Aerosol 3D printing quality real-time monitoring and system drift intelligent correction method

Abstract

The invention discloses an aerosol 3D printing quality real-time monitoring and system drift intelligent correction method, and belongs to the field of deep learning, and the method comprises the steps: carrying out the numerical sampling of aerosol 3D printing technological parameters, and obtaining a printing work window based on the technological parameter data; the relation between the technological parameters and the line quality characteristics is fitted through a Gaussian process regression method, a process model is obtained, and technological parameter control data are generated through the process model to control the printing process; acquiring image data of an aerosol printing result, and evaluating the image data by replacing a backbone network of the ViT model with a DINOv2 model to obtain a drift state result of a printing system; when the printing system drifts, the printing work window and the process model are corrected through transfer learning, the printing process is controlled through the corrected printing work window and process model, rapid correction of the drifting system can be achieved, and the aerosol printing precision is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of additive manufacturing and deep learning technology, and in particular relates to a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift. Background Art

[0002] Aerosol Jet 3D Printing (AJP) plays a crucial role in the fabrication of microelectronics, flexible electronics, and functionally graded materials. As a high-precision additive manufacturing technology, AJP atomizes nano / micron-scale materials into an aerosol and then focuses them for deposition. This enables the fabrication of complex electronic component structures with micron-level resolution. AJP is also compatible with a wide range of functional materials (such as conductive inks and ceramic slurries), demonstrating unique advantages in the fields of microsensors, embedded circuits, and biomedical devices.

[0003] However, its industrial application is limited by the dynamic coupled optimization of process parameters, especially the coordinated control of sheath gas flow rate, carrier gas flow rate, printing speed, nozzle pressure, and nozzle-substrate distance. For example, insufficient atomization pressure or carrier gas flow fluctuations can lead to uneven distribution of aerosol particles, causing film thickness deviations or microcracks; mismatch between the nozzle and substrate spacing can cause edge diffusion or "coffee ring" phenomena, affecting the conductivity of high-density interconnects; and mismatch between deposition speed and drying rate can lead to insufficient particle fusion, reducing the mechanical strength and electrical properties of functional layers (such as conductive lines). These defects jointly restrict the miniaturization of devices, functional integration, and the reliability of mass production. Therefore, multi-parameter coordinated optimization and real-time optimization of process parameters are necessary to ensure that the deposition morphology accuracy, functional characteristics, and process stability meet the stringent requirements of micro-nano devices, ensure the quality of electronic components, and improve production efficiency.

[0004] Traditional methods for optimizing aerosol printing process quality often rely on operator experience to adjust parameters and ignore the impact of system offsets, leading to a dynamic mismatch between material flow characteristics and the molding environment during the printing process. For example, phase transition hysteresis of aerosol particles and the coupled effects of multiple physical fields (thermal, mechanical, and rheological) can easily lead to interlayer bonding strength fluctuations and residual stress accumulation during material deposition, which in turn can cause quality defects such as deformation. Furthermore, traditional quality inspection methods (such as coordinate measurement and CT scanning) are mostly performed only in the post-processing stage and are unable to monitor and capture defects during the dynamic molding process in real time, resulting in delayed problem detection and additional material waste. In contrast, numerical simulation techniques based on the finite element method, while able to predict aerosol flow and deposition behavior during aerosol printing to a certain extent, are computationally complex and lack the ability to incorporate external perturbations in real time, making them unable to fully meet the requirements of real-time quality control and optimization. Therefore, developing more accurate simulation methods that can respond to external perturbations in real time will be key to the widespread application of aerosol printing technology.

[0005] In addition, system drift may also occur during aerosol jet printing, greatly affecting the printing quality. System drift refers to the phenomenon that the key process parameters or equipment component performance in the printing process deviate gradually and unexpectedly over time, resulting in a dynamic deviation between the output characteristics and the preset target. This drift is mainly caused by the coupling of the following multiple factors: (1) hardware performance degradation, such as the gradual blockage of the nozzle micropores due to material residue and the decrease in the stability of the atomizer pump pressure output; (2) thermodynamic parameter fluctuations caused by environmental disturbances, such as temperature and humidity changes leading to abnormal evaporation rate of aerosol solvents, destroying the gas-liquid two-phase flow equilibrium; (3) rheological property deviation caused by material batch differences, especially the change of nanoparticle concentration distribution or ink viscoelasticity with storage time; (4) real-time monitoring data distortion caused by sensor calibration drift, which makes the feedback signal of the closed-loop control system mismatch with the actual physical process. System drift is cumulative and time-varying in its impact on printing quality. Initially, it may manifest as random fluctuations of ±5% in line width, but over time, it can lead to cross-scale defects. At the microscopic level, it can cause nanoparticle deposition trajectory deviations, resulting in submicron-level layer thickness unevenness. At the macroscopic level, it can exacerbate the "coffee ring" edge effect, causing resistivity distribution variations of over 30% for conductive lines with a width of 10-50μm. Furthermore, it can cause interlayer misalignment in three-dimensional stacked structures, reducing the fatigue life of 100μm-scale functional devices by over 40%. Furthermore, this drift exhibits nonlinear propagation characteristics. A single parameter drift (e.g., a 0.1L / min sheath gas flow rate drift) can exponentially degrade printing resolution through multi-physics coupling within the aerosol focusing field. Traditional control strategies based on fixed parameter sets struggle to adapt to these time-varying disturbances, often requiring frequent manual compensation adjustments by the operator. This not only causes fluctuations in production line yields by 15%-20%, but also results in an additional 30% in material verification losses for each batch, severely limiting the economic and reliability of continuous production.

[0006] In summary, aerosol printing, as an important branch of additive manufacturing technology, has shown great potential in many high-end fields. However, in the process of realizing industrial-scale production, it is still necessary to solve the problems of inability to perform parameter coupling optimization, offset detection and offset correction, and further improve the effectiveness and robustness of the printing system. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention proposes a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift to solve the problems existing in the above-mentioned prior art.

[0008] To achieve the above objectives, the present invention provides a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift, comprising:

[0009] Numerical sampling of aerosol 3D printing process parameters is performed to obtain process parameter data. Based on the process parameter data, a machine learning model is used to obtain a printing working window, which includes the value range of the optimal process parameters. The relationship between the process parameters and the line quality characteristics is fitted using the Gaussian process regression method to obtain a process model. Based on the printing working window, the process model is used to generate corresponding process parameter control data to control the printing process.

[0010] During the printing process, image data of the aerosol printing result is obtained, and the image data is evaluated using a drift evaluation model to obtain a drift result of the printing system; wherein the drift evaluation model adopts a ViT model in which the backbone network is replaced by a DINOv2 model;

[0011] When there is drift in the printing system, the printing working window and process model are corrected through transfer learning, and the printing process is controlled by the corrected printing working window and process model.

[0012] Optionally, a Latin hypercube sampling method is used to numerically sample the process parameters of aerosol 3D printing.

[0013] Optionally, the machine learning model adopts a support vector machine, wherein the comprehensive printing quality corresponding to the process parameter data is obtained, the support vector machine is supervised learning is performed through the process parameter data and the corresponding comprehensive printing quality, and the optimal hyperplane of the support vector machine is obtained as the printing working window; wherein the comprehensive printing quality is the quantitative analysis result of the printing result corresponding to the process parameter data.

[0014] Optionally, the process model generation process includes:

[0015] The relationship between process parameters and line quality characteristics is modeled using the Gaussian process regression method, wherein the process parameters include sheath gas flow rate, carrier gas flow rate, and printing speed, and the line quality characteristics include line width and line thickness, to obtain a process model. The process model is fitted to obtain corresponding process parameters according to the required line quality characteristics, and the printing process is controlled by the fitted process parameters.

[0016] Optionally, before evaluating the image data using the drift estimation model, the following steps are also included:

[0017] Unsupervised learning is performed on the DINOv2 model, and the backbone network of the ViT model is replaced by the DINOv2 model after unsupervised learning. The replaced ViT model is trained with small-sample supervised learning to obtain a drift evaluation model.

[0018] Optionally, the process of evaluating the image data includes:

[0019] The DINOv2 model is used to extract features from image data to obtain key features. The remaining parts of the ViT model are used to fuse and identify the key features to obtain the final classification results. The final classification results are evaluated to determine whether there is drift in the printing system.

[0020] Optionally, the process of unsupervised learning of the DINOv2 model includes:

[0021] Obtaining sample image data of aerosol printing results, wherein the sample image data is unlabeled; preprocessing the sample image data; the DINOv2 model includes a teacher network and a student network;

[0022] The teacher network and the student network are trained, wherein the network parameters of the teacher network are updated by sliding average according to the network parameters of the student network, and the student network is trained by contrastive learning through the output of the teacher network; during the training process, a weighted sum of the contrastive learning loss and the teacher network loss is used as the loss function; and an optimized DINOv2 model is obtained.

[0023] Optionally, the process of training the replaced ViT model for small-sample supervised learning includes:

[0024] Freeze the model parameters of the DINOv2 model after unsupervised learning, obtain labeled sample images, extract labeled sample images through the DINOv2 model after unsupervised learning, recognize them through the remaining part of the ViT model, optimize the remaining part of the ViT model, and obtain the drift evaluation model.

[0025] Optionally, transfer learning uses the Instance Transfer transfer learning method.

[0026] On the other hand, the present invention provides a system for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift, which is used to execute the above method.

[0027] Compared with the prior art, the present invention has the following advantages and technical effects:

[0028] (1) System optimization of printing parameters based on machine learning: At the beginning of printing, the printing parameters can be systematically optimized to improve printing quality; (2) Online monitoring of system drift based on deep learning and machine vision: At the beginning of printing, the AJP printing system status is analyzed in real time to see if it has drifted; (3) After drift of the printing system is detected, a transfer learning algorithm is used to effectively integrate the knowledge of the original working window and process model with the newly sampled samples to achieve rapid correction of the drift system, reduce the impact of system drift during the printing process, and effectively improve the stability and repeatability of aerosol printing.

[0029] The organic combination of machine learning-based printing parameter system optimization, deep learning-based system drift online monitoring, transfer learning-based drift system rapid correction and machine vision optimizes printing quality from different angles and stages, and improves the effectiveness and robustness of the printing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0031] Figure 1 This is a flow chart of a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift according to an embodiment of the present invention;

[0032] Figure 2 This is a diagram of the main hardware configuration of an embodiment of the present invention, wherein: Figure 2 (a) is the hardware configuration diagram for the initial working window and printing process model recognition based on the coaxial camera. Figure 2 (b) shows the hardware configuration diagram for online monitoring of printing system drift based on process cameras during printing.

[0033] Figure 3 is a schematic diagram of a Latin Hypercube Sampling (LHS) experimental design in a workspace for the main working parameters of the embodiment of the present invention, wherein: Figure 3 (a) is a schematic diagram of the experimental point design in a three-dimensional workspace based on the Latin hypercube sampling method. Figure 3 (b) is a schematic diagram of the experimental point design in a two-dimensional workspace based on the Latin hypercube sampling method;

[0034] Figure 4 is a schematic diagram of the quality analysis of printed lines according to an embodiment of the present invention, wherein: Figure 4 (a) is a schematic diagram for defining the average width and average roughness of printed lines. Figure 4 (b) is a schematic diagram for defining the average thickness of printed lines;

[0035] Figure 5 1 is a schematic diagram of an original working parameter window of aerosol printing based on SVM classification according to an embodiment of the present invention;

[0036] Figure 6 is a main working parameter-print quality relationship process model based on GPR in an embodiment of the present invention, wherein: Figure 6 (a) is a schematic diagram of the process model of the relationship between sheath gas flow rate, carrier gas flow rate and printed line width. Figure 6(b) is a schematic diagram of the process model of the relationship between printing speed, carrier gas flow rate and printing line width. Figure 6 (c) is a schematic diagram of the process model of the relationship between printing speed, sheath gas flow rate and printed line thickness;

[0037] Figure 7 Schematic diagram of feature extraction of different printed line samples based on the DINOv2 unsupervised learning algorithm according to an embodiment of the present invention, wherein: Figure 7 (a1-a6) in the figure are schematic diagrams of different forms of printed original line samples. Figure 7 (b1-b6) in FIG. 3 is a schematic diagram of line morphological features extracted from the printed line samples corresponding to (a1-a6);

[0038] Figure 8 1 is a performance diagram of a classification model according to an embodiment of the present invention that is based on the fusion of DINOv2 and ViT models and is trained using a large number of automatically labeled samples;

[0039] Figure 9 is a comparison chart of defect recognition performance analysis of the ViT classification model under a small sample size embodiment of the present invention, wherein: Figure 9 (a) is a schematic diagram showing the comparison of the generalization ability of the ViT model with the YOLO and CNN models under different sample numbers. Figure 9 (b) is a schematic diagram showing the comparison of the robust performance of the ViT model, YOLO and CNN models under the condition of adding noise interference. Figure 9 (c) is a schematic diagram of the sample obtained after applying different types of noise interference to the printed sample;

[0040] Figure 10 This is a schematic diagram of analyzing whether the printing system has drifted based on the trained ViT classification model. At the same time, combined with the attention mechanism, the gradient information visualization method is used to highlight the line features and enhance the interpretability of the model judgment. Figure 10 (a1-a6) in the figure are schematic diagrams of different forms of printed original line samples. Figure 10 (b1-b6) in the figure are schematic diagrams of printed line sample features corresponding to (a1–a6) obtained by combining the attention mechanism with the gradient information visualization method;

[0041] Figure 11 This is a schematic diagram of the drift causes and impact analysis of the aerosol printing process, where: Figure 11 (a) Schematic diagram of the analysis of the causes of drift during the printing process. Figure 11 (b) is a schematic diagram showing the effect of system drift caused by long-term printing on the cross-sectional morphology of printed lines. Figure 11(c) is a schematic diagram showing the effect of system drift caused by long-term printing on the printed line width. Figure 11 (d) is a schematic diagram showing the effect of system drift caused by long-term printing on the capacitance characteristics of printed components;

[0042] Figure 12 This is a schematic diagram of selecting representative experimental points based on the incremental clustering algorithm, where: Figure 12 (a) is a schematic diagram of the selection of representative experimental points in the source model space of the printed line width. Figure 12 (b) is a schematic diagram of the selection of representative experimental points in the source model space of printed line thickness;

[0043] Figure 13 This is a schematic diagram of the working window after the drifted system working window is quickly corrected based on the transfer learning algorithm;

[0044] Figure 14 This is a schematic diagram of the process model after rapid correction of the drifting system process model based on the transfer learning algorithm, where: Figure 14 (a) is a schematic diagram comparing the line width model obtained based on transfer learning and the actual measurement results. Figure 14 (b) is a schematic diagram comparing the line thickness model obtained based on transfer learning and the actual measurement results. DETAILED DESCRIPTION

[0045] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0047] The core technology area of ​​this invention patent can be categorized as "intelligent additive manufacturing quality control and process optimization," specifically involving aerosol jet 3D printing process optimization, real-time quality monitoring, and rapid correction of system drift. The technologies involved include aerosol jet printing (AJP) additive manufacturing, machine learning, deep learning, transfer learning, and machine vision.

[0048] In response to the problems existing in the prior art, the present invention proposes a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift. This method organically combines machine learning, deep learning technology, and transfer learning with real-time computer vision to optimize AJP process parameters, while also enabling online monitoring of system drift and rapid correction of drift systems. Specifically, in this method, an integrated machine learning strategy is first used to optimize the main operating parameters of the print, identify the initial optimal workspace and process model, and ensure that the operating parameters are systematically optimized at the beginning of the initial printing process. Subsequently, in order to reduce the impact of system drift on printing defects and dimensional inaccuracies during the AJP process, as well as on the functionality and reliability of the final product, the invention deploys a deep learning model based on the fusion of Self-Distilled Vision Transformer 2 (DINOv2) and Vision Transformer (ViT) to monitor system drift and analyze the status of the AJP printing system in real time. In addition, after monitoring the system drift, a transfer learning algorithm is used to effectively integrate the knowledge of the original working window and process model with the newly sampled samples to achieve rapid correction of the drift system and effectively improve the stability and repeatability of aerosol printing.

[0049] The present invention can combine advanced machine learning, deep learning, transfer learning, machine vision and other technologies to perform fine process parameter optimization, real-time monitoring and rapid migration correction of the aerosol printing system, which can effectively improve the stability and repeatability of aerosol printing, ensure that the functional components produced can meet strict mechanical properties, geometric accuracy and surface quality requirements, and thus promote its application in high-precision fields such as aerospace and biomedicine.

[0050] Describe the technical process of the above technical solution:

[0051] To address the aforementioned technical issues, the present invention proposes a method for real-time monitoring of aerosol 3D printing quality and intelligently correcting system drift. This method systematically optimizes AJP print quality during the initial print job and simultaneously analyzes the AJP printing system status in real time during the initial printing process. Furthermore, upon detecting system drift, a transfer learning algorithm is employed to effectively integrate knowledge from the original working window and process model with newly sampled data, enabling rapid correction of drift and effectively improving the stability and repeatability of aerosol printing.

[0052] This method combines machine learning, deep learning techniques, transfer learning, and real-time computer vision to systematically optimize AJP process parameters while enabling online monitoring of system drift and rapid correction of drift. Specifically, it involves the following three parts:

[0053] S1: Initial working window and printing process model identification based on machine learning:

[0054] (S1.1) In the process parameter space, the Latin hypercube sampling (LHS) experimental design method is used for experimental design;

[0055] (S1.2) Establish a quantitative analysis standard for the quality of aerosol printed lines and systematically analyze the quality of printed lines;

[0056] (S1.3) Quantify the topographic features of printed line samples generated under the LHS experimental design framework and use support vector machines (SVM) to define the initial optimal operating window to achieve robust classification of normal / abnormal lines;

[0057] (S1.4) Use the Gaussian process regression method to establish an initial mapping model of aerosol printing working parameters and line features to guide autonomous decision-making and robust regulation of quality optimization during the printing process.

[0058] The initial working window and process model identified in the first part can systematically optimize the print quality of aerosol jet 3D printing during the initial printing work.

[0059] S2: Online monitoring of system drift based on deep learning and machine vision:

[0060] (S2.1) In the workspace, use an aerosol printer to randomly print a large number of unlabeled samples and collect images related to the samples;

[0061] (S2.2) Subsequently, the DIBNOv2 model is trained for unsupervised learning based on a large number of unlabeled samples. The trained model is used to extract unsupervised features of lines.

[0062] (S2.3) Fuse the DINOv2 and ViT models and train them for supervised learning based on a small number of labeled printed line samples;

[0063] (S2.4) The trained ViT model can efficiently and accurately identify abnormal samples and their abnormal locations, and perform online evaluation of whether the printing status is drifting in real time.

[0064] S3: Rapid drift correction of aerosol printing system based on transfer learning:

[0065] Based on the developed online monitoring model, when the printing status is diagnosed as drifting, the system's initial working window and process model will be quickly corrected using transfer learning methods. The specific steps are as follows:

[0066] (S3.1) Design and select representative experimental points in the printing space for the drift system;

[0067] (S3.2) Based on the selected experimental points, perform new sample collection and feature extraction in the drift system;

[0068] (S3.3) Use the transfer learning method to extract the printed physical knowledge in the initial working window and process model identified in the S1 stage, and combine it with the new sample data to migrate and update the target working window and process model to achieve rapid correction of the drift system.

[0069] In view of the above, the technical solution of the present invention includes the following components and their principles, actions and functional uses:

[0070] Coaxial camera: Set up as Figure 2 As shown in (a), Figure 2 (a) is a hardware configuration diagram for systematic optimization of printing process parameters based on a coaxial camera before starting aerosol printing / changing printing materials / changing the working environment. A mobile printing platform is used to transport the printed sample placed on the platform to the vertical position directly below the lens of the coaxial camera. The coaxial camera is then used to capture images of the printed sample, and the print quality of the sample is quantitatively analyzed based on formulas (1)-(6).

[0071] Process Camera: Set up as Figure 2 As shown in (b), Figure 2 (b) shows the hardware configuration for real-time evaluation and online quality correction of printed lines during aerosol printing, using a process camera. The camera is installed near the nozzle of the aerosol 3D printer and maintains relative motion with the nozzle during printing. It simultaneously captures line images of printed samples at an angle of approximately 45-60 degrees. The developed quality monitoring model is then used to assess the real-time printing status based on these captured sample images.

[0072] During the printing process, the main feature images of the printed sample are collected at a certain angle to make print quality decisions;

[0073] Latin Hypercube Sampling (LHS) is an efficient experimental design method with the following advantages: efficient coverage of the input space, reduction of the number of samples, avoidance of sample centralization, applicability to high-dimensional problems, and global exploration and optimization of the AJP multi-factor working parameter space with minimal experimental effort.

