Method and system for predicting dimensional error of face projection microstereography three-dimensional printing molding

Abstract

The application provides a face projection microstereography three-dimensional printing forming size error prediction method and system, the method comprises the following steps: inputting the parameter values of each target process parameter and the design size of the printed object into a target stacking generalization integrated model to obtain a target forming size error; the target process parameter is a process parameter that has a significant influence on the forming size error; the first layer of the target stacking generalization integrated model comprises at least two target prediction models, and the second layer is a meta-learner; the target prediction model is used to determine an initial forming size error based on the parameter values of each target process parameter and the design size; and the meta-learner is used to determine the target forming size error based on the output weights of each target prediction model and each initial forming size error. The application can avoid inductive bias, take into account global trend fitting and local detail capture, combine the secondary error correction of the meta-learner, and improve the prediction error, the generalization ability and the robustness of the target stacking generalization integrated model.

Description

Technical Field

[0001] This invention relates to the fields of micro-nano additive manufacturing and intelligent manufacturing technology, and in particular to a method and system for predicting the dimensional error of surface projection micro-stereolithography 3D printing. Background Technology

[0002] Surface projection micro-stereolithography technology uses digital light processing to project and solidify photosensitive resin layer by layer, enabling the fabrication of complex structures at the micrometer level. However, the forming accuracy of surface projection micro-stereolithography is highly susceptible to the combined effects of various factors, leading to deviations between the actual printed dimensions and the design dimensions.

[0003] Common influencing factors include: uneven exposure energy distribution caused by optical edge diffraction and scattering; deformation caused by polymerization shrinkage and internal stress of photosensitive resin; and the influence of resin leveling time and coating uniformity on interlayer stacking errors. These factors typically exhibit nonlinear, time-varying, and strongly coupled characteristics, making it difficult for traditional analytical modeling or finite element methods to simultaneously achieve both modeling efficiency and prediction accuracy under engineering conditions.

[0004] Existing control technologies for the accuracy of surface projection micro-stereolithography printing mainly include physical / finite element modeling methods, trial and error methods, and methods based on a single machine learning model.

[0005] Physical / finite element modeling methods attempt to simulate the photocuring process through analytical formulas or simulations, but they are difficult to fully cover all nonlinear coupling effects such as edge diffraction, resin shrinkage, and leveling, resulting in limited prediction accuracy and huge computational demands.

[0006] Trial and error methods rely heavily on engineers' experience to conduct repeated experiments and adjustments, which is not only inefficient and wasteful of materials, but also makes it difficult to find the global optimal solution in a multidimensional parameter space.

[0007] Existing single machine learning models (such as those using only linear regression or a single decision tree) often suffer from inductive bias, making it difficult to simultaneously capture both global trend fitting and local details, resulting in insufficient generalization ability and robustness of the model.

[0008] It is evident that existing technologies are inadequate in terms of modeling accuracy, efficiency, and generalization ability, and there is an urgent need for an error prediction method with high prediction accuracy and stability. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a method and system for predicting the dimensional error of surface projection micro-stereolithography 3D printing.

[0010] This invention provides a method for predicting dimensional errors in surface projection micro-stereolithography 3D printing, comprising: The parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design dimensions of the printed object are input into the trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

[0011] According to the method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided by the present invention, the training process of the target stacking generalized ensemble model includes: Based on a pre-built error sample set, at least two candidate prediction models are selected from the regression model library. The error sample set includes multiple error samples, which contain the standard size, actual molding size error, and process parameter combination corresponding to the standard test sample. The at least two candidate prediction models are trained based on the error sample set to obtain at least two trained target prediction models; Based on the at least two trained target prediction models, an initial stacked generalization ensemble model is constructed using the two-layer stacked generalization architecture. Based on the error sample set, the initial stacked generalized ensemble model is trained to adjust the output weights of each target prediction model in the initial stacked generalized ensemble model, thereby obtaining the target stacked generalized ensemble model.

[0012] According to the present invention, a method for predicting the dimensional error of surface projection micro-stereolithography 3D printing is provided, wherein the method involves training an initial stacked generalized ensemble model based on the error sample set to adjust the output weights of each target prediction model in the initial stacked generalized ensemble model, thereby obtaining the target stacked generalized ensemble model, comprising: For each error sample, the standard dimensions and process parameters in the error sample are combined and input into each of the target prediction models in the initial stacking generalization ensemble model to obtain the first predicted forming size error output by each target prediction model; Each of the first predicted molding size errors is input into the meta-learner in the initial stacked generalization ensemble model. The meta-learner performs a weighted summation of each of the first predicted molding size errors based on the output weights of each of the target prediction models to obtain the second predicted molding size error. Based on the second predicted molding size error and the actual molding size error in the error sample, the loss value is determined; The output weights of each target prediction model in the meta-learner are adjusted based on the loss value. Continue training the initial stacked generalized ensemble model until the training stopping condition is met, and obtain the target stacked generalized ensemble model.

[0013] According to the method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided by the present invention, before selecting at least two candidate prediction models from a regression model library based on a pre-constructed error sample set, the method further includes: Based on each of the target process parameters, a sampling algorithm is used to generate different combinations of process parameters, and the combination of process parameters includes the parameter values ​​corresponding to each of the target process parameters. Based on the combination of process parameters, different standard test samples are printed using the surface projection micro stereolithography 3D printing equipment to obtain the actual molding size of each standard test sample. For each of the standard test specimens, the error of the actual molding size corresponding to the standard test specimen is determined based on the standard size corresponding to the standard test specimen and the actual molding size. The error sample set is constructed based on the standard dimensions, actual molding size errors, and process parameter combinations corresponding to each standard test sample.

[0014] According to the method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided by the present invention, before generating different combinations of process parameters based on each of the target process parameters using a sampling algorithm, the method further includes: A significance analysis was performed on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to select at least one target process parameter.

[0015] According to the present invention, a method for predicting the molding size error of surface projection micro-stereolithography 3D printing is provided, wherein different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment based on various combinations of process parameters to obtain the actual molding size corresponding to each standard test sample, including: Based on the combination of process parameters, different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment, and images corresponding to each standard test sample are obtained. Image processing algorithms are used to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size of each of the standard test samples. Data cleaning is performed on the initial molding dimensions corresponding to each of the standard test specimens to obtain the actual molding dimensions corresponding to each of the standard test specimens.

[0016] According to the present invention, a method for predicting the molding size error of surface projection micro-stereolithography 3D printing is provided, wherein the error sample set is constructed based on the standard size corresponding to each standard test sample, the actual molding size error, and the combination of process parameters, including: The errors of the standard dimensions and actual molding dimensions corresponding to each of the standard test specimens are respectively standardized; Based on the standard dimensions and actual molding size errors corresponding to each of the standardized standard test specimens, and the process parameter combinations corresponding to each of the standard test specimens, an error sample set is constructed, which includes a training set and a test set.

