Real-time monitoring of aerosol 3D printing quality and intelligent system drift correction method

Abstract

This invention discloses a method for real-time quality monitoring and intelligent system drift correction in aerosol 3D printing, belonging to the field of deep learning. The method includes: numerically sampling the process parameters of aerosol 3D printing to obtain and, based on the process parameter data, acquiring a printing working window; fitting the relationship between process parameters and line quality characteristics using Gaussian process regression to obtain a process model; generating process parameter control data through the process model to control the printing process; acquiring image data of the aerosol printing results; evaluating the image data by replacing the backbone network of the ViT model with the DINOv2 model to obtain the drift state result of the printing system; when drift exists in the printing system, correcting the printing working window and process model through transfer learning; and controlling the printing process through the corrected printing working window and process model, thereby achieving rapid correction of the drift system and effectively improving the accuracy of aerosol printing.

Description

Technical Field

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

[0002] Aerosol Jet 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 / micro-scale materials into aerosols and focuses their deposition, enabling the formation of complex electronic component structures with micron-level resolution. It is also compatible with various functional materials (such as conductive inks and ceramic pastes), demonstrating unique advantages in the fields of microsensors, embedded circuits, and biomedical devices.

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

[0004] In optimizing the quality of aerosol printing processes, traditional methods often rely on operator experience for parameter adjustments, neglecting the impact of system offsets. This leads to a dynamic mismatch between material flow characteristics and the forming environment during printing. For example, the phase transition hysteresis of aerosol particles and the multi-physics (thermo-mechanical-rheological) coupling effects make it easy for interlayer bonding strength fluctuations and residual stress accumulation to occur during material deposition, resulting in quality defects such as deformation. Meanwhile, traditional quality inspection methods (such as coordinate measurement and CT scanning) are mostly performed only in the post-processing stage, unable to monitor and capture defects in the dynamic forming process in real time, leading to delayed problem detection and additional material waste. In contrast, while numerical simulation technology based on the finite element method can predict aerosol flow and deposition behavior to some extent during aerosol printing, its computational complexity is high, and it is difficult to couple external disturbances in real time, failing to fully meet the needs of real-time quality control and optimization. Therefore, developing a more accurate simulation method that can respond to external disturbances in real time will be key to the widespread application of aerosol printing technology.

[0005] In addition, system drift can also occur in aerosol jet printing, which greatly affects print quality. System drift refers to the unexpected gradual deviation of key process parameters or equipment component performance over time during the printing process, resulting in dynamic deviation of output characteristics from the preset target. This drift is mainly caused by the following multi-factor coupling effects: (1) Hardware performance degradation, such as gradual blockage of nozzle micro-orifices due to material residue and decreased stability of atomizing pump pressure output; (2) Fluctuations in thermodynamic parameters caused by environmental disturbances, such as abnormal evaporation rate of aerosol solvent due to temperature and humidity changes, which disrupts the gas-liquid two-phase flow balance; (3) Rheological characteristic deviation caused by batch differences in materials, especially the change in nanoparticle concentration distribution or ink viscoelasticity with storage time; (4) Distortion of real-time monitoring data caused by sensor calibration drift, which makes the feedback signal of the closed-loop control system mismatched with the actual physical process. System drift has a cumulative and time-varying effect on print quality: initially, it may only manifest as random fluctuations in linewidth of ±5%, but over time it can lead to cross-scale defects—at the microscopic level, it causes deviations in the deposition trajectory of nanoparticles, resulting in submicron-level layer thickness unevenness; at the macroscopic level, it exacerbates the "coffee ring" edge effect, causing resistivity distribution differences of more than 30% in conductive lines with a width of 10-50μm; and at the macroscopic level, it leads to interlayer misalignment in three-dimensional stacked structures, reducing the fatigue life of 100μm-level functional devices by more than 40%. More seriously, this drift has nonlinear propagation characteristics; a single parameter (such as a drift of 0.1L / min in sheath gas velocity) can cause an exponential degradation in print resolution through multi-physical coupling effects of the aerosol focusing field. Traditional control strategies based on fixed parameter sets are difficult to adapt to such time-varying disturbances, often requiring frequent manual compensation adjustments by operators. This not only causes fluctuations in production line yield of 15%-20%, but also results in an additional 30% material verification loss per batch, severely restricting the economy and reliability of continuous production.

[0006] In conclusion, although aerosol printing, as an important branch of additive manufacturing technology, has shown great potential in many high-end fields, it still needs to solve the problems of inability to perform parameter coupling optimization, offset detection and offset correction in the process of achieving industrial-scale production. Further improvements are needed to enhance the effectiveness and robustness of the printing system. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method for real-time quality monitoring and intelligent system drift correction in aerosol 3D printing, thereby resolving the issues present in the prior art.

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

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

[0010] During the printing process, image data of the aerosol printing results are acquired, and the image data is evaluated using a drift evaluation model to obtain the drift result of the printing system; wherein the drift evaluation model adopts the ViT model by replacing the backbone network with the DINOv2 model.

[0011] When the printing system experiences drift, the printing work window and process model are corrected through transfer learning, and the printing process is controlled using the corrected printing work window and process model.

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

[0013] Optionally, the machine learning model employs a support vector machine, wherein the overall printing quality corresponding to the process parameter data is obtained, and the support vector machine is subjected to supervised learning using the process parameter data and the corresponding overall printing quality to obtain the optimal hyperplane of the support vector machine as the printing working window; wherein the overall printing quality is the quantitative analysis result of the printing result corresponding to the process parameter data.

[0014] Optionally, the process of generating the process model includes:

[0015] The relationship between process parameters and line quality characteristics is modeled using Gaussian process regression. The process parameters include sheath gas velocity, carrier gas velocity, and printing speed, while the line quality characteristics include line width and line thickness. A process model is obtained, in which the process model is fitted to obtain the corresponding process parameters according to the required line quality characteristics. The fitted process parameters are then used to control the printing process.

[0016] Optionally, the following steps may be included before evaluating the image data using the drift evaluation model:

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

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

[0019] Feature extraction is performed on image data using the DINOv2 model to obtain key features. The remaining part of the ViT model is then used to fuse and identify the key features to obtain the final classification result. The final classification result is evaluated to determine whether the printing system has drift.

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

[0021] Acquire sample image data of aerosol printing results, wherein the sample image data is unlabeled; preprocess 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 moving average based on 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, the weighted sum of the loss of contrastive learning and the loss of the teacher network is used as the loss function; thus, the optimized DINOv2 model is obtained.

[0023] Optionally, the process of training the replaced ViT model using few-shot supervised learning includes:

[0024] The model parameters of the DINOv2 model after unsupervised learning are frozen, labeled sample images are obtained, labeled sample images are extracted using the unsupervised DINOv2 model, identification is performed using the remaining part of the ViT model, and the remaining part of the ViT model is optimized to obtain the drift evaluation model.

[0025] Optionally, the transfer learning method is adopted.

[0026] On the other hand, the present invention provides a real-time quality monitoring and intelligent system drift correction system for aerosol 3D printing, for performing the above-described method.

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

[0028] (1) Machine learning-based printing work parameter system optimization: At the start of printing work, the printing work parameters can be optimized in a systematic way to improve printing quality; (2) System drift online monitoring based on deep learning and machine vision: During the start of printing work, the state of the AJP printing system is analyzed in real time to see if it shifts; (3) After the printing system drift is detected, the 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 printing, 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 online drift monitoring, transfer learning-based rapid drift correction, and machine vision optimizes print quality from different angles and stages, improving the effectiveness and robustness of the printing system. Attached Figure Description

[0030] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

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

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

[0033] Figure 3 This is a schematic diagram of the Latin hypercube sampling (LHS) experiment design in the workspace for the main working parameters of identification according to an embodiment of the present invention. Figure 3 (a) in the figure is a schematic diagram of the experimental point design in a three-dimensional workