CN-121562843-B - Multi-workpiece parallel 3D printing time prediction method based on machine learning fusion model

CN121562843BCN 121562843 BCN121562843 BCN 121562843BCN-121562843-B

Abstract

The invention relates to a multi-workpiece parallel 3D printing time prediction method based on a machine learning fusion model, which relates to the field of 3D printing technology and machine learning fusion, and comprises the steps of collecting multi-workpiece parallel 3D printing data, preprocessing, and forming a standardized data set through standardization; the method comprises the steps of establishing a feature association probability model based on Bayesian network structure learning and a maximum likelihood estimation method, mining potential association among features in a standardized dataset based on the feature association probability model to generate association features, screening the generated association features to form a final feature set, establishing LightGBM a model, establishing a multi-workpiece parallel 3D printing time prediction base model based on Bayesian optimization on the basis of the final feature set, and improving the precision and stability of multi-workpiece parallel 3D printing time prediction.

Inventors

WU CHUAN
SHEN ZHE
YANG BOYAN
WANG MIN
Meng Tingyao
Dai Jingcen
ZHANG XIAO

Assignees

徐州医科大学

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (9)

1. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model is characterized by comprising the following steps of: collecting parallel 3D printing data of multiple workpieces, preprocessing, and standardizing to form a standardized data set; Constructing a feature association probability model based on Bayesian network structure learning and a maximum likelihood estimation method, wherein the feature association probability model is as follows: determining the structure of a Bayesian network by using a K2 algorithm by taking a feature vector of a standardized data set as input, sorting according to the classification of geometric parameters, technological parameters and layout parameters, setting a feature node sequence, taking a Bayesian information criterion as a scoring function, setting the maximum number of parent nodes to be 3 by traversing the parent node combination of the feature nodes, and selecting the network structure with the maximum Bayesian information criterion value as an optimal Bayesian network structure; for each feature node in the optimal bayesian network structure Based on its parent node Based on maximum likelihood estimation method, parameter learning is carried out on the conditional probability table of the Bayesian network to obtain conditional probability distribution among features Wherein, the Is characterized by Is a parent node feature set of (a); Mining potential associations between features in a standardized dataset based on a feature association probability model, generating associated features, and The generated association features are filtered to form a final feature set; And constructing LightGBM a model, and establishing a basic model of multi-workpiece parallel 3D printing time prediction based on Bayesian optimization on the basis of the final feature set.
2. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 1 is characterized in that a single scene minimum sample size is calculated based on statistical sample sizes, and according to the minimum sample size, the total sample size is obtained by combining actual scene acquisition sample sizes, wherein the data set covers multi-workpiece parallel printing scenes of different workpiece types, different printing equipment models and different production batches and comprises workpiece geometric parameters, printing process parameters and workpiece layout parameters.
3. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 2, wherein the preprocessing comprises the following steps: carrying out missing value statistics on collected data and calculating the missing quantity of each characteristic, wherein for continuous characteristics, interpolation is carried out by adopting a method for constructing a regression model based on characteristic correlation, a multiple linear regression model is constructed by carrying out correlation analysis by calculating Pearson correlation coefficients among the continuous characteristics, and the missing values are filled by utilizing the trained multiple linear regression model; adopting a box diagram method and combining 3D printing process constraint to correct abnormal values, namely calculating quartiles of each feature And According to the formula Determining abnormal value judgment threshold value to be smaller than Or is greater than And correcting by adopting a mode of replacing the average value of adjacent samples, and checking the numerical value through process constraint.
4. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 3, wherein the method is characterized in that potential correlations among features in a standardized dataset are mined based on a feature correlation probability model, and specifically comprises the following steps: For any two features And According to conditional probability Calculating the association coefficient of the two The calculation formula is as follows: ; Wherein, the As a result of the covariance, And Standard deviations of the j and k th features, respectively; screening out feature pairs with absolute values of association coefficients larger than 0.6 to generate feature interaction items And ; Dividing the workpiece geometric parameter set, the printing process parameter set and the workpiece layout parameter set into three groups of characteristics respectively; For each group of characteristics, the Bayesian network posterior probability is firstly based Screening for effective characteristics, wherein As a virtual target variable, the virtual target variable, And (3) reducing the dimension of each group of features by adopting PCA as the features in the parameter group, and extracting the first 3 main components as feature combination associated features.
5. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 4, wherein the screening of the associated features comprises the following steps: combining features in the standardized dataset with the generated associated features to form an initial feature set ; The RFE method is adopted for the initial feature set Screening; taking RMSE of LightGBM model as evaluation index, gradually eliminating features based on model performance contribution, and reserving features with feature importance ranking 20 before ranking to form final feature set 。
6. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 5, wherein the Bayesian optimization comprises: randomly sampling M groups of super parameter combinations in a super parameter space, respectively training LightGBM models, and calculating objective function values corresponding to each group of super parameters to form an initial sample set; Based on the initial sample set, repeatedly executing the following operations until the iteration number reaches a preset maximum iteration number or the variation of the objective function value is smaller than a preset threshold value, and obtaining a super-parameter combination, wherein the operations comprise: adopting GP as a proxy model of Bayesian optimization, and fitting the GP model based on an initial sample set to obtain a probability mapping relation between the super-parameters and the objective function values; adopting an EI function as an Acquisition function, and calculating the EI value of each candidate super parameter in the super parameter space; Selecting a candidate super-parameter combination with the maximum EI value, training LightGBM a model, calculating an objective function value of the candidate super-parameter combination, adding the super-parameter combination and the objective function value into a sample set, and updating a proxy model; based on the obtained optimal super-parameter combination, training LightGBM a model, and constructing a multi-workpiece parallel 3D printing time prediction basic model.
7. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 6, wherein the final feature set is obtained by The corresponding dataset is divided into a training set and a test set with the aim of minimizing CVRMSE using the LightGBM model K-fold cross-validation CVRMSE on the training set as an objective function.
8. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 1, further comprising the step of performing transfer learning on a basic model, wherein the method is specifically as follows: For each scene, sample data in the scene is screened from the data set to form a scene data set; Adopting a transfer learning method, taking parameters of a basic model as priori knowledge, fine-tuning the model based on a scene data set, adjusting the learning rate and setting training rounds; and executing fine adjustment operation to finally obtain the scene-adaptive prediction model.
9. The multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model of claim 8 is characterized in that for each scene-adapted prediction model, a corresponding scene test set is obtained by dividing a scene data set, the performance of the prediction model is verified by using the test set, if the verification result meets the preset requirement, the scene-adapted prediction model can be put into practical application, and if the verification result does not meet the preset requirement, the following operations are executed until the performance requirement is met: Readjusting the super parameter optimization range; Increasing the amount of sample data in the scene; The base model is retrained.