[0074] Support Vector Machine (SVM): A classification algorithm based on core techniques and the maximum margin principle, which achieves robust binary classification of normal / abnormal printing patterns by constructing a high-dimensional hyperplane.

[0075] Gaussian process regression model (GPR): is a probability-based non-parametric regression method that uses Gaussian processes to model the distribution of data, thereby making predictions on unknown data points and giving the uncertainty of the predictions.

[0076] Because the established GPR model not only generates predicted output responses across the entire workspace but also provides quantitative information about the uncertainty of the corresponding predictions, the parameter solution with the least uncertainty can be selected from multiple candidate solutions during the parameter adjustment process, thereby achieving autonomous decision-making for quality optimization and effectively improving the corresponding robustness.

[0077] Self-Distilled Vision Transformer 2 (DINOv2): is an unsupervised learning feature extraction model that mainly uses self-distillation technology for unsupervised learning and has demonstrated very strong performance in vision tasks.

[0078] Vision Transformer (ViT): is an image classification model based on the Transformer architecture. ViT can effectively model the global dependencies of images through the Transformer's self-attention mechanism, which makes ViT advantageous in handling long-distance dependencies and capturing global features.

[0079] Incremental clustering algorithms select a few representative points from a large number of experimental points in the design space. These points effectively capture the diversity of the design space and represent different potential regions or regional characteristics. These points are selected to reduce the number of experiments and computational effort while ensuring that the selected points are representative of the characteristics of the entire design space.

[0080] Transfer learning: In transfer learning, the original task and the target task often share some similarities, so knowledge from the original task can be leveraged to accelerate or improve the learning process of the target task. The basic idea of ​​transfer learning is that existing knowledge can be transferred between different tasks or domains without having to train the model from scratch.

[0081] The above technical solution is described in detail with reference to the relevant drawings:

[0082] The overall workflow of this invention is as follows Figure 1 As shown, combined with Figure 2 The dual-imaging system shown in the figure involves technical fields including machine learning-based system optimization of operating parameters, deep learning-based online drift system monitoring, and transfer learning-based rapid correction of system drift. The specific steps are described as follows:

[0083] Step S1. Identification of the initial working window and printing process model based on machine learning:

[0084] (S1.1) In the process parameter space, the Latin Hypercube Sampling (LHS) experimental design method is used for experimental design:

[0085] The key advantages of the Latin Hypercube Sampling (LHS) method include efficient coverage of the input space, reduced sample size, avoidance of sample concentration, applicability to high-dimensional problems, and the absence of assumptions about the distribution of the input variables. By evenly partitioning the input variable interval and randomly sampling from it, LHS ensures that each dimension is fully represented in the input space, thereby providing more accurate results with fewer samples. Its excellent sample distribution characteristics make LHS widely used in fields such as computer experiments, optimization problems, and uncertainty analysis, particularly when dealing with complex models, significantly improving computational efficiency and accuracy.

[0086] The present invention performs reasonable sampling in the process parameter space through Latin hypercube sampling, wherein the process parameters include sheath gas flow rate, carrier gas flow rate, and printing speed.

[0087] like Figure 3 As shown, the left figure shows the LHS sampling experiment design in a three-dimensional workspace; the right figure shows the LHS sampling experiment design in a two-dimensional workspace with a fixed printing speed. The core idea of ​​Latin hypercube sampling is to divide the range of each input parameter into intervals of equal probability and then randomly select a sample point from each interval. This ensures that the value range of each input variable is evenly covered, effectively avoids concentration between samples, and provides a more comprehensive sample distribution. In the multidimensional case, LHS can reduce the number of sampling points while maintaining good coverage of the input space.

[0088] The main steps are as follows:

[0089] (1) Define variable ranges: Assume that the problem has Sk input variables (or factors), and each variable has a defined range. For example, the possible range of values ​​for the variable SHGFR (sheath gas flow rate) is [a1, b1], the possible range of values ​​for the variable CGFR (carrier gas flow rate) is [a2, b2], the possible range of values ​​for the print speed is [a3, b3], and so on.

[0090] (2) Dividing intervals: Divide the value range of each input variable evenly to obtain different evenly divided intervals.

[0091] (3) Random sampling: For each variable, a sample point is randomly selected from each corresponding interval, that is, a value is selected. To ensure uniform coverage, the sample positions in each interval of each variable are usually not repeated, that is, only one sample point is selected in each interval.

[0092] (4) Constructing the sample matrix: Combine the sample points of each variable to form a Sn×Sk matrix, where Sn is the number of sample points and Sk is the dimension of the input variable. Each row represents a sample point and each column represents an input variable.

[0093] (5) Arrangement and optimization: In order to increase the uniformity of the samples, the sampling points of each column, i.e., different values ​​of the same process parameters, are randomly arranged. This can avoid the repetitive patterns of certain variables in all samples, thereby enhancing the diversity of the samples.

[0094] (S1.2) Establish a quantitative analysis standard for the quality of aerosol printed lines and systematically analyze the quality of printed lines:

[0095] The schematic diagram of print line quality analysis is as follows Figure 4 As shown, the print line quality analysis is as follows:

[0096] (1) Average value of scattered line width ( ): Calculated by spatially averaging the pixel intensity data by column:

[0097] = (1)

[0098] in Represents the discretized line width of the i-th column, and N is the total number of columns.

[0099] (2) Average line thickness ( ): By calculating the cross-sectional area (c) and the line width ( ) is calculated and the mean is taken M times:

[0100] (2)

[0101] (3) Average linear density ( ): Based on the identified print line edge, calculate the average grayscale intensity of the pixels in the surrounding area and deduct the background light intensity:

[0102] (3)

[0103] in is the discretized average grayscale intensity of the i-th column, is the average background intensity (used to offset the interference of illumination on pixel intensity).

[0104] (4) Average line edge roughness ( ): Calculate the standard deviation of the actual edge fluctuation based on the reference line (average line):

[0105] = (4)

[0106] Among them are and Respectively represent the deviation of the upper and lower edges of the i-th column relative to the mean line.

[0107] (5) Line discontinuity ( ): Statistical average of the number of upper and lower edge detection failures:

[0108] = (5)

[0109] in and The number of failed upper and lower edge detections, respectively.

[0110] (6) Overall print quality ( ): Obtained by weighted summation of main line feature parameters, used for binary classification of sample quality:

[0111] (6)

[0112] in , and is the weight coefficient of line density, edge roughness and discontinuity.

[0113] Subsequently, because line width and roughness are the main morphological features of printed lines, the width of the quantified lines ( ) and roughness ( ), use analysis of variance (ANOVA) to calculate the F value and compare it with the critical value. If the F value is greater than the critical value, or the P value is less than α, the null hypothesis is rejected. This identifies statistically significant parameters from all process parameters, ignoring those that are not significant, and simplifying the model.