[0017] According to the present invention, a method for predicting the dimensional error of surface projection micro-stereolithography 3D printing is provided, wherein the step of performing significance analysis on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment and screening out at least one target process parameter includes: The Plackett-Burman experimental design was used to perform significance analysis and two-level screening on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to obtain at least one target process parameter. Each of the candidate process parameters includes exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time; Each of the target process parameters includes at least one of the exposure time, the exposure power density, the printing layer thickness, and the leveling time.

[0018] According to the present invention, a method for predicting the dimensional error of surface projection micro-stereolithography 3D printing is provided, wherein the sampling algorithm is the Latin hypercube sampling algorithm; the standard test sample includes micropillar arrays with different design diameter gradients, and each group of micropillar arrays is repeatedly arranged in a preset number.

[0019] According to the present invention, a method for predicting the molding size error of surface projection micro-stereolithography 3D printing is provided. The method involves using image processing algorithms to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each standard test sample to obtain the initial molding size corresponding to each standard test sample. This includes: For each image, the image is converted to the HSV color space, and a scale outline is extracted based on the color threshold range. The pixel equivalent is then determined based on the scale outline. The image is converted to grayscale and Gaussian blurred for noise reduction. An edge detection operator is used to extract the edges. The center and radius of the micro-pillar are determined by voting in the accumulator space using the Hough gradient circular transform. The physical diameter is calculated by combining the pixel equivalent, and the initial forming size corresponding to the standard test sample in the image is obtained.

[0020] According to the present invention, a method for predicting the dimensional error of surface projection micro-stereolithography 3D printing is provided, wherein the at least two trained target prediction models include at least two of random forest, gradient boosting tree, optimized distributed gradient boosting library, Gaussian process regression, multilayer perceptron and ridge regression.

[0021] This invention also provides a surface projection micro-stereolithography 3D printing molding size error prediction system, comprising: The molding size error prediction module is configured to input the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design size of the printed object into the trained target stacking generalization ensemble model to predict the molding size error and obtain the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

[0022] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the surface projection micro-stereolithography three-dimensional printing molding size error prediction method as described above.

[0023] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the surface projection micro-stereolithography three-dimensional printing molding size error prediction method as described above.

[0024] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the surface projection micro-stereolithography three-dimensional printing molding size error prediction method as described above.

[0025] The present invention provides a method and system for predicting the molding size error in surface projection micro-stereolithography 3D printing. This method and system inputs the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design dimensions of the printed object into a trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error. The target stacking generalization ensemble model adopts a two-layer stacking generalization architecture. The first layer of the two-layer stacking generalization architecture is a base learner, and the second layer is a meta-learner. The base learner contains at least two trained target prediction models. The target prediction models are used to determine the initial molding size error based on the parameter values ​​of each target process parameter and the design dimensions. The meta-learner is used to determine the target molding size error based on the output weights of each target prediction model and the initial molding size error output by each target prediction model. This invention avoids inductive bias by using at least two different target prediction models, while simultaneously considering both global trend fitting and local detail capture. It combines the outputs of each target prediction model with a meta-learner for secondary error correction, significantly improving prediction accuracy and enhancing the generalization ability and robustness of the target stacking generalization ensemble model. Furthermore, it employs process parameters that significantly impact molding size errors, reducing data processing volume while maintaining prediction accuracy. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is one of the flowcharts illustrating the method for predicting dimensional errors in surface projection micro-stereolithography 3D printing provided by the present invention.

[0028] Figure 2 This is a schematic diagram of the stacked generalization ensemble model provided by the present invention.

[0029] Figure 3 This is the second flowchart of the method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided by the present invention.

[0030] Figure 4 This is a schematic diagram of the structure of the surface projection micro-stereolithography 3D printing molding size error prediction system provided by the present invention.

[0031] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] First, the relevant content of this invention will be explained.

[0034] The existing surface projection micro-stereolithography 3D printing technology has the following problems in terms of forming accuracy control: Nonlinear errors caused by multi-physics coupling are difficult to predict: the precision of surface projection micro-stereolithography is affected by multiple factors such as optical edge diffraction, resin polymerization shrinkage, and leveling time. These factors have nonlinear and time-varying characteristics, which are difficult to fully cover by traditional physical / finite element modeling methods, and the computational load is huge, making it difficult to balance efficiency and accuracy. Traditional trial-and-error methods are inefficient: repeated experiments and adjustments based on experience not only waste materials, but also make it difficult to find the global optimal solution in a multidimensional parameter space. Insufficient generalization ability of single models: Existing single machine learning models (such as linear regression or single decision tree) have inductive bias, making it difficult to simultaneously take into account global trend fitting and local detail capture, resulting in insufficient robustness of the model. To this end, the present invention provides a method and system for predicting the dimensional error of surface projection micro-stereolithography 3D printing. It can accurately establish the mapping relationship between key process parameters and dimensional error without significantly increasing experimental complexity, and achieve high-precision and robust dimensional error prediction. It can also be used for process parameter optimization / recommendation.

[0035] The following is combined Figures 1 to 5 This invention describes a method and system for predicting dimensional errors in surface projection micro-stereolithography 3D printing.

[0036] Figure 1 This is one of the flowcharts illustrating the method for predicting dimensional errors in surface projection micro-stereolithography 3D printing provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Input the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design size of the printed object into the trained target stacking generalization ensemble model to predict the molding size error and obtain the target molding size corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

[0037] Specifically, the target process parameters are process parameters that have a significant impact on the molding dimensional error. Each of the target process parameters includes at least one of the exposure time, the exposure power density, the printing layer thickness, and the leveling time. The parameter values ​​of the target process parameters refer to the actual values ​​of the target process parameters, such as an exposure time of 0.5 to 3 seconds (s) and a power density of 40 to 120 milliwatts per square centimeter (mW·cm²). -2 The printing layer thickness is 5 / 10 / 20 micrometers (μm), and the leveling time is 3 to 150 seconds.

[0038] Specifically, see Figure 2 , Figure 2 This is a schematic diagram of the stacked generalization ensemble model provided by the present invention: The target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner, wherein the meta-learner is a linear regression model; the base learner contains at least two trained target prediction models. The at least two trained target prediction models include at least two of Random Forest, Gradient Boosting, Optimized Distributed Gradient Boosting Library (XGBoost), Gaussian Process Regression, Multilayer Perceptron (MLP), and Ridge Regression.

[0039] Specifically, the target prediction model is used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design dimensions; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

[0040] Specifically, the object to be printed is the object that needs to be printed using surface projection micro-stereolithography 3D printing. The design size refers to the target size of the object to be printed.