Description

Multi-workpiece parallel 3D printing time prediction method based on machine learning fusion model Technical Field The invention relates to the field of 3D printing technology and machine learning fusion, in particular to a multi-workpiece parallel 3D printing time prediction method based on a machine learning fusion model, which is suitable for accurately predicting printing time in a multi-workpiece parallel 3D printing scene and provides technical support for 3D printing production scheduling, cost control and efficiency optimization. Background The 3D printing technology is used as a core technology for additive manufacturing, and is widely applied to the fields of aerospace, medical treatment, automobiles and the like by virtue of the advantages of no need of a die, customizable production and the like. With the continuous improvement of production demands, multi-workpiece parallel 3D printing becomes an important mode for improving production efficiency. However, in the process of parallel printing multiple workpieces, the printing time is affected by various factors, such as geometric parameters (volume, surface area, complexity, etc.) of the workpiece, printing process parameters (layer height, printing speed, filling rate, etc.), equipment parameters (nozzle temperature, platform temperature, etc.), and layout parameters (spacing, arrangement mode, etc.) of the workpiece on the printing platform, so that the printing time is difficult to accurately predict, and the conventional 3D printing time prediction method mainly comprises an empirical formula method and a simple machine learning method. The empirical formula method is based on a large number of experimental data fitting to obtain a relation between printing time and related parameters, but the method can only consider a small number of key parameters, cannot comprehensively cover complex influence factors under a multi-workpiece parallel printing scene, has poor generalization capability and has larger prediction errors under different scenes. The simple machine learning method such as a support vector machine, a common decision tree and the like can consider more parameters, but the characteristic engineering links adopt manual screening or simple statistical methods to generate the characteristics, so that potential association among the characteristics is difficult to mine, and meanwhile, the model super-parameters are adjusted by manual experience, so that the optimal performance of the model cannot be ensured, and the prediction precision and stability are difficult to meet the actual production requirements. In other prior art, for example, chinese patent publication No. CN112380716B discloses a SLA 3D printing time prediction method and system based on a learning algorithm, the content is based on an SLA process, the method is not suitable for other 3D printing technologies, a large amount of historical data is required to train time deviation coefficients, the adaptability to novel materials and processes is poor, the method is not suitable for a multi-task scene, chinese patent publication No. CN109648856B discloses a 3D printer processing time estimation method, the content depends on standard component simulation, but the average speed assumption ignores individual differences, the cooperative effect of workpiece combination is not considered, the prediction result cannot be dynamically adjusted, the traditional method cannot effectively process the space interference and thermodynamic coupling effect among workpieces, the quantitative modeling is lacking in resource competition during multi-task parallel, the cooperative effect of workpiece combination is not considered in characteristic engineering, and super-parameter optimization is not specially optimized for the multi-task scene. Disclosure of Invention The invention aims to solve the problems of imperfect characteristic engineering, insufficient model super-parameter optimization, low prediction precision and the like in the existing multi-workpiece parallel 3D printing time prediction method, provides a multi-workpiece parallel 3D printing time prediction method based on a machine learning fusion model, improves the precision and stability of multi-workpiece parallel 3D printing time prediction, and provides a reliable basis for 3D printing production management. In order to solve the technical problems, the invention adopts the following technical scheme that the multi-workpiece parallel 3D printing time prediction method based on the machine learning fusion model comprises the following steps: collecting parallel 3D printing data of multiple workpieces, preprocessing, and standardizing to form a standardized data set; constructing a feature association probability model based on Bayesian network structure learning and a maximum likelihood estimation method; Mining potential associations between features in a standardized dataset based on a feature association probability