDM printing, operators often make empirical compensations based on the nominal hole diameter. However, this only optimizes the nominal hole diameter. When process parameters such as printing temperature and printing speed change, due to the non-monotonicity of the deformation mechanism, the experience cannot be extended across parameters, which consumes a lot of trial and error time.

[0004] To predict the dimensional accuracy of FDM printing and optimize process parameters, some existing technologies attempt to use multiple linear regression, orthogonal experiments, or response surface methodology to establish the coupling relationship between printing process parameters. However, these technologies have extremely limited ability to express the strong nonlinear coupling relationship between process parameters in FDM printing and the dimensional effect caused by small aperture holes, resulting in low generalization accuracy for cross-size prediction. Other existing technologies attempt to use machine learning methods to predict FDM printing hole size deviations. However, the convolutional networks, dense neural networks, or shallow backpropagation models used are prone to severe overfitting in scenarios with limited sample size of process parameters, resulting in limited prediction accuracy. Furthermore, they cannot utilize the inherent process mechanism of FDM printing and lack interpretability.

[0005] Therefore, it is necessary to provide a method for predicting the accuracy of FDM printing hole size and optimizing process parameters. This method can be based on the inherent process mechanism of FDM printing, utilize a limited sample of process parameters, more accurately predict the accuracy of FDM printing hole size, and perform reverse optimization of process parameters. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method and application for predicting FDM printing hole size accuracy and optimizing process parameters, thereby resolving the technical issues raised in the background section.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: A method for predicting the size accuracy of FDM printed holes and optimizing process parameters includes the following steps: S1: Collect process parameters and aperture errors of different printed samples, combine process parameters to construct process parameter feature vectors, build FDM aperture size benchmark dataset, create supervision labels based on aperture error, and divide the FDM aperture size benchmark dataset into k training folds. S2: Within each training iteration, construct and train a forward weighted process mechanism model, input process parameter feature vectors, and based on the hole forming process mechanism in FDM printing, generate a 19-dimensional candidate mechanism feature vector matrix through nonlinear transformation and algebraic mapping, and generate an enhanced mechanism feature matrix after weighting. S3: Within each training iteration, construct and train a weighted hierarchical target decomposition model, decompose the supervision labels, calculate the overall trend term through internal cross-validation, and separate the process fluctuation residual labels. S4: Within each training iteration, based on the enhancement mechanism feature matrix and the process fluctuation residual label, four tree model expert branches are trained respectively. The optimal deployment weights of each tree model expert branch are solved. The output results of each tree model expert branch are weighted and combined using the optimal deployment weights, and the overall trend term is accumulated to output the aperture error prediction value. S5: Allocate the best deployment weights to the expert branches of each tree model, construct the final multi-expert prediction model, and combine the forward weighted process mechanism model, the weighted hierarchical target decomposition model, and the final multi-expert prediction model to construct the forward joint prediction model; S6: Set the target aperture of the FDM printed part. For each preset process parameter combination, use the reverse two-stage grid search method, call the forward joint prediction model to perform reverse solution, construct a multi-dimensional collaborative comprehensive scoring function, and calculate and sort the optimal process parameter combination based on the aperture error prediction value.

[0008] Furthermore, the specific method of S2 is as follows: A forward weighted process mechanism model was constructed, classifying the hole forming process mechanism in FDM printing into size effect mechanism, thermal-velocity coupling mechanism, and geometry-process interaction mechanism. By utilizing the size effect mechanism, the thermal-velocity coupling mechanism, and the geometry-process interaction mechanism, and through nonlinear transformation and algebraic mapping, the process parameter features in the process parameter feature vector are extended, resulting in 4 size effect extended features, 4 thermal-velocity coupling features, and 6 geometry-process interaction features. The process parameter features include the design nominal aperture D, layer height H, printing temperature T, printing speed V, and fill rate F. Process parameter features, size effect extension features, thermal-velocity coupling features, and geometry-process interaction features are used as candidate mechanism features. Within each training iteration, 19 candidate mechanism features from different samples are combined to form a 19-dimensional candidate mechanism feature vector matrix. After scaling transformation, the baseline mechanism feature matrix is ​​obtained. The ridge regression method is used to learn the normalized continuous weights of the candidate mechanism feature vector matrix. The forward weighted process mechanism model is trained within the training process and the enhanced mechanism feature matrix is ​​output.