[0114] The P-value is a probability value used in statistical hypothesis testing, measuring the probability that the observed data would produce the current or more extreme result if the null hypothesis were true. The α value is a pre-set threshold used to decide whether to reject the null hypothesis. It is typically set at 0.05 (5%), indicating a 5% probability of incorrectly rejecting the null hypothesis. If the P-value is less than the α value, the result is considered statistically significant, and the null hypothesis is rejected.

[0115] (7)

[0116] Specifically, MSB (mean square between groups) measures the degree of variation in the means of different groups, reflecting the treatment effect or the difference between groups; MSW (mean square within groups) represents the random fluctuation of the data within each group, reflecting individual errors or intra-group variation.

[0117] (8)

[0118] (9)

[0119] in, is the number of groups, is the sample size of group i, is the mean of group i, is the overall mean, is the jth observation value in the i-th group, and N is the total sample size.

[0120] (S1.3) Support vector machine (SVM) is used to define the initial optimal operating window to achieve robust classification of normal / abnormal lines.

[0121] The comprehensive printing quality of the line is calculated based on formula (6): ), where the weight coefficients in formula (6) can be assigned according to the quality requirements and the weight sum is 1. Then, based on the set judgment threshold of the comprehensive quality of the printed lines, the quality of the printed line samples is divided into two categories: normal and abnormal. Then, the support vector machine (SVM) and kernel function method are used to identify the optimal decision boundary of the classification statistics and use it as the printing working window, as shown in the figure. Figure 5 shown.

[0122] The input of the support vector machine in this content is the process parameters, and its output is whether the line quality is normal. By finding the corresponding optimal hyperplane, the range of process parameters belonging to the normal classification can be clearly defined as the printing working window.

[0123] Specifically, a support vector machine is a supervised learning algorithm primarily used for classification and regression tasks. Its core concept is to find an optimal hyperplane that maximizes the margin between data of different categories, thereby improving the model's generalization ability. In the above description, since normal print samples and abnormal samples are not linearly separable in the operating space, a Gaussian kernel function K is used to map the training features into a high-dimensional space:

[0124] (10)

[0125] in and Two samples from the training dataset, is the Gaussian kernel variance.

[0126] At the same time, slack variables are introduced into the objective function to relax the constraints, making the training data approximately linearly separable in the new space and identifying nonlinear decision boundaries. Therefore, the problem of identifying the optimal operating window in the design space is transformed into the following convex quadratic programming problem to find the classification hyperplane, i.e., the optimal working space.

[0127] (11)

[0128] (12)

[0129] Among them, st represents the condition. is the normal vector of the decision hyperplane, is the bias term, is the mapping function of sample x in high-dimensional space, Represents the distance between the interval and the data point: If the sample is correctly classified by the hyperplane, then ;like is on the wrong side of the hyperplane, then c is a custom penalty coefficient used to control the training error. The present invention uses a uniform grid set as the initial parameter to search for the optimal hyperparameters c and The schematic diagram of the optimal parameter working space (optimal working window surrounded by the decision boundary) identified by the source model (support vector machine) is shown in Figure 5 As shown, the working parameters within the optimal working space can achieve the best printing quality and ensure the printing of normal line samples. Parameters outside the printing working space will cause abnormal printing quality.

[0130] The support vector machine (SVM) classifies print line quality by finding an optimal hyperplane in the workspace. First, the SVM aims to find a hyperplane that maximizes the separation between the two data points. This means maximizing the distance from the closest point (i.e., the support vector) to the hyperplane, while ensuring that every data point is correctly classified. For linearly separable data, the SVM uses a simple linear function to find the optimal hyperplane. However, for nonlinearly separable data, the SVM uses a kernel function to project the data into a higher-dimensional feature space, making it linearly separable in this space. In this new, higher-dimensional space, the SVM can find a linear hyperplane to separate the data, despite the nonlinear distribution of the data in the original space. Common kernel functions include polynomial kernels and Gaussian radial basis kernels (RBFs), which help the SVM find the optimal segmentation boundary in high-dimensional space. In this way, the SVM can handle complex nonlinear problems and ensure accurate classification.

[0131] (S1.4) Use the Gaussian process regression method to establish an initial model of the aerosol printing working parameter-line feature process.

[0132] GPR (Gaussian Process Regression) is a non-parametric regression method based on Bayesian theory. Unlike traditional parametric regression models, it does not build a model by fitting parameters, but directly models by defining a potential function space. GPR updates the model through Bayesian reasoning, that is, calculating the posterior distribution of the potential function through known data. Because it is a non-parametric model, GPR does not require a specific distribution form of the data in advance. It can adapt to complex, unknown relationships and is very suitable for processing high-dimensional and nonlinear data. Given a set of observations, GPR calculates the conditional probability distribution, predicts the function value of the new input point and gives the uncertainty of the prediction. The specific modeling conclusions are as follows:

[0133] Assuming implicit function With a zero-mean Gaussian process (GP) prior distribution, the covariance matrix is ,but The expression is:

[0134] (13)

[0135] is the prediction function, is Gaussian noise, is the training response, X is all the input data points in the training data, then the predicted points With training data They obey the joint Gaussian distribution:

[0136] (14)

[0137] Posterior predictive distribution of line features of aerosol prints based on Bayesian inference , p represents the probability distribution function:

[0138] (15)

[0139] in, Indicates that after given observation data y, the prediction function and the joint distribution of the training function f. represents the marginal distribution of the observed data y, Represents the prediction function and the joint prior distribution of the training function f. represents the distribution of observed responses y given a training function f.

[0140] The predicted mean of the line feature is printed for:

[0141] (16)

[0142] Prediction variance for:

[0143] (17)

[0144] in, Represents the predicted data point The covariance matrix between . Represents the predicted data point and training data points The covariance matrix between . Represents the covariance matrix between training data points. Represents the identity matrix.

[0145] In the present invention, sheath gas flow rate, carrier gas flow rate, and printing speed are used as the main process parameters, and line width and line thickness are used as the main line quality characteristics for GPR modeling. The schematic diagram of the obtained process model is shown in FIG. Figure 6 As shown, in Figure 6 In the figure, (a), (b), and (c) correspond to the process models of the sheath gas flow rate and carrier gas flow rate, the printing speed and carrier gas flow rate, and the sheath gas flow rate and printing speed combination, respectively. Figure 6 It includes the mean of the model prediction and the corresponding prediction variance, which helps to select and combine different working parameters within the above-mentioned optimal workspace during the parameter adjustment process to generate candidate solutions. The parameter solution with the least uncertainty can be selected from multiple candidate solutions and used as the real-time printing solution to control printing, thereby achieving autonomous decision-making for quality optimization and effective improvement of corresponding robustness.

[0146] Step S2. Online monitoring of system drift based on deep learning and machine vision:

[0147] Online monitoring of system drift is based on the fusion of the Self-Distilled Vision Transformer 2 (DINOv2) and Vision Transformer (ViT) models. The intelligent print quality monitoring model includes:

[0148] Backbone network, feature fusion module and classification head module;

[0149] The backbone network is used to extract key features from the image. Specifically, it uses the DINOv2 model based on self-supervised learning to optimize image feature representation through self-distillation and contrastive learning, thereby improving the quality and robustness of visual representation and replacing the backbone network module of ViT.