[0041] In practical applications, see Figure 2 The parameter values ​​of each target process parameter and the design dimensions (printing data) of the printed object in the surface projection micro-stereolithography 3D printing equipment can be input into the trained target stacking generalization ensemble model. Each target prediction model in the target stacking generalization ensemble model receives the parameter values ​​of each target process parameter and the design dimensions.

[0042] For each target prediction model, the molding size error is predicted based on the parameter values ​​of each target process parameter and the design dimensions, resulting in the initial molding size error corresponding to the printed object. The initial molding size error output by each target prediction model (as a new feature matrix) is input into the meta-learner. The meta-learner, based on the output weights of each target prediction model, performs a weighted summation of the initial molding size errors to obtain the target molding size error. The formula for determining the target molding size error is as follows: E = a1·e1 + ... + a n ·e n Where E is the target forming size error, e1 to e n These represent the target prediction models from the 1st to the nth, a1 to a... n These are the output weights of the first target prediction model to the nth target prediction model, where n is the number of target prediction models included in the target stacking generalization ensemble model.

[0043] It should be noted that the target stacking generalized ensemble model can also be used for process parameter optimization / recommendation. That is, the design dimensions of the printed object are input into the target stacking generalized ensemble model, and the target stacking generalized ensemble model outputs recommended parameter values ​​for each target process parameter in the surface projection micro-stereolithography 3D printing equipment.

[0044] This invention avoids inductive bias by using at least two different target prediction models, while simultaneously considering both global trend fitting and local detail capture. It combines the outputs of each target prediction model with a meta-learner for secondary error correction, significantly improving prediction accuracy and enhancing the generalization ability and robustness of the target stacking generalization ensemble model. Furthermore, it employs process parameters that significantly impact molding size errors, reducing data processing volume while maintaining prediction accuracy.

[0045] Optionally, the training process of the target stacked generalization ensemble model includes: Based on a pre-built error sample set, at least two candidate prediction models are selected from the regression model library. The error sample set includes multiple error samples, which contain the standard size, actual molding size error, and process parameter combination corresponding to the standard test sample. The at least two candidate prediction models are trained based on the error sample set to obtain at least two trained target prediction models; Based on the at least two trained target prediction models, an initial stacked generalization ensemble model is constructed using the two-layer stacked generalization architecture. Based on the error sample set, the initial stacked generalized ensemble model is trained to adjust the output weights of each target prediction model in the initial stacked generalized ensemble model, thereby obtaining the target stacked generalized ensemble model.

[0046] Specifically, candidate prediction models include, but are not limited to, random forest, gradient boosting tree, XGBoost, Gaussian process regression, multilayer perceptron, ridge regression, etc.

[0047] Specifically, the standard test sample comprises micropillar arrays with different designed diameter gradients, and each group of micropillar arrays is repeatedly arranged in a preset number. For example, the standard test sample can be a standard test sample containing micropillar arrays with different diameter gradients (e.g., 5μm, 10μm, 20μm, 40μm, 60μm, 80μm), with each group of micropillars repeatedly arranged in a preset number (e.g., 9) to obtain sufficient statistical data. The standard test sample comprises micropillar arrays with different designed diameter gradients, and is repeatedly arranged in a preset number, thus enabling the statistical and stable acquisition of molding error data under various working conditions.

[0048] In practical applications, at least two candidate prediction models are selected from the regression model library based on the error sample set: the error sample set is input into each prediction model in the regression model library to perform error prediction, and the prediction results of each prediction model are obtained. Based on the prediction results, the model performance of each prediction model is determined, such as at least one of accuracy and response rate, and at least two prediction models whose model performance meets the performance conditions are selected as candidate prediction models.

[0049] Furthermore, each candidate prediction model is trained based on the error sample set to obtain the target prediction model corresponding to each candidate model. This involves hyperparameter optimization based on the candidate error sample set to determine the base model set. Candidate models include, but are not limited to, random forest, gradient boosting tree, XGBoost, Gaussian process regression, multilayer perceptron, and ridge regression. Hyperparameter optimization can employ methods such as Bayesian optimization, using cross-validation error as the objective function to search for the optimal parameter combination.

[0050] The training process of the candidate prediction model includes: for each error sample, combining the standard dimensions and process parameters in the error sample and inputting them into the candidate prediction model to obtain the predicted molding size error output by the candidate prediction model; calculating the error loss based on the predicted molding size error and the actual molding size error in the error sample; adjusting the model parameters of the candidate prediction model based on the error loss; and continuing to train the adjusted candidate prediction model until the training termination condition is met to obtain the target prediction model corresponding to the candidate prediction model.

[0051] It should be noted that the error sample set can be divided into a training set and a test set, i.e., the error sample set includes a training set and a test set. Therefore, for each candidate prediction model, the candidate prediction model is trained based on the training set to obtain a trained candidate prediction model; the trained candidate prediction model is then tested based on the test set; if the test passes, the trained candidate prediction model is used as the target prediction model corresponding to the candidate prediction model; if the test fails, the step of training the candidate prediction model based on the training set and subsequent steps are continued.

[0052] Subsequently, a stacking two-layer generalization architecture was adopted to construct an initial stacked generalization ensemble model based on each target prediction model, so as to improve prediction performance by leveraging the complementary inductive biases of heterogeneous models.

[0053] The first layer (base learner) selects at least two differentiated target prediction models, including at least two of the following: random forest, gradient boosting tree, XGBoost, Gaussian process regression, multilayer perceptron, and ridge regression. These target prediction models cover tree models, probabilistic models, neural networks, and linear models, ensuring the diversity of the ensemble.

[0054] The second layer (meta-learner) employs a linear regression model. The meta-learner learns the output weights of the base learner to dynamically correct the error.

[0055] Training strategy: Using the error sample set as input to each target prediction model, a 5-fold cross-validation out-of-fold prediction mechanism is employed to generate a meta-feature matrix (the output of each target prediction model), which serves as the input to the meta-learner. The meta-learner is trained only on the out-of-fold meta-features, effectively preventing data leakage and overfitting.

[0056] It should be noted that the error sample set can be divided into a training set and a test set. Therefore, the initial stacked generalized ensemble model is trained based on the training set to obtain a trained initial stacked generalized ensemble model; the trained initial stacked generalized ensemble model is then tested based on the test set; if the test passes, the trained initial stacked generalized ensemble model is used as the target stacked generalized ensemble model; if the test fails, the step of training the initial stacked generalized ensemble model based on the training set and subsequent steps are continued. This further ensures the robustness of the target stacked generalized ensemble model.