[0009] Furthermore, the specific method of S3 is as follows: Within each training iteration, a weighted hierarchical target decomposition model is constructed, and the mean value of the supervised label under each designed nominal aperture D is calculated in groups to obtain discrete mean curves. A cubic polynomial is used to fit the discrete mean curve to the overall trend curve. ; Define the overall trend item ,

[0010] In the formula, λ is the fusion weight. Within the training layer, a weighted hierarchical target decomposition model is trained through cross-validation adaptive learning to obtain fusion weights that make the overall trend estimate approximate the aperture error. Separate calculation of process fluctuation residual label r, y is the supervisory label.

[0011] Furthermore, the specific method for S4 to construct the final multi-expert prediction model is as follows: Four tree model expert branches are built in parallel, including the basic CART branch, the random forest Bagging branch, the GBR-Boostin branch, and the GBR residual branch. The process parameter feature vector is used as the input to the basic CART branch and the random forest Bagging branch, the candidate mechanism feature vector matrix is ​​used as the input to the GBR-Boostin branch, the enhancement mechanism feature matrix is ​​used as the input to the GBR residual branch, and the process fluctuation residual label is used as the supervision constraint. The four tree model expert branches are trained in each training layer. Outside of the training process, each tree model expert branch is run to obtain its corresponding predicted process fluctuation residual. The optimal deployment weights of each tree model expert branch are then determined by non-negative least squares algorithm constraints. The predicted process fluctuation residuals are weighted and combined using the optimal deployment weights to obtain the final predicted value of the process fluctuation residuals. The final predicted value of the process fluctuation residuals is then added to the overall trend estimation term to obtain the final predicted value of the aperture error.

[0012] Furthermore, the reverse two-stage grid search method in S6 is specifically as follows: Set the target aperture desired by the user and construct a multi-dimensional discrete search space. The multi-dimensional discrete search space is set with multiple preset printing parameter combinations formed by orthogonal combinations of different layer heights, different printing temperatures, different printing speeds, and different fill rates. Each preset printing parameter combination and the discrete design nominal aperture combination form a preset process parameter combination. The design nominal aperture range is 1.5mm to 20.8mm. Within the multidimensional discrete search space, a coarse grid step size of 0.02 mm is first used as the nominal aperture of the design. The forward joint prediction model is then called to calculate the predicted values ​​of the actual aperture under each preset process parameter combination in batches, and the coarse search grid points that make the predicted values ​​of the actual aperture closest to the target aperture are located. Within a local neighborhood of 0.035 mm before and after the coarse search grid point, with a fine grid step size of 0.001 mm as the nominal aperture, the forward joint prediction model is called again to locate the fine search grid point in the local neighborhood that makes the actual aperture prediction value close to the target aperture. The preset process parameter combination at the fine search grid point is used as the preferred process parameter combination.

[0013] Furthermore, the specific method for calling the forward joint prediction model in S6 is as follows: Input the preset process parameter combination into the forward joint prediction model; The forward-weighted process mechanism model generates its corresponding enhancement mechanism feature matrix; The weighted hierarchical target decomposition model generates its corresponding overall trend term; Finally, the multi-expert prediction model generates the corresponding final aperture error prediction value based on the enhancement mechanism feature matrix and the overall trend term. The final aperture error prediction value is added to the design nominal aperture in the preset process parameter combination to obtain the actual aperture prediction value.

[0014] Furthermore, the specific method for obtaining the optimal combination of process parameters in S6 is as follows: Construct a multidimensional collaborative comprehensive scoring function J.

[0015] In the formula, To assess the predictive reliability of preset process parameter combinations, R is the final predicted aperture error, R is the numerical inverse solution residual, and Q is the extrapolation penalty distance for exceeding the model training range. Calculate the multidimensional synergistic comprehensive scoring function for each preferred combination of process parameters, sort them in ascending order, and take the preferred combination of process parameters with the lowest score as the optimal combination of process parameters.

[0016] Furthermore, the method for calculating the prediction confidence is as follows: Store the out-of-place prediction absolute residuals of the feature vectors of different process parameters during the training phase of the forward joint prediction model, and construct the out-of-place prediction absolute residual matrix; Between the process parameter feature vectors where the designed nominal aperture is close to the preset process parameter combination, the out-of-place prediction absolute residual corresponding to the designed nominal aperture of the preset process parameter combination is obtained by interpolation and used as the local out-of-place prediction absolute residual. The global average absolute residual of out-of-place prediction is obtained by averaging the out-of-place prediction absolute residual matrix. The prediction reliability is obtained by weighted summing of the absolute residuals of the local out-of-range prediction and the absolute residuals of the global average out-of-range prediction.