[0150] The feature fusion module is used to combine features at different levels or from different sources;

[0151] The classification head module is used for fully connected layer connection, converting feature maps into final classification results, and completing the final classification task.

[0152] (S2.1) Automatic collection of large quantities of unlabeled samples and labeling of small quantities of samples using an aerosol printer.

[0153] (1) First, the Latin hypercube sampling method is used to randomly design the process parameter space to ensure that the experimental points are evenly distributed in the entire design space, thereby providing a diverse sample basis for subsequent experiments.

[0154] (2) The aerosol printer performs random printing according to the designed experimental points, generating a large amount of different unlabeled sample data, which includes a variety of possible experimental results and variable combinations.

[0155] (3) Collect these generated samples, and the large amount of unlabeled data collected provides rich input for unsupervised model training.

[0156] (4) At the same time, a small number of samples are selected for manual labeling for subsequent supervised model training.

[0157] Through the above method, the efficiency of constructing the ViT model training dataset is improved while ensuring the diversity of data distribution.

[0158] (S2.2) Training of unsupervised learning model for feature extraction based on DINOv2.

[0159] DINOv2 is an unsupervised feature extraction model that primarily uses self-distillation for unsupervised learning and has demonstrated strong performance in visual tasks. DINOv2's training principle is based on the idea of ​​self-distillation, where training is performed through mutual guidance between a "teacher model" and a "student model." Specifically, it uses a "contrastive learning" approach. Unlike traditional contrastive learning, DINOv2's teacher and student are the states of the same model at different time steps, with no manual labels between them. The main principles are as follows:

[0160] (1) Self-distillation process: During training, the model’s student network learns the output features of the teacher network. The parameters of the teacher network are a sliding average of the student network’s parameters. In this way, the student network gradually learns from the teacher network. As training progresses, the performance of the student network becomes closer and closer to that of the teacher network.

[0161] (2) Feature contrast learning: DINOv2 uses a self-supervised contrastive loss function, which maximizes the similarity between different views while minimizing the similarity between different images. In this process, images are processed with different data augmentation processes, and each image is treated as a different "view", thereby generating a set of positive and negative samples.

[0162] (3) Clustering objective: DINOv2 encourages the model to learn discriminative visual features by comparing each image sample to itself. In this way, it not only captures global information but also grasps local details, thereby extracting more discriminative features.

[0163] The specific training process steps of DINOv2 are as follows:

[0164] (1) Data preparation: The unlabeled sample image dataset of aerosol printing is used. The data does not require manual annotation and only needs some preprocessing, such as image normalization. The unlabeled sample image set includes printing images under different process parameters collected by the Latin hypercube sampling method;

[0165] (2) Data augmentation: Apply two different random data augmentations (such as random cropping, color perturbation, random flipping, etc.) to each input image. This allows the model to perform comparative learning between different transformations of the same image.

[0166] (3) Initializing the model: In order to train DINOv2’s unsupervised learning capability to extract image features, standard pre-training methods or random initialization methods are usually used to initialize the parameters of the training model.

[0167] (4) Teacher network and student network:

[0168] Teacher Network: Initially, the parameters of the teacher network and the student network are the same. As training progresses, the parameters of the teacher network are updated by the sliding average of the student network.

[0169] Student network: The student network learns by comparing its output with the teacher network, with the goal of making the feature distribution of the student network similar to that of the teacher network.

[0170] (5) Self-distillation training: The teacher network and the student network are trained in parallel. The teacher network is updated by sliding average, and the student network is optimized by contrastive loss function. The goal of the student network is to replicate the output feature representation of the teacher network as much as possible. The parameter update formula of the teacher network is:

[0171] (18)

[0172] in, represents the parameters of the teacher network, represents the parameters of the student network, is the sliding mean coefficient, which is usually close to 1.

[0173] (6) Feature contrast learning: The contrast loss is used to minimize the distance between different enhanced views of the same image and maximize the distance between different images. Each pair of views is used to calculate their similarity to train the student network. The calculation formula is:

[0174] (19)

[0175] in They are respectively the student network outputs enhanced by different data, The cosine similarity of two feature vectors, Id represents the index of the sample.

[0176] (7) The training goal is to minimize the total loss function , the loss function includes contrast loss and distillation losses The total loss function can be written as:

[0177] (20)

[0178] in , are the weight coefficients for contrast loss and distillation loss, which are usually adjusted according to the experimental settings.

[0179] (8) Training iteration: This process is iterated continuously, and the model gradually improves its feature extraction ability through multiple training rounds (epochs), and finally learns discriminative visual representations on unlabeled data.

[0180] In summary, DINOv2 uses unsupervised learning and self-distillation to learn efficient and discriminative visual features without relying on labeled data. Its training process continuously optimizes the model's feature extraction capabilities and improves its learning performance through comparative learning between the teacher and student networks.

[0181] Figure 7 The printed line sample features extracted based on the DINOv2 model are shown, where (a1)-(a6) are the original images of the printed line samples, and (b1)-(b6) are the printed line sample features corresponding to the original images.

[0182] (S2.3) Training of ViT supervised learning models based on a small number of labeled samples.

[0183] ViT is an image classification model based on the Transformer architecture that has achieved remarkable results in the field of computer vision in recent years. Specifically, ViT effectively models global image dependencies through the Transformer's self-attention mechanism, making it advantageous in handling long-range dependencies and capturing global features. The core concept of ViT is to divide the image into small patches of fixed size and input the pixel information of these patches into the Transformer model for processing. Specifically, it includes modules such as image segmentation, linear mapping, position encoding, Transformer encoder, and classification head output.

[0184] In this invention, since DINOv2 has been trained as the backbone module of ViT and is ready to continue training the ViT model based on a small number of labeled samples, there is no need to retrain the entire ViT model. Instead, we only need to freeze the parameters of DINOv2 and use its trained features for ViT, avoiding training the feature extraction part from scratch. Then, we perform further supervised learning optimization on the entire model. Specifically, the training steps include the following:

[0185] (1) Freezing the backbone weights of DINOv2: Since DINOv2 has been trained on a large-scale dataset and has learned rich feature representations, the parameters of the DINOv2 part are frozen during the supervised learning training process. The weights of the DINOv2 module are not updated, and only the parameters of the ViT part are trained. This can avoid overfitting and reduce training time.

[0186] (2) Input and data preparation: Label samples are automatically collected and annotated for supervised learning tasks. The image features extracted by the DINOv2 model will serve as the input of ViT. Assuming that the output of DINOv2 is a high-dimensional feature vector of the image, ViT will process these features for further classification.

[0187] (3) Supervised learning in the ViT part: The training focus of the model shifts to the ViT part, especially the last few layers of Transformer and the classification head. The output of the Transformer is passed through a set of classification heads (fully connected layers) to make the final classification prediction, classifying the quality of the aerosol printing lines and judging whether the printing quality is normal or abnormal. The line shape of abnormal printing quality is as follows: Figure 7 As shown in (a1-a5), they are: line discontinuity, line overspray, line low thickness, line high edge roughness, and line high diffusion. The line shape of normal printing quality is as follows Figure 7 As shown in (a6) in .