[0057] Optionally, training the initial stacked generalization ensemble model based on the error sample set to adjust the output weights of each target prediction model in the initial stacked generalization ensemble model to obtain the target stacked generalization ensemble model includes: For each error sample, the standard dimensions and process parameters in the error sample are combined and input into each of the target prediction models in the initial stacking generalization ensemble model to obtain the first predicted forming size error output by each target prediction model; Each of the first predicted molding size errors is input into the meta-learner in the initial stacked generalization ensemble model. The meta-learner performs a weighted summation of each of the first predicted molding size errors based on the output weights of each of the target prediction models to obtain the second predicted molding size error. Based on the second predicted molding size error and the actual molding size error in the error sample, the loss value is determined; The output weights of each target prediction model in the meta-learner are adjusted based on the loss value. Continue training the initial stacked generalized ensemble model until the training stopping condition is met, and obtain the target stacked generalized ensemble model.

[0058] Specifically, the training stopping condition can be at least one of the following: the loss value is less than the loss threshold, the rate of change of the loss value is less than the rate of change threshold, or the number of iterations reaches the iteration threshold.

[0059] In practical applications, for each error sample, the standard dimensions and process parameters in the error sample are combined and input into the initial stacked generalization ensemble model. Each target prediction model in the initial stacked generalization ensemble model receives the standard dimensions and process parameters from the error sample.

[0060] Furthermore, for each target prediction model, the target prediction model performs dimensional error prediction based on the standard dimensions and process parameter combinations in the error sample, and obtains the first predicted molding dimensional error corresponding to the error sample output by the target prediction model.

[0061] Then, the first predicted molding size error corresponding to the error sample output by each target prediction model is input into the meta-learner in the initial stacked generalization ensemble model. The meta-learner performs a weighted summation of each first predicted molding size error based on the output weights of each target prediction model to obtain the second predicted molding size error.

[0062] Next, the second predicted molding size error and the actual molding size error in the error sample are calculated according to the set loss function to obtain the loss value, and the output weights of each target prediction model in the meta-learner are adjusted based on the loss value; the initial stacked generalization ensemble model is trained until the training stopping condition is reached, thereby obtaining the trained target stacked generalization ensemble model.

[0063] In this embodiment of the invention, Stacking heterogeneous integration utilizes the complementary inductive biases of different target prediction models and achieves two-layer error correction by learning dynamic weights through a meta-learner, which significantly improves prediction accuracy and generalization stability.

[0064] Optionally, before selecting at least two candidate prediction models from the regression model library based on a pre-built error sample set, the method further includes: Based on each of the target process parameters, a sampling algorithm is used to generate different combinations of process parameters, and the combination of process parameters includes the parameter values ​​corresponding to each of the target process parameters. Based on the combination of process parameters, different standard test samples are printed using the surface projection micro stereolithography 3D printing equipment to obtain the actual molding size of each standard test sample. For each of the standard test specimens, the error of the actual molding size corresponding to the standard test specimen is determined based on the standard size corresponding to the standard test specimen and the actual molding size. The error sample set is constructed based on the standard dimensions, actual molding size errors, and process parameter combinations corresponding to each standard test sample.

[0065] In practical applications, a sampling algorithm can be used to generate a combination of process parameters based on each target process parameter. The control surface projection micro-stereolithography 3D printing equipment can then print standard test samples to obtain the actual molding size corresponding to the standard test samples.

[0066] The sampling algorithm is Latin hypercube sampling (LHS) to ensure sufficient coverage and uniform distribution of the parameter space.

[0067] For example, a coverage process window is generated (e.g., exposure time 0.5–3 s, power density 40–120 mW·cm). -2 The process parameter combinations (printing layer thickness 5 / 10 / 20μm, leveling time 3~150 s) can be selected, such as 60 sets of process parameter combinations.

[0068] After obtaining the actual molded dimensions, for each standard test specimen, the deviation between the standard dimension and the actual molded dimension is taken as the actual molded dimension error for that standard test specimen. Further, the standard dimension, actual molded dimension error, and process parameters of the standard test specimen are combined to form an error sample. The error samples corresponding to each standard test specimen constitute the error sample set. This ensures the reliability and accuracy of the error sample set, which is beneficial for improving the prediction accuracy of the stacked generalization ensemble model.

[0069] In this embodiment of the invention, the LHS sampling algorithm can perform stratified sampling, avoiding sample concentration or gaps, and improving the representativeness of the samples and the generalization ability of the model.

[0070] Optionally, the step of printing different standard test samples using the surface projection micro-stereolithography 3D printing equipment based on each of the aforementioned process parameter combinations to obtain the actual molded size corresponding to each of the aforementioned standard test samples includes: Based on the combination of process parameters, different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment, and images corresponding to each standard test sample are obtained. Image processing algorithms are used to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size of each of the standard test samples. Data cleaning is performed on the initial molding dimensions corresponding to each of the standard test specimens to obtain the actual molding dimensions corresponding to each of the standard test specimens.

[0071] Specifically, the image can be a microscopic image.

[0072] In practical applications, a combination of process parameters can be generated based on each target process parameter using a sampling algorithm. The control plane projection micro-stereolithography 3D printing equipment can then print standard test samples and obtain the corresponding microscopic images of the standard test samples.

[0073] Then, automatic image measurement and data preprocessing are performed: image processing algorithms are used to complete scale recognition, micro-pillar extraction, and size measurement to obtain the initial molding dimensions corresponding to the standard test samples; then, the initial molding dimensions corresponding to the standard test samples are cleaned to remove printing failure samples, and outliers are removed using the Standard Score (Z-Score) Laplace's Rule (3σ) criterion to obtain the actual molding dimensions corresponding to each standard test sample. This ensures the accuracy and reliability of the data.

[0074] Optionally, the step of using image processing algorithms to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size corresponding to each of the standard test samples includes: For each image, the image is converted to the HSV color space, and a scale outline is extracted based on the color threshold range. The pixel equivalent is then determined based on the scale outline. The image is converted to grayscale and Gaussian blurred for noise reduction. An edge detection operator is used to extract the edges. The center and radius of the micro-pillar are determined by voting in the accumulator space using the Hough gradient circular transform. The physical diameter is calculated by combining the pixel equivalent, and the initial forming size corresponding to the standard test sample in the image is obtained.

[0075] Specifically, HSV (Hue, Saturation, Value) is a color space created based on the intuitive characteristics of color, also known as the Hexcone Model. The HSV color model refers to a subset of visible light in the H, S, V three-dimensional color space, which contains all colors in a certain color gamut.

[0076] In practical applications, the process of ruler recognition is as follows: convert the microscopic image to the HSV color space, extract the ruler outline (the number of pixels occupied by the ruler in the microscopic image) according to the color threshold range, and calculate the pixel equivalent (the size of each pixel).