[0017] Furthermore, the specific method for extending the mechanism using the size effect mechanism, the thermal-velocity coupling mechanism, and the geometry-process interaction mechanism is as follows: Based on the size effect mechanism, the size effect extension characteristics are derived, including: aperture squared. reciprocal of aperture Reciprocal of the square of the aperture Logarithm of aperture ln(D); Based on the heat-velocity coupling mechanism, the heat-velocity coupling characteristics are extended to include: temperature-velocity ratio. The product of heat rate T×V, and the square of temperature velocity square ; Based on the geometry-process interaction mechanism, the extended geometry-process interaction features include: velocity-layer-height product V×H, and layer-height-aperture ratio. velocity aperture ratio , temperature-to-heat ratio The product of the filling layer height F×H and the product of the filling speed F×V.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention provides a method for predicting FDM printing hole size accuracy and optimizing process parameters. Through a forward weighted process mechanism model, it can fully utilize the hole forming process mechanism in FDM printing, providing robust input rich in physical mechanism semantics for the final multi-expert prediction model. Through a weighted hierarchical target decomposition model, the hole diameter error is decomposed into an overall trend term and a process fluctuation residual label, helping the final multi-expert prediction model overcome the problems of sensitivity to sample changes and unsmooth segmentation. The final multi-expert prediction model complements and superimposes the prediction advantages of each tree model expert branch, has better cross-size prediction capability, and can output more reliable FDM printing hole size accuracy prediction results.

[0019] 2. The present invention provides a method for predicting the accuracy of FDM printing hole size and optimizing process parameters. It adopts a reverse two-stage grid search method to perform reverse solving on the forward joint prediction model. Under the premise of ensuring that the searched preset process parameter combination can be printed in reality, high-resolution search optimization is performed, which can accurately express the size effect under small hole diameter and ensure that the printed hole size can be effectively assembled.

[0020] 3. The present invention provides a method for predicting the accuracy of FDM printing hole size and optimizing process parameters. During the training phase, the forward joint prediction model is always performed within the training fold, and cross-validation is performed in different training folds to avoid information leakage and prevent overfitting due to the limited number of printing samples in PLA material printing data. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the FDM printing hole size accuracy prediction and process parameter optimization method described in this embodiment; Figure 2 This is a schematic diagram of the forward joint prediction model structure described in this embodiment; Figure 3 This is a schematic diagram of the process of S6 described in this embodiment. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention provides a method for predicting the accuracy of FDM printing hole dimensions and optimizing process parameters, such as... Figure 1 As shown, the method includes the following steps: S1: Collect process parameters and aperture errors of different printed samples, combine process parameters to construct process parameter feature vectors, build FDM aperture size benchmark dataset, create supervision labels based on aperture error, and divide the FDM aperture size benchmark dataset into k training folds. Furthermore, the specific method for constructing the FDM pore size benchmark dataset in S1 is as follows: taking the same design nominal pore diameter as an experimental batch, collecting PLA material printing data based on a four-factor, three-level full factorial experimental design for each experimental batch, and using the interquartile range method grouped by the design nominal pore diameter to perform normalization preprocessing on the PLA material printing data to obtain the FDM pore size benchmark dataset, including the process parameters and pore diameter errors of different printing samples; A process parameter feature vector is constructed by combining process parameters to build an FDM aperture size benchmark dataset. Aperture error is used as a supervision label y. The process parameter features in the process parameter feature vector include the designed nominal aperture D, layer height H, printing temperature T, printing speed V, and fill rate F. The aperture error is the difference between the measured aperture and the designed nominal aperture. The interquartile range method can identify and retain the real physical response generated by specific extreme combinations of process parameters, prohibit global blind tailing, and maintain the strict orthogonality and balance of the FDM aperture size benchmark dataset.

[0025] Furthermore, the FDM aperture size benchmark dataset is divided into k training folds, so that subsequent training is carried out in different training folds and cross-validation is performed in different training folds. This avoids information leakage from the FDM aperture size benchmark dataset and prevents overfitting due to the limited number of printing samples in the PLA material printing data. In this embodiment, k=5.