[0188] (4) Loss function and optimization: Use cross entropy loss as the loss function. Specific loss function The expression is:

[0189] (twenty one)

[0190] in The category is The true label, The category predicted by the model is The probability of Indicates the total number of categories.

[0191] During training, an optimization algorithm (such as Adam or AdamW) can be used to update the weights of the ViT part to minimize the loss function and complete the training of the ViT model.

[0192] The trained DINOv2 is used as the backbone of ViT, and then the ViT model is trained based on labeled samples, combining the advantages of self-supervised learning and transfer learning. It enables the model to achieve better performance with less labeled data, while reducing computing resources and training time, effectively improving the task adaptability and performance of the model. The training process and performance of the ViT model are shown in the figure. Figure 8 In addition, further experiments (see Figure 9 ) Through the evaluation of modeling performance on small sample data sets, Figure 9 (a) in the figure is a comparison chart of the classification performance of the ViT classification model, the YOLO model, and the CNN model in the case of small samples. Figure 9 (b) shows the robustness comparison of the classification performance of the ViT classification model, the YOLO model, and the CNN model in the case of small samples. Figure 9(c) in the figure represents the image type used for robustness comparison, including processing methods such as contrast, blur, saturation, rotation, and noise. Comparative experiments demonstrate that the ViT model achieves good classification performance and robustness even with small sample sizes. The classification performance of various models was compared by introducing different forms of noise (such as brightness, contrast, saturation, hue, mirroring, and rotation). Experimental results demonstrate that the ViT model trained using this method outperforms the classic CNN convolutional neural network and YOLO object detection model in small sample size generalization ability, and also demonstrates superior noise robustness and stronger interference resistance.

[0193] In addition, the attention heat map is a visualization tool used to show how the model allocates attention to different parts of the input data when processing a specific task. It identifies the attention weight of each part of the input by changing the color. The attention heat map helps understand which parts of the input the model "pays attention to" when making decisions. The heat map can clearly see which input features or positions the model relies on when judging whether the printing system has shifted. This is very important for improving the interpretability of the model. Fusion of the trained ViT model with the attention heat map for online monitoring is beneficial for the AJP printing online evaluation system to suppress background noise during monitoring, enhance the feature response of the target area, accurately locate defects such as low density of printed lines and overspray, and help to robustly judge whether the printing system has shifted through heat map analysis, as shown in the figure. Figure 10 As shown, (a1)-(a6) are the original images of printed line samples, and (b1)-(b6) are the heat maps of printed line samples corresponding to the original images.

[0194] (S2.4) Real-time monitoring of printing system drift based on the ViT model:

[0195] The modal detection and threshold triggering algorithm ensures the operational stability of the system by real-time monitoring and continuous analysis of the predicted frequency of anomalies in successive printed layers.

[0196] First, a statistical modal threshold is calculated based on historical data. This threshold serves as a criterion for identifying system drift. Specifically, an anomaly threshold can be set by analyzing the normal fluctuation range of historical data (such as maximum and minimum fluctuation values). For example, if the frequency fluctuation of defects detected during printing exceeds a certain multiple (e.g., 1.5 times) of the maximum fluctuation value in the historical data, it is considered a systematic deviation. If the number of anomalies detected in the system exceeds the set modal threshold in multiple consecutive judgments, the algorithm determines that the anomaly is no longer simple random noise but rather a system drift phenomenon. At this point, the system triggers a rapid correction mechanism to quickly restore normal operation. This precise anomaly detection and correction triggering mechanism effectively avoids print quality issues or production errors caused by system drift, thereby ensuring ultimate accuracy and quality. This mechanism also reduces human intervention while improving production efficiency and the system's self-healing capabilities.

[0197] Step S3. Rapid drift correction of aerosol printing system based on transfer learning:

[0198] (S3.1) Design and selection of representative experimental points for the drift system:

[0199] Figure 11 Schematic diagram of the causes and effects of system drift. In order to quickly correct the drift system, the selected experimental points should be able to capture the key features of the target printing process to improve the predictive performance of the target model. A direct approach is to repeatedly perform Latin hypercube (LHS) experimental design on the target process. Although LHS can ensure uniform coverage of the target input feature space, this method may be suboptimal for nonlinear target processes with sufficient original data sets. Therefore, under the assumption that there is process similarity between the two fields, two rules are used to select representative experimental points of the target process:

[0200] (1) The representative target dataset should cover the boundaries and key local extreme points of the source domain;

[0201] (2) The distribution of target data in the design space should be selected by jointly analyzing the input space and the response space. Therefore, more experimental points should be allocated in areas where the response changes rapidly with the input. Based on this idea, an incremental clustering algorithm will be used to select representative experimental points in the design space. The specific design method is as follows:

[0202] By jointly analyzing the working parameters of aerosol printing - line characteristics, every point in the printing space The significance P of a representative experimental point is calculated as:

[0203] (twenty two)

[0204] in , are the dimensions of the input and output spaces, respectively. Indicates the point with the largest P value at the i-th coordinate in different directions of the printing space It is selected as a representative experimental point, and then the next representative experimental point is selected based on the iterative method:

[0205] (twenty three)

[0206] in is an adjustment parameter, which is set to 1.6 in the present invention. This update-selection process will be iterated continuously until the required number of representative experimental points are selected. Figure 12 Schematic diagram of selected representative experimental points.

[0207] (S3.2) New sample collection and feature extraction based on representative experimental points:

[0208] Because the inherent deposition mechanisms of the AJP process typically remain unchanged, certain key printing characteristics and process behaviors remain consistent even when system drift occurs. System drift typically refers to deviations in the performance of the printing process due to changes in materials, equipment, environmental conditions, and other factors. However, because these changes may not completely disrupt the underlying deposition mechanisms, certain characteristics will continue to exhibit similar patterns and trends. Therefore, transfer learning methods can be used to address this challenge. Transfer learning can leverage existing source domain model knowledge, combined with a small number of newly sampled samples, to rapidly correct the system after drift. Specifically, the source domain model provides a stable process model before drift, and through transfer learning, this knowledge can be transferred to the system after drift. Only a small amount of data from the target domain is required to adjust the model parameters to adapt to the new operating conditions. This approach effectively reduces the need for new data, reduces the time and cost of model training, and improves the system's ability to quickly adapt to drift, thereby ensuring printing quality and process stability.

[0209] In the present invention, the proportion of the selected representative experimental points can be set to 10%-30% of the initial model experimental points, and the specific proportion can be flexibly adjusted according to the actual effect and system requirements. The selection of representative experimental points is to maximize the potential of existing data while avoiding excessive data collection and computing overhead. By accurately selecting representative experimental points, the system characteristics and their changing trends can be more effectively captured, providing more accurate data support for subsequent model updates and optimizations. After feature extraction, the collected new samples can be effectively fused with the initial model. The fused model can combine the information in the new samples with the knowledge in the source model, thereby achieving rapid correction of the system after drift. This process not only accelerates the adaptation speed of the model, but also improves its accuracy and robustness. Through this method, a small amount of new data can be effectively utilized to achieve continuous optimization of complex systems, so that the model can still maintain good performance and stability when facing system changes or environmental changes.