[0077] Micropillar identification: After converting the microscopic image to grayscale and denoising it with Gaussian blur, the edge detection (Canny) operator is used to extract the edge, and the center and radius (diameter) of the micropillar are determined by Hough gradient circle transform. Then, the physical diameter is calculated by combining pixel equivalents, so as to obtain the initial molding size corresponding to each standard test sample.

[0078] In this embodiment of the invention, the automated image calibration and recognition algorithm unifies the measurement rules, reduces human subjective error, and ensures the reliability and efficiency of the data by combining it with visual verification.

[0079] Optionally, before generating different combinations of process parameters using a sampling algorithm based on each of the target process parameters, the method further includes: A significance analysis was performed on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to select at least one target process parameter.

[0080] Specifically, each of the candidate process parameters includes exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time.

[0081] In practical applications, in order to focus on the main influencing factors, it is necessary to screen the target process parameters: a significance analysis can be performed on the candidate process parameters, and the target process parameters that have a significant impact on the molding size error can be screened from the candidate process parameters based on the significance analysis results.

[0082] Optionally, the step of performing significance analysis on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to screen out at least one of the target process parameters includes: The Plackett-Burman experimental design was used to perform significance analysis and two-level screening on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to obtain at least one target process parameter.

[0083] Specifically, the Plackett-Burman design is a screening experimental design method primarily used in multi-factor systems to quickly identify key factors that significantly influence the response variable. Its core objective is to efficiently screen key factors by minimizing the number of trials, thereby reducing resource consumption in subsequent optimization phases.

[0084] In practical applications, in order to reduce the modeling dimensionality and focus on the main influencing factors, the Plackett-Burman (PB) experimental design can be used to conduct significance analysis on candidate process parameters.

[0085] For example, candidate process parameters (candidate factors) are set to include exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time. Through a two-level partial factorial design experiment, target process parameters that have a significant impact on the molding size error (P<0.05) are screened out: exposure time, exposure power density, printing layer thickness, and leveling time.

[0086] In this embodiment of the invention, Plackett-Burman screening can first eliminate factors (candidate process parameters) with weaker influence, reducing the modeling dimensionality and experimental cost, while the core parameters retained can better characterize the error formation mechanism.

[0087] Optionally, constructing the error sample set based on the standard dimensions corresponding to each of the standard test samples, the actual molding size error, and the combination of process parameters includes: The errors of the standard dimensions and actual molding dimensions corresponding to each of the standard test specimens are respectively standardized; The error sample set is constructed based on the standard dimensions and actual molding size errors corresponding to each of the standardized standard test samples, as well as the process parameter combinations corresponding to each of the standard test samples.

[0088] In practical applications, to prevent data leakage, Z-Score standardization can be used to standardize the errors of each standard size and each actual molding size. Then, an error sample set can be constructed based on the standardized standard size, each actual molding size error, and the actual molding size error.

[0089] Z-Score standardization is a commonly used data preprocessing method that transforms data into a standard normal distribution with a mean of 0 and a standard deviation of 1. It can eliminate the dimensional differences between different features, allowing data to be compared on the same scale, thereby improving model performance and the fairness of analysis.

[0090] Preferably, Z-score normalization can be calculated based on the training set parameters to avoid data leakage.

[0091] The following is combined Figure 3 and Figure 4 The method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided by the present invention will be further explained.

[0092] See Figure 3 , Figure 3 This is the second flowchart of the method for predicting the molding size error of surface projection micro-stereolithography 3D printing provided by the present invention. The method includes six steps: target process parameter screening, dataset construction, automatic image measurement and data preprocessing, candidate model training and optimization, construction of stacking ensemble model and molding size error prediction.

[0093] Target process parameter screening: Candidate process parameters were set as exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time. The response variable was the micropillar diameter error after molding. A Plackett-Burman (PB) two-level partial factorial design was used, and a total of 15 experiments were performed. Significance analysis was conducted at the 95% confidence level. The results showed that exposure time, exposure power density, printing layer thickness, and leveling time were the core input features that significantly affected the molding error; while cleaning time and drying time had a relatively small impact within the current range and were therefore excluded. Therefore, the target process parameters include exposure time, exposure power density, printing layer thickness, and leveling time.

[0094] Dataset Construction: For the four selected target process parameters, Latin hypercube sampling was used to generate 60 sets of process parameter points, ensuring sufficient coverage and uniform distribution. The target process parameter values ​​ranged from 0.5 to 3 seconds for exposure time and 40 to 120 mW·cm³ for exposure power density. -2 The printing layer thickness was 5 / 10 / 20 μm, and the leveling time ranged from 3 to 150 s. The designed standard test specimen structure was a micropillar array with dimensions of 3.84 mm × 2.16 mm, containing six groups of micropillar arrays with designed diameters of 5, 10, 20, 40, 60, and 80 μm, an aspect ratio of 4:1, and 9 repeating units in each group. Four specimens were printed in a 2×2 layout per batch to evaluate repeatability. After printing, the specimens were statically cleaned with isopropanol for 30 minutes and dried in a fume hood for 30 minutes before microscopic imaging.

[0095] Automatic image measurement and data preprocessing includes automated image measurement and data cleaning and preprocessing.

[0096] Automated image measurement: In order to quickly and accurately obtain dimensional errors, an automatic measurement algorithm based on Python OpenCV can be used for scale recognition, pixel equivalent calculation, and micro-pillar recognition.

[0097] Scale recognition and pixel equivalent calculation: The microscopic image is converted from the Blue, Green, Red (BGR) space to the HSV space, and a mask is created using the hue range of the red scale. The width of the bounding rectangle of the scale outline is extracted. Pixel equivalent The formula for calculating (μm / pixel) is: K= / ,in This refers to the actual physical length of the scale (e.g., 200 μm), i.e., the scale outline.

[0098] Micropillar recognition: The microscopic image is converted to grayscale, Gaussian blurring with a 9x9 convolution kernel is used for noise reduction, and the Canny operator is used to extract edges. Then, the Hough gradient circular transform is used to find the center and radius in the accumulator space through voting. The radius search range and threshold are set according to the design dimensions, and the circle radius is output and converted into the physical diameter. Micropillar diameter measurement. ,in This represents the pixel size corresponding to the radius. Annotated images are output for quick review. For each diameter gradient, the average of the nine measurements after removing outliers is taken as the final response.

[0099] Data cleaning and preprocessing: The measurement data was processed by reorganizing the results of 60 process combinations × 6 diameters into a long table. Printing failure samples (78) with unmeasurable errors were removed. Six outlier samples were identified by the criteria, and 276 valid samples were retained. Z-score standardization was used to eliminate the influence of units of measurement; the formula is: =(x-μ) / σ, where, Here, x represents the standardized sample, and μ represents the sample mean. The standard deviation is given. Standardized parameters are calculated only from the training set. The dataset is divided into training and test sets at 70% and 30% respectively.