[0026] S2: Within each training iteration, construct and train a forward weighted process mechanism model, input process parameter feature vectors, and based on the hole forming process mechanism in FDM printing, generate a 19-dimensional candidate mechanism feature vector matrix through nonlinear transformation and algebraic mapping, and generate an enhanced mechanism feature matrix after weighting. A forward weighted process mechanism model was constructed, classifying the hole forming process mechanism in FDM printing into size effect mechanism, thermal-velocity coupling mechanism, and geometry-process interaction mechanism. By utilizing the size effect mechanism, the thermal-velocity coupling mechanism, and the geometry-process interaction mechanism, respectively, and through nonlinear transformation and algebraic mapping, the process parameter features in the process parameter feature vector are extended, resulting in 4 size effect extended features, 4 thermal-velocity coupling features, and 6 geometry-process interaction features. Among them, the size effect extension features include: aperture squared reciprocal of aperture Reciprocal of the square of the aperture The logarithm of aperture ln(D) and the size effect extension feature are used to characterize the significant deformation size effect and nonlinear shrinkage trend under small aperture. Thermal-velocity coupling characteristics include: temperature-velocity ratio The product of heat rate T×V, and the square of temperature velocity square Among them, the temperature-rate ratio is used to characterize the amount of heat input per unit time, and the square of the rate is used to characterize the rheological behavior and shear thinning effect of PLA material during extrusion. Geometric-process interaction features include velocity-to-layer-height product V×H and layer-height-to-aperture ratio. velocity aperture ratio , temperature-to-heat ratio The product of the filling layer height F×H and the product of the filling speed F×V are used to characterize the interaction boundary between the interlayer stacking extrusion pressure and the macroscopic geometry.

[0027] Process parameter features, size effect extension features, thermal-velocity coupling features, and geometry-process interaction features are used as candidate mechanism features. Within each training iteration, 19 candidate mechanism features from different samples are combined to form a 19-dimensional candidate mechanism feature vector matrix. After scaling transformation, the baseline mechanism feature matrix is ​​obtained. The ridge regression method is used to learn the normalized continuous weights of the candidate mechanism feature vector matrix, train the forward weighted process mechanism model, and output the enhanced mechanism feature matrix. ,

[0028] In the formula, i is the sample number, i=1,2,3,…, j is the candidate mechanism feature number, j=1,2,…19, z is the candidate mechanism feature, W is the normalized continuous weight, its value range is [0,1], the sum of each normalized continuous weight is 1, and x is the element in the feature matrix, that is, the enhancement mechanism feature.

[0029] Furthermore, by constructing a forward weighted process mechanism model, the hole forming process mechanism in FDM printing can be fully utilized, and the input process parameter features can be extended and enhanced into enhanced mechanism features. This provides a robust input rich in physical mechanism semantics for the subsequent prediction of hole size accuracy, reducing the risk of overfitting in hole size accuracy prediction.

[0030] S3: Within each training iteration, construct and train a weighted hierarchical target decomposition model, decompose the supervision labels, calculate the overall trend term through internal cross-validation, and separate the process fluctuation residual labels. Furthermore, the specific method of S3 is as follows: A weighted hierarchical target decomposition model is constructed. Within the training iteration, groups with the same nominal aperture D are grouped together, and the mean value of the supervision label under each nominal aperture is calculated to obtain the discrete mean curve. A cubic polynomial is used to fit the discrete mean curve to the overall trend curve. ; Define the overall trend term dominated by the design nominal aperture D. ,

[0031] In the formula, λ is the fusion weight. The fusion weights are discretized into {0.0, 0.1, ..., 1.0}. Within the training fold, a weighted hierarchical target decomposition model is trained through cross-validation adaptive learning to obtain the fusion weights that make the overall trend estimate approximate the aperture error. In this embodiment, λ=0.7 is obtained. Subtracting the aperture error from the overall trend term, the residual label r of the process fluctuation triggered by the specific process environment is calculated separately. Running a weighted hierarchical target decomposition model can output the overall trend term and process fluctuation residual label under different design nominal apertures.

[0032] Furthermore, the aperture error is decomposed into an overall trend term and a process fluctuation residual label by a weighted hierarchical target decomposition model. The process fluctuation residual label can purely reflect the fluctuation of process parameters, which helps to overcome the problems of sensitivity to sample changes and unsmooth segmentation in subsequent aperture size accuracy prediction.