[0210] (S3.3) Rapid correction of target working window and process model based on transfer learning:

[0211] (1) Rapid migration correction of the original working window:

[0212] Instance Transfer is a transfer learning method used to rapidly transfer and calibrate the original target working window. This method adapts the source task dataset to the target task by adjusting instance weights or directly modifying the source task data. The core idea behind this method is to "align" the source and target task data, using certain techniques to minimize the differences between them. This allows the source model's training to be directly transferred to the target task.

[0213] The main steps of migration are:

[0214] 1) Adjust the weights of the samples used when developing the original working window model:

[0215] By weighting or reselecting data samples in the source task, the data in the source task is made more suitable for the target task.

[0216] Weighted source task instances: Assign a weight to each source task instance so that more important instances (those that are closer to the target task data distribution) have a greater impact on model training. The specific weight assignment formula is:

[0217] (twenty four)

[0218] in, is a task instance of the original model The weight of Is the original task instance and target task instance The distance between them (usually using Euclidean distance or other similarity metrics), Ns is the total number of instances in the original model. The weights can be used to weight the training of the source model so that similar data points have a more significant transfer effect on the target task.

[0219] 2) Migrate the original model to the target model:

[0220] Fine-tuning a model trained on the source task is performed on the target task. Fine-tuning the source model on the target task data can further adjust the source model to adapt it to the target task. During fine-tuning, the target task dataset is usually used to optimize the parameters of the source task model:

[0221] (25)

[0222] in, is the predicted output of the target model, is the loss function, is the true label of the target task, are the parameters of the original model, is the total number of target tasks, and after fine-tuning, these parameters will adapt to the target tasks.

[0223] By assigning appropriate weights to instances in the original task through Instance Transfer, the data of the original task is aligned with the data of the target task. By training and weighting the source data on the original model and then fine-tuning it, the original model can be effectively transferred. The new working window after the specific transfer correction is shown in the schematic diagram. Figure 13 .

[0224] (2) Rapid migration correction of the original Gaussian process regression model:

[0225] Since the AJP process will follow similar material deposition mechanisms under different operating conditions, the trend similarities exhibited in the input-output relationships can be regarded as prior knowledge in different domains, and the input-output relationship changes in the target domain can be corrected by applying output slope / bias corrections.

[0226] (26)

[0227] in and denote the translation vector and scaling vector in the source domain response space, respectively. and represent the initial model and target model respectively, and They represent the input dataset of the initial model and the input dataset of the target model respectively.

[0228] In addition, considering the similarity of inputs, such as translation similarity and scaling similarity, input slope / bias correction is adopted to further adjust the working range differences between input feature spaces.

[0229] (27)

[0230] in, and Denote the translation vector and scaling vector in the source domain feature space, respectively. Therefore, based on the model transfer and the established initial model, the target model can be obtained in the following way.

[0231] (28)

[0232] Finally, by optimizing the following equations and combining a representative dataset collected from the target domain , the correction vector can be estimated :

[0233] (29)

[0234] in Represents the error between the actual observation value and the target model prediction output. The correction vector is obtained through genetic algorithm optimization. The new process model after migration correction is shown in the schematic diagram. Figure 14 .

[0235] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. Aerosol 3D printing quality real-time monitoring and system drift intelligent correction method, characterized by: include: Numerical sampling of aerosol 3D printing process parameters is performed to obtain process parameter data. Based on the process parameter data, a machine learning model is used to obtain a printing working window, which includes the value range of the optimal process parameters. The relationship between the process parameters and the line quality characteristics is fitted using the Gaussian process regression method to obtain a process model. Based on the printing working window, the process model is used to generate corresponding process parameter control data to control the printing process. During the printing process, image data of the aerosol printing result is obtained, and the image data is evaluated using a drift evaluation model to obtain a drift result of the printing system; wherein the drift evaluation model adopts a ViT model in which the backbone network is replaced by a DINOv2 model; When there is drift in the printing system, the printing working window and process model are corrected through transfer learning, and the printing process is controlled by the corrected printing working window and process model.

2. The method according to claim 1, characterized in that The Latin hypercube sampling method was used to numerically sample the process parameters of aerosol 3D printing.

3. The method according to claim 1, characterized in that The machine learning model adopts a support vector machine, wherein the comprehensive printing quality corresponding to the process parameter data is obtained, the support vector machine is supervised by learning the process parameter data and the corresponding comprehensive printing quality, and the optimal hyperplane of the support vector machine is obtained as the printing working window; wherein the comprehensive printing quality is a quantitative analysis result of the printing result corresponding to the process parameter data.

4. The method according to claim 1, wherein The process of generating a process model includes: The relationship between process parameters and line quality characteristics is modeled using the Gaussian process regression method, wherein the process parameters include sheath gas flow rate, carrier gas flow rate, and printing speed, and the line quality characteristics include line width and line thickness, to obtain a process model. The process model is fitted to obtain corresponding process parameters according to the required line quality characteristics, and the printing process is controlled by the fitted process parameters.

5. The method according to claim 1, wherein The image data is evaluated by the drift estimation model before: Unsupervised learning is performed on the DINOv2 model, and the backbone network of the ViT model is replaced by the DINOv2 model after unsupervised learning. The replaced ViT model is trained with small-sample supervised learning to obtain a drift evaluation model.

6. The method according to claim 1, characterized in that The process of evaluating image data includes: The DINOv2 model is used to extract features from image data to obtain key features. The remaining parts of the ViT model are used to fuse and identify the key features to obtain the final classification results. The final classification results are evaluated to determine whether there is drift in the printing system.

7. The method according to claim 5, characterized in that The process of unsupervised learning for the DINOv2 model includes: Obtaining sample image data of aerosol printing results, wherein the sample image data is unlabeled; preprocessing the sample image data; the DINOv2 model includes a teacher network and a student network; The teacher network and the student network are trained, wherein the network parameters of the teacher network are updated by sliding average according to the network parameters of the student network, and the student network is trained by contrastive learning through the output of the teacher network; during the training process, a weighted sum of the contrastive learning loss and the teacher network loss is used as the loss function; and an optimized DINOv2 model is obtained.

8. The method according to claim 1, characterized in that The process of training the replaced ViT model for small-sample supervised learning includes: Freeze the model parameters of the DINOv2 model after unsupervised learning, obtain labeled sample images, extract labeled sample images through the DINOv2 model after unsupervised learning, recognize them through the remaining part of the ViT model, optimize the remaining part of the ViT model, and obtain the drift evaluation model.

9. The method according to claim 1, characterized in that Transfer learning uses the Instance Transfer transfer learning method.

10. Aerosol 3D printing quality real-time monitoring and system drift intelligent correction system, characterized by: Used to execute the method according to any one of claims 1 to 9.

Images