[0100] Building a Stacking integration model: such as Figure 2 As shown, a Stacking two-layer generalization architecture is constructed. The preferred base learners for the first layer include: Random Forest, Gradient Boosting Tree, XGBoost, Gaussian Process Regression, Multilayer Perceptron, and Ridge Regression. The second-layer meta-learner uses a linear regression model. The Stacking two-layer generalization architecture employs a hierarchical cascading strategy, achieving non-linear fusion of prediction results from multiple models by constructing base learners and meta-learners. This architecture leverages the differentiated inductive biases of heterogeneous base learners for complementary advantages, fully exploits data features through a hierarchical training mechanism, effectively overcomes the limitations of a single model, and thus significantly improves the overall model's generalization performance and prediction accuracy.

[0101] The first layer of base learners selected six heterogeneous models with different inductive biases to ensure diversity in the ensemble. Random Forest, Gradient Boosting Tree, and XGBoost are tree-based ensemble models, adept at handling nonlinear relationships and high-dimensional features. Gaussian Process Regression is a kernel-based probabilistic model that can provide estimates of prediction uncertainty and is suitable for small sample data. Multilayer Perceptron is a fully connected neural network with strong global fitting capabilities. Ridge Regression is a linear model with L2 regularization, used to capture linear trends and prevent overfitting.

[0102] In order to obtain optimal generalization performance during model building and training, this embodiment adopts a "hierarchical optimization + joint fine-tuning" strategy. First, independent Bayesian hyperparameter tuning is performed on each base learner to initially lock the optimal parameter space of each model. Then, joint tuning is performed on the stacking ensemble model to determine the best hyperparameter combination through global optimization, thereby significantly improving model performance by adapting to the characteristics of the error prediction problem. At the same time, existing machine learning computing libraries and their built-in efficient optimization algorithms are used to automatically complete the entire model training process, which greatly reduces the cost of manual debugging while ensuring training efficiency.

[0103] Hyperparameters include the number of decision trees, maximum depth, minimum number of samples for node splits, and maximum number of features for random forests; the learning rate, regularization term, and subsampling ratio for gradient boosting trees; and the noise level and optimizer restart count for Gaussian process regression.

[0104] Linear regression is used as the meta-learner. The reason for choosing a linear model is that after processing by the first strong learner layer, the output predictions are usually highly correlated with the true labels; if a complex nonlinear model is used as the meta-model at this point, it is very easy to overfit the meta-features. The linear meta-model essentially learns a set of "optimal dynamic weights," and its mathematical expression is as follows: .

[0105] in The metamodel assigns trust weights to each base model. For example, when dealing with samples with smooth features, the weight of the GPR may be automatically amplified; while when dealing with complex step-changes, the contribution of the tree model is more significant.

[0106] To prevent data leakage during model training, this embodiment employs a 5-fold cross-validation out-of-bag prediction mechanism. Specifically, the training set is randomly divided into 5 parts. In each training round, the base learner is trained using 4 parts, and predictions are made using the remaining part. The out-of-bag predictions obtained from the 5 rounds are concatenated to generate a meta-feature matrix. The second-layer meta-learner is trained solely based on this meta-feature matrix, ensuring that the features input to the meta-learner are not utilized during the training phase of the base learner, effectively improving the robustness of the final print error prediction model.

[0107] Then, model evaluation, verification, and interpretability analysis are performed: the constructed Stacking ensemble learning model is comprehensively evaluated, verified, and physically interpretable to demonstrate the superiority of this invention in the task of predicting errors in surface projection micro-stereolithography 3D printing.

[0108] To quantify the predictive performance of the model, three core evaluation metrics were selected: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (CDE). ).

[0109] The root mean square error (RMSE) measures the standard deviation between the predicted and actual values ​​and is sensitive to large error samples; the mean absolute error (MAE) measures the average level of prediction error and better reflects the actual deviation of the predicted values; the coefficient of determination (...) The value is used to characterize the model's ability to explain data variability; the closer it is to 1, the better the model fits.

[0110] Under the same test set conditions, the Stacking ensemble model outperforms the single base learner across all key evaluation metrics. Specific quantitative data shows that the model's coefficient of determination (COP) is significantly higher than that of the single base learner. The score is as high as 0.897, compared to the best-performing single model (XGBoost). The accuracy was improved by approximately 3.7% (=0.865); simultaneously, its root mean square error (RMSE) was significantly reduced to 1.958 μm, meaning that the deviation between the predicted value and the actual printing error was effectively controlled within the micrometer-level accuracy range. Further residual analysis showed that the model exhibited extremely strong robustness when dealing with regions with strong nonlinear coupling characteristics (such as the significant error region caused by edge diffraction effects in large-sized micropillars). This is mainly due to the meta-learner in the two-layer stacked architecture being able to dynamically learn and integrate the decision weights of different target prediction models, achieving complementary advantages between heterogeneous models. Based on the above high-precision prediction results, the Stacking ensemble model can accurately predict potential forming errors according to the user-input design dimensions and process parameters, and generate process compensation values ​​accordingly (such as correcting exposure time or scaling the design model), thereby significantly improving the one-time forming accuracy and yield of surface projection micro-stereolithography 3D printing without repeated trial and error.

[0111] Molding size error prediction: The parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design size of the printed object are input into the trained target stacking generalization ensemble model to predict the molding size error and obtain the target molding size error corresponding to the printed object.

[0112] The method for predicting the molding size error of surface projection micro-stereolithography 3D printing provided in this invention achieves accurate prediction of molding error through a complete process of target process parameter screening, dataset construction, automated measurement, and integrated model construction.

[0113] The method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided in this invention has high prediction accuracy and strong robustness. The Stacking ensemble model combines the advantages of multiple heterogeneous models and significantly reduces prediction error through secondary error correction by a meta-learner. Experiments show that the coefficient of determination (R²) of this model is high. 2 The mean square error (RMSE) can reach 0.897, and the root mean square error (RMSE) is reduced to 1.958 μm, which is about 3.7% better than the single optimal model (XGBoost).

[0114] The method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided in this invention has low experimental cost and high modeling efficiency: it adopts Plackett-Burman screening to eliminate minor factors, thereby avoiding full coverage of all process parameters. Combined with the LHS sampling strategy, it significantly reduces the number of experiments and material consumption while ensuring the coverage of the sample space.

[0115] The method for predicting the dimensional error of surface projection micro-stereolithography 3D printing provided in this invention provides objective and reliable data acquisition: the proposed automated image calibration and micro-pillar recognition algorithm unifies the measurement rules, eliminates the subjective error of human reading, and has a processing speed much faster than manual measurement, ensuring the feasibility of constructing a large dataset.