[0033] S4: Within each training iteration, based on the enhancement mechanism feature matrix and the process fluctuation residual label, four tree model expert branches are trained respectively. The optimal deployment weights of each tree model expert branch are solved. The output results of each tree model expert branch are weighted and combined using the optimal deployment weights, and the overall trend term is accumulated to output the aperture error prediction value. Furthermore, the specific method of S4 is as follows: Four tree model expert branches with different bias-variance fitting characteristics were constructed in parallel: a basic CART branch, a random forest Bagging branch, a GBR-Boostin branch, and a GBR residual branch. The process parameter feature vectors were used as input to the basic CART branch to preserve baseline boundaries; the process parameter feature vectors were used as input to the random forest Bagging branch to reduce model variance; the candidate mechanism feature vector matrix was used as input to the GBR-Boostin branch to reduce bias; and the enhanced mechanism feature matrix was used as input to the GBR residual branch to focus on the details of the FDM printing process mechanism. The process fluctuation residual labels were used as supervisory constraints to train the four tree model expert branches within the training period, avoiding data leakage and preventing overfitting. Outside of the training process, each tree model expert branch is run to obtain its corresponding predicted process fluctuation residual. The optimal deployment weights of each tree model expert branch are then determined using a non-negative least squares algorithm. Simultaneously, a 2% performance prior smoothing constraint is injected into the determination of the optimal deployment weights to prevent unordered focusing of weights due to occasional interference. In this embodiment, the optimal deployment weights of the base CART branch, random forest Bagging branch, GBR-Boostin branch, and GBR residual branch are 0.005, 0.158, 0.530, and 0.307, respectively. The predicted process fluctuation residuals are weighted and combined using the optimal deployment weights to obtain the final predicted value of the process fluctuation residuals. The final predicted value of the process fluctuation residuals is then added to the overall trend estimation term to obtain the final predicted value of the aperture error.

[0034] S5: Allocate the best deployment weights to the expert branches of each tree model, construct the final multi-expert prediction model, and combine the forward weighted process mechanism model, the weighted hierarchical target decomposition model, and the final multi-expert prediction model to construct the forward joint prediction model; By using the optimal deployment weights of each tree model expert branch, a final multi-expert prediction model including four tree model expert branches is constructed. The final multi-expert prediction model complements and superimposes the prediction advantages of each tree model expert branch, and can output more reliable FDM printing hole size accuracy prediction results. By combining the forward weighted process mechanism model, the weighted hierarchical target decomposition model, and the final multi-expert prediction model, a model is constructed as follows: Figure 2 The forward joint prediction model is shown below. When the forward joint prediction model is invoked, for a set of preset process parameter combinations, after inputting them into the forward joint prediction model, the forward weighted process mechanism model generates its corresponding candidate mechanism feature vector matrix and enhanced mechanism feature matrix. The weighted hierarchical target decomposition model generates its corresponding overall trend term. Finally, the multi-expert prediction model generates its corresponding final process fluctuation residual prediction value based on the preset process parameter features, candidate mechanism feature vector matrix, and enhanced mechanism feature matrix. The final process fluctuation residual prediction value is added to the overall trend estimation term to obtain the final aperture error prediction value. This allows us to obtain the predicted true aperture value for the combination of process parameters. , In the formula, This refers to the nominal aperture design in the preset combination of process parameters; Furthermore, to verify the prediction accuracy of the forward joint prediction model for FDM printing hole size, this embodiment compares the prediction model using only the basic CART branch with the forward joint prediction model. The comparison results are as follows: the determination coefficient of the forward joint prediction model increased from 0.803 to 0.874 compared to the basic CART branch, an increase of 8.8%, indicating that the forward joint prediction model effectively overcomes the problems of traditional decision tree or single tree models being sensitive to sample changes and having uneven segmentation; the root mean square error and mean absolute error of the forward joint prediction model were significantly reduced by 20.03% and 19.21% respectively compared to the basic CART branch, indicating that the generalization risk of the forward joint prediction model under large error conditions is greatly reduced, improving the cross-size prediction capability and helping to ensure the stability of prediction results across batches under complex FDM printing processes.

[0035] S6: Set the target aperture of the FDM printed part. For each preset process parameter combination, use the reverse two-stage grid search method, call the forward joint prediction model to perform reverse solution, construct a multi-dimensional collaborative comprehensive scoring function, and calculate and sort the optimal process parameter combination based on the aperture error prediction value.