[0116] The surface projection micro-stereolithography 3D printing molding size error prediction method provided in this embodiment of the invention has strong generalization ability: the double-layer stacked architecture is trained through an out-of-bag prediction mechanism, which effectively avoids overfitting and makes it more stable when dealing with nonlinear coupling feature regions (such as errors caused by edge diffraction).

[0117] The following describes the surface projection micro-stereolithography 3D printing dimensional error prediction system provided by the present invention. The surface projection micro-stereolithography 3D printing dimensional error prediction system described below can be referred to in correspondence with the surface projection micro-stereolithography 3D printing dimensional error prediction method described above.

[0118] Figure 4 This is a schematic diagram of the structure of the surface projection micro-stereolithography 3D printing molding size error prediction system provided by the present invention, as shown below. Figure 4 As shown, the surface projection micro-stereolithography 3D printing molding dimensional error prediction system includes: The molding size error prediction module 401 is configured to input the parameter values ​​of each target process parameter in the surface projection micro stereolithography 3D printing equipment and the design size of the printed object into the trained target stacking generalization ensemble model to predict the molding size error and obtain the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

[0119] This invention avoids inductive bias by using at least two different target prediction models, while simultaneously considering both global trend fitting and local detail capture. It combines the outputs of each target prediction model with a meta-learner for secondary error correction, significantly improving prediction accuracy and enhancing the generalization ability and robustness of the target stacking generalization ensemble model. Furthermore, it employs process parameters that significantly impact molding size errors, reducing data processing volume while maintaining prediction accuracy.

[0120] Optionally, the surface projection micro-stereolithography 3D printing molding size error prediction system further includes a training module, configured as follows: Based on a pre-built error sample set, at least two candidate prediction models are selected from the regression model library. The error sample set includes multiple error samples, which contain the standard size, actual molding size error, and process parameter combination corresponding to the standard test sample. The at least two candidate prediction models are trained based on the error sample set to obtain at least two trained target prediction models; Based on the at least two trained target prediction models, an initial stacked generalization ensemble model is constructed using the two-layer stacked generalization architecture. Based on the error sample set, the initial stacked generalized ensemble model is trained to adjust the output weights of each target prediction model in the initial stacked generalized ensemble model, thereby obtaining the target stacked generalized ensemble model.

[0121] Optionally, the training module is specifically configured as follows: For each error sample, the standard dimensions and process parameters in the error sample are combined and input into each of the target prediction models in the initial stacking generalization ensemble model to obtain the first predicted forming size error output by each target prediction model; Each of the first predicted molding size errors is input into the meta-learner in the initial stacked generalization ensemble model. The meta-learner performs a weighted summation of each of the first predicted molding size errors based on the output weights of each of the target prediction models to obtain the second predicted molding size error. Based on the second predicted molding size error and the actual molding size error in the error sample, the loss value is determined; The output weights of each target prediction model in the meta-learner are adjusted based on the loss value. Continue training the initial stacked generalized ensemble model until the training stopping condition is met, and obtain the target stacked generalized ensemble model.

[0122] Optionally, the surface projection micro-stereolithography 3D printing molding size error prediction system further includes a construction module configured as follows: Based on each of the target process parameters, a sampling algorithm is used to generate different combinations of process parameters, and the combination of process parameters includes the parameter values ​​corresponding to each of the target process parameters. Based on the combination of process parameters, different standard test samples are printed using the surface projection micro stereolithography 3D printing equipment to obtain the actual molding size of each standard test sample. For each of the standard test specimens, the error of the actual molding size corresponding to the standard test specimen is determined based on the standard size corresponding to the standard test specimen and the actual molding size. The error sample set is constructed based on the standard dimensions, actual molding size errors, and process parameter combinations corresponding to each standard test sample.

[0123] Optionally, the surface projection micro-stereolithography 3D printing molding size error prediction system further includes a screening module, configured as follows: A significance analysis was performed on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to select at least one target process parameter.

[0124] Optionally, the building module is specifically configured as follows: Based on the combination of process parameters, different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment, and images corresponding to each standard test sample are obtained. Image processing algorithms are used to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size of each of the standard test samples. Data cleaning is performed on the initial molding dimensions corresponding to each of the standard test specimens to obtain the actual molding dimensions corresponding to each of the standard test specimens.

[0125] Optionally, the building module is specifically configured as follows: The errors of the standard dimensions and actual molding dimensions corresponding to each of the standard test specimens are respectively standardized; Based on the standard dimensions and actual molding size errors corresponding to each of the standardized standard test specimens, and the process parameter combinations corresponding to each of the standard test specimens, an error sample set is constructed, which includes a training set and a test set.

[0126] Optionally, the filtering module is specifically configured as follows: The Plackett-Burman experimental design was used to perform significance analysis and two-level screening on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to obtain at least one target process parameter. Each of the candidate process parameters includes exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time; Each of the target process parameters includes at least one of the exposure time, the exposure power density, the printing layer thickness, and the leveling time.

[0127] Optionally, the sampling algorithm is the Latin hypercube sampling algorithm; the standard test sample includes micropillar arrays with different design diameter gradients, and each group of micropillar arrays is repeatedly arranged in a preset number.

[0128] Optionally, the building module is specifically configured as follows: For each image, the image is converted to the HSV color space, and a scale outline is extracted based on the color threshold range. The pixel equivalent is then determined based on the scale outline. The image is converted to grayscale and Gaussian blurred for noise reduction. An edge detection operator is used to extract the edges. The center and radius of the micro-pillar are determined by voting in the accumulator space using the Hough gradient circular transform. The physical diameter is calculated by combining the pixel equivalent, and the initial forming size corresponding to the standard test sample in the image is obtained.

[0129] Optionally, the at least two trained target prediction models include at least two of random forest, gradient boosting tree, optimized distributed gradient boosting library, Gaussian process regression, multilayer perceptron and ridge regression.

[0130] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can call logic instructions in the memory 530 to execute a method for predicting the molding size error in surface projection micro-stereolithography 3D printing. This method includes: inputting the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design dimensions of the printed object into a trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object; the target process parameters are process parameters that have a significant impact on the molding size error; the target stacking generalization ensemble model adopts a two-layer stacking generalization architecture; the first layer of the two-layer stacking generalization architecture is a base learner, and the second layer of the two-layer stacking generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each target process parameter and the design dimensions; the meta-learner is used to determine the target molding size error based on the output weights of each target prediction model and the initial molding size error output by each target prediction model.

[0131] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0132] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the surface projection micro-stereolithography 3D printing molding size error prediction method provided by the above methods. The method includes: inputting the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design size of the printed object into a trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object; the target process parameter is a process parameter that has a significant impact on the molding size error; the target stacking generalization ensemble model adopts a two-layer stacking generalization architecture; the first layer of the two-layer stacking generalization architecture is a base learner, and the second layer of the two-layer stacking generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction model is used to determine the initial molding size error based on the parameter values ​​of each target process parameter and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each target prediction model and the initial molding size error output by each target prediction model.