[0036] Furthermore, such as Figure 3 As shown, the specific steps of S6 include a progressively unfolding initialization phase, a coarse search phase, a fine search phase, and a scoring and ranking phase. Specifically: Initialization Phase: The user's desired target aperture is set, and a multi-dimensional discrete search space is constructed. This space contains multiple preset printing parameter combinations formed by orthogonal combinations of different layer heights, printing temperatures, printing speeds, and fill rates. Each preset printing parameter combination, along with a discrete design nominal aperture combination, forms a preset process parameter combination. In this embodiment, the preset layer heights are 0.12mm, 0.20mm, and 0.28mm; the preset printing temperatures are 190°C, 205°C, and 220°C; the preset printing speeds are 30mm / s, 50mm / s, and 70mm / s; and the preset fill rates are 20%, 50%, and 80%. There are 81 corresponding preset printing parameter combinations, with the design nominal aperture ranging from 1.5mm to 20.8mm. The multi-dimensional discrete search space confines the process parameter combinations within a discrete experimentally feasible domain, avoiding invalid process parameter combinations that cannot be printed accurately. Coarse search stage: In the multidimensional discrete search space, with a coarse grid step size of 0.02mm as the nominal aperture of the design, the forward joint prediction model is called to traverse and solve the actual aperture prediction value under each preset process parameter combination, and locate the coarse search grid point that makes the actual aperture prediction value closest to the target aperture. Fine search stage: In the multidimensional discrete search space, a local neighborhood is set within 0.035 mm before and after the coarse search grid point. Within the local neighborhood, the fine grid step size is 0.001 mm as the design nominal aperture. The forward joint prediction model is called again to perform a local secondary traversal. The predicted value of the actual aperture under each preset process parameter combination in the local neighborhood is calculated. The fine search grid point in the local neighborhood that makes the predicted value of the actual aperture closest to the target aperture is located. The preset process parameter combination at the fine search grid point is taken as the preferred process parameter combination. Scoring and Ranking Stage: Constructing a multi-dimensional collaborative comprehensive scoring function J.

[0037] In the formula, To determine the prediction reliability of a preset combination of process parameters, the prediction reliability is calculated as follows: Store the out-of-place absolute prediction residuals of the feature vectors of different process parameters during the training phase of the forward joint prediction model, constructing an out-of-place absolute prediction residual matrix; between the feature vectors of process parameters whose designed nominal aperture is close to the preset combination of process parameters, interpolation is used to calculate the corresponding out-of-place absolute prediction residual under the designed nominal aperture of the preset combination of process parameters, serving as the local out-of-place absolute prediction residual; the mean of the out-of-place absolute prediction residual matrix is ​​calculated to obtain the global average out-of-place absolute prediction residual; the weighted sum of the local out-of-place absolute prediction residual and the global average out-of-place absolute prediction residual yields the prediction reliability. R represents the numerical inverse solution residual, and Q represents the extrapolation penalty distance for exceeding the model training range. Operators can print calibration samples in advance when the extrapolation penalty distance is close to or exceeds the boundary, thus avoiding blindly generating extreme process conclusions that lack physical basis.

[0038] Calculate the multidimensional collaborative comprehensive scoring function for each preferred process parameter combination, sort them in ascending order, and output the preferred process parameter combination with the lowest score as the optimal process parameter combination, and output the remaining preferred process parameter combinations as alternative process parameter combinations.

[0039] Specifically, in one embodiment, the target aperture of a certain FDM printed part is set to 6.0 mm, and the nominal aperture at the located coarse search grid point is 6.12 mm. The final optimal process parameter combination is: nominal aperture D = 6.122 mm, layer height H = 0.20 mm, printing temperature T = 205℃, printing speed V = 30 mm / s, and fill rate F = 50%. Based on this optimal process parameter combination, the operator can modify the slicing settings to achieve one-time high-precision assembly. This optimal process parameter combination is solved inversely by the forward joint prediction model to obtain the corresponding overall trend estimate of -0.188 mm, the final process fluctuation residual prediction value of +0.066 mm, and the final aperture error prediction value. Therefore, the compensation amount of the nominal aperture to the target aperture is 0.122 mm; In another embodiment, the target aperture of a certain FDM printed part is set to 2.0 mm. In the optimal combination of process parameters, the compensation amount of the nominal aperture to the target aperture is 0.263 mm. It can be seen that the compensation amount of the present invention is greater than that of the larger aperture. This indicates that the reverse two-stage mesh search method adopted by the present invention can perform high-resolution search optimization to accurately reveal and compensate for the size effect that is more sensitive to melt deformation of small apertures, avoid the problem of large aperture interference and small aperture cannot be assembled, and ensure that the printed aperture size can be effectively assembled.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the accuracy of FDM printing hole dimensions and optimizing process parameters, characterized in that, Includes the following steps: S1: Collect process parameters and aperture errors of different printed samples, combine process parameters to construct process parameter feature vectors, build FDM aperture size benchmark dataset, create supervision labels based on aperture error, and divide the FDM aperture size benchmark dataset into k training folds. S2: Within each training iteration, construct and train a forward weighted process mechanism model, input process parameter feature vectors, and based on the hole forming process mechanism in FDM printing, generate a 19-dimensional candidate mechanism feature vector matrix through nonlinear transformation and algebraic mapping, and generate an enhanced mechanism feature matrix after weighting. S3: Within each training iteration, construct and train a weighted hierarchical target decomposition model, decompose the supervision labels, calculate the overall trend term through internal cross-validation, and separate the process fluctuation residual labels. S4: Within each training iteration, based on the enhancement mechanism feature matrix and the process fluctuation residual label, four tree model expert branches are trained respectively. The optimal deployment weights of each tree model expert branch are solved. The output results of each tree model expert branch are weighted and combined using the optimal deployment weights, and the overall trend term is accumulated to output the aperture error prediction value. S5: Allocate the best deployment weights to the expert branches of each tree model, construct the final multi-expert prediction model, and combine the forward weighted process mechanism model, the weighted hierarchical target decomposition model, and the final multi-expert prediction model to construct the forward joint prediction model; S6: Set the target aperture of the FDM printed part. For each preset process parameter combination, use the reverse two-stage grid search method, call the forward joint prediction model to perform reverse solution, construct a multi-dimensional collaborative comprehensive scoring function, and calculate and sort the optimal process parameter combination based on the aperture error prediction value.

2. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 1, characterized in that, The specific method of S2 is as follows: A forward weighted process mechanism model was constructed, classifying the hole forming process mechanism in FDM printing into size effect mechanism, thermal-velocity coupling mechanism, and geometry-process interaction mechanism. Through nonlinear transformation and algebraic mapping, the process parameter features in the process parameter feature vector are extended using the size effect mechanism, the thermal-velocity coupling mechanism, and the geometry-process interaction mechanism, respectively, resulting in 4 size effect extended features, 4 thermal-velocity coupling features, and 6 geometry-process interaction features. The process parameter features include the design nominal aperture D, layer height H, printing temperature T, printing speed V, and fill rate F. Process parameter features, size effect extension features, thermal-velocity coupling features, and geometry-process interaction features are used as candidate mechanism features. Within each training iteration, 19 candidate mechanism features from different samples are combined to form a 19-dimensional candidate mechanism feature vector matrix. After scaling transformation, the baseline mechanism feature matrix is ​​obtained. The ridge regression method is used to learn the normalized continuous weights of the candidate mechanism feature vector matrix. The forward weighted process mechanism model is trained within the training process and the enhanced mechanism feature matrix is ​​output.

3. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 2, characterized in that, The specific method of S3 is as follows: Within each training iteration, a weighted hierarchical target decomposition model is constructed, and the mean value of the supervised label under each designed nominal aperture D is calculated in groups to obtain discrete mean curves. A cubic polynomial is used to fit the discrete mean curve to the overall trend curve. ; Define the overall trend item , ； In the formula, λ is the fusion weight. Within the training layer, a weighted hierarchical target decomposition model is trained through cross-validation adaptive learning to obtain fusion weights that make the overall trend estimate approximate the aperture error. Separate calculation of process fluctuation residual label r, y is the supervisory label.

4. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 3, characterized in that, The specific method for constructing the final multi-expert prediction model using S4 is as follows: Four tree model expert branches are built in parallel, including the basic CART branch, the random forest Bagging branch, the GBR-Boostin branch, and the GBR residual branch. The process parameter feature vector is used as the input to the basic CART branch and the random forest Bagging branch, the candidate mechanism feature vector matrix is ​​used as the input to the GBR-Boostin branch, the enhancement mechanism feature matrix is ​​used as the input to the GBR residual branch, and the process fluctuation residual label is used as the supervision constraint. The four tree model expert branches are trained in each training layer. Outside of the training process, each tree model expert branch is run to obtain its corresponding predicted process fluctuation residual. The optimal deployment weights of each tree model expert branch are then determined by non-negative least squares algorithm constraints. The predicted process fluctuation residuals are weighted and combined using the optimal deployment weights to obtain the final predicted value of the process fluctuation residuals. The final predicted value of the process fluctuation residuals is then added to the overall trend estimation term to obtain the final predicted value of the aperture error.

5. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 4, characterized in that, The reverse two-stage grid search method in S6 is specifically as follows: Set the target aperture desired by the user and construct a multi-dimensional discrete search space. The multi-dimensional discrete search space is set with multiple preset printing parameter combinations formed by orthogonal combinations of different layer heights, different printing temperatures, different printing speeds, and different fill rates. Each preset printing parameter combination and the discrete design nominal aperture combination form a preset process parameter combination. The design nominal aperture range is 1.5mm to 20.8mm. Within the multidimensional discrete search space, a coarse grid step size of 0.02 mm is first used as the nominal aperture of the design. The forward joint prediction model is then called to calculate the predicted values ​​of the actual aperture under each preset process parameter combination in batches, and the coarse search grid points that make the predicted values ​​of the actual aperture closest to the target aperture are located. Within a local neighborhood of 0.035 mm before and after the coarse search grid point, with a fine grid step size of 0.001 mm as the nominal aperture, the forward joint prediction model is called again to locate the fine search grid point in the local neighborhood that makes the actual aperture prediction value close to the target aperture. The preset process parameter combination at the fine search grid point is used as the preferred process parameter combination.

6. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 5, characterized in that, The specific method for calling the forward joint prediction model in S6 is as follows: Input the preset process parameter combination into the forward joint prediction model; The forward-weighted process mechanism model generates its corresponding enhancement mechanism feature matrix; The weighted hierarchical target decomposition model generates its corresponding overall trend term; Finally, the multi-expert prediction model generates the corresponding final aperture error prediction value based on the enhancement mechanism feature matrix and the overall trend term. The final aperture error prediction value is added to the design nominal aperture in the preset process parameter combination to obtain the actual aperture prediction value.

7. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 6, characterized in that, The specific method for obtaining the optimal combination of process parameters in S6 is as follows: Construct a multidimensional collaborative comprehensive scoring function J. ； In the formula, To assess the predictive reliability of preset process parameter combinations, R is the final predicted aperture error, R is the numerical inverse solution residual, and Q is the extrapolation penalty distance for exceeding the model training range. Calculate the multidimensional synergistic comprehensive scoring function for each preferred combination of process parameters, sort them in ascending order, and take the preferred combination of process parameters with the lowest score as the optimal combination of process parameters.

8. The method for predicting FDM printing hole size accuracy and optimizing process parameters according to claim 7, characterized in that, The method for calculating the prediction confidence level is as follows: Store the out-of-place prediction absolute residuals of the feature vectors of different process parameters during the training phase of the forward joint prediction model, and construct the out-of-place prediction absolute residual matrix; Between the process parameter feature vectors where the designed nominal aperture is close to the preset process parameter combination, the out-of-place prediction absolute residual corresponding to the designed nominal aperture of the preset process parameter combination is obtained by interpolation and used as the local out-of-place prediction absolute residual. The global average absolute residual of out-of-place prediction is obtained by averaging the out-of-place prediction absolute residual matrix. The prediction reliability is obtained by weighted summing of the absolute residuals of the local out-of-range prediction and the absolute residuals of the global average out-of-range prediction.

9. The method for predicting the accuracy of FDM printing hole size and optimizing process parameters according to claim 2, characterized in that, The specific methods for extending the mechanism using the size effect mechanism, the thermal-velocity coupling mechanism, and the geometry-process interaction mechanism are as follows: Based on the size effect mechanism, the size effect extension characteristics are derived, including: aperture squared. reciprocal of aperture Reciprocal of the square of the aperture Logarithm of aperture ln(D); Based on the heat-velocity coupling mechanism, the heat-velocity coupling characteristics are extended to include: temperature-velocity ratio. The product of heat rate T×V, and the square of temperature velocity square ; Based on the geometry-process interaction mechanism, the extended geometry-process interaction features include: velocity-layer-height product V×H, and layer-height-aperture ratio. velocity aperture ratio , temperature-to-heat ratio The product of the filling layer height F×H and the product of the filling speed F×V.

Images