[0133] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the method for predicting the molding size error of surface projection micro-stereolithography 3D printing provided by the above methods. The method includes: inputting the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design dimensions of the printed object into a trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object; the target process parameters are process parameters that have a significant impact on the molding size error; the target stacking generalization ensemble model adopts a two-layer stacking generalization architecture; the first layer of the two-layer stacking generalization architecture is a base learner, and the second layer of the two-layer stacking generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each target process parameter and the design dimensions; the meta-learner is used to determine the target molding size error based on the output weights of each target prediction model and the initial molding size error output by each target prediction model.

[0134] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0135] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting dimensional errors in surface projection micro-stereolithography 3D printing, characterized in that, include: The parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design dimensions of the printed object are input into the trained target stacking generalization ensemble model to predict the molding size error, thereby obtaining the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

2. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 1, characterized in that, The training process of the target stacked generalized ensemble model includes: Based on a pre-built error sample set, at least two candidate prediction models are selected from the regression model library. The error sample set includes multiple error samples, which contain the standard size, actual molding size error, and process parameter combination corresponding to the standard test sample. The at least two candidate prediction models are trained based on the error sample set to obtain at least two trained target prediction models; Based on the at least two trained target prediction models, an initial stacked generalization ensemble model is constructed using the two-layer stacked generalization architecture. Based on the error sample set, the initial stacked generalized ensemble model is trained to adjust the output weights of each target prediction model in the initial stacked generalized ensemble model, thereby obtaining the target stacked generalized ensemble model.

3. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 2, characterized in that, The step of training the initial stacked generalization ensemble model based on the error sample set, and adjusting the output weights of each target prediction model in the initial stacked generalization ensemble model to obtain the target stacked generalization ensemble model, includes: For each error sample, the standard dimensions and process parameters in the error sample are combined and input into each of the target prediction models in the initial stacking generalization ensemble model to obtain the first predicted forming size error output by each target prediction model; Each of the first predicted molding size errors is input into the meta-learner in the initial stacked generalization ensemble model. The meta-learner performs a weighted summation of each of the first predicted molding size errors based on the output weights of each of the target prediction models to obtain the second predicted molding size error. Based on the second predicted molding size error and the actual molding size error in the error sample, the loss value is determined; The output weights of each target prediction model in the meta-learner are adjusted based on the loss value. Continue training the initial stacked generalized ensemble model until the training stopping condition is met, and obtain the target stacked generalized ensemble model.

4. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 2 or 3, characterized in that, Before selecting at least two candidate prediction models from the regression model library based on a pre-built error sample set, the process also includes: Based on each of the target process parameters, a sampling algorithm is used to generate different combinations of process parameters, and the combination of process parameters includes the parameter values ​​corresponding to each of the target process parameters. Based on the combination of process parameters, different standard test samples are printed using the surface projection micro stereolithography 3D printing equipment to obtain the actual molding size of each standard test sample. For each of the standard test specimens, the error of the actual molding size corresponding to the standard test specimen is determined based on the standard size corresponding to the standard test specimen and the actual molding size. The error sample set is constructed based on the standard dimensions, actual molding size errors, and process parameter combinations corresponding to each standard test sample.

5. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 4, characterized in that, Before generating different combinations of process parameters based on the target process parameters using a sampling algorithm, the process further includes: A significance analysis was performed on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to select at least one target process parameter.

6. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 4, characterized in that, Based on the various combinations of process parameters, different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment to obtain the actual molded dimensions of each standard test sample, including: Based on the combination of process parameters, different standard test samples are printed using the surface projection micro-stereolithography 3D printing equipment, and images corresponding to each standard test sample are obtained. Image processing algorithms are used to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size of each of the standard test samples. Data cleaning is performed on the initial molding dimensions corresponding to each of the standard test specimens to obtain the actual molding dimensions corresponding to each of the standard test specimens.

7. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 4, characterized in that, The error sample set is constructed based on the standard dimensions corresponding to each standard test sample, the actual molding size error, and the combination of process parameters, including: The errors of the standard dimensions and actual molding dimensions corresponding to each of the standard test specimens are respectively standardized; Based on the standard dimensions and actual molding size errors corresponding to each of the standardized standard test specimens, and the process parameter combinations corresponding to each of the standard test specimens, an error sample set is constructed, which includes a training set and a test set.

8. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 5, characterized in that, The significance analysis of each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment is performed to screen out at least one target process parameter, including: The Plackett-Burman experimental design was used to perform significance analysis and two-level screening on each candidate process parameter of the surface projection micro-stereolithography 3D printing equipment to obtain at least one target process parameter. Each of the candidate process parameters includes exposure time, exposure power density, printing layer thickness, leveling time, cleaning time, and drying time; Each of the target process parameters includes at least one of the exposure time, the exposure power density, the printing layer thickness, and the leveling time.

9. The method for predicting dimensional errors in surface projection micro-stereolithography 3D printing according to claim 6, characterized in that, The step of using image processing algorithms to perform scale recognition, micro-pillar extraction, and size measurement on the images corresponding to each of the standard test samples to obtain the initial molding size corresponding to each of the standard test samples includes: For each image, the image is converted to the HSV color space, and a scale outline is extracted based on the color threshold range. The pixel equivalent is then determined based on the scale outline. The image is converted to grayscale and Gaussian blurred for noise reduction. An edge detection operator is used to extract the edges. The center and radius of the micro-pillar are determined by voting in the accumulator space using the Hough gradient circular transform. The physical diameter is calculated by combining the pixel equivalent, and the initial forming size corresponding to the standard test sample in the image is obtained.

10. A system for predicting dimensional errors in surface projection micro-stereolithography 3D printing, characterized in that, include: The molding size error prediction module is configured to input the parameter values ​​of each target process parameter in the surface projection micro-stereolithography 3D printing equipment and the design size of the printed object into the trained target stacking generalization ensemble model to predict the molding size error and obtain the target molding size error corresponding to the printed object. The target process parameters are process parameters that have a significant impact on the molding size error; the target stacked generalization ensemble model adopts a two-layer stacked generalization architecture; the first layer of the two-layer stacked generalization architecture is a base learner, and the second layer of the two-layer stacked generalization architecture is a meta-learner; the base learner contains at least two trained target prediction models; the target prediction models are used to determine the initial molding size error based on the parameter values ​​of each of the target process parameters and the design size; the meta-learner is used to determine the target molding size error based on the output weights of each of the target prediction models and the initial molding size error output by each of the target prediction models.

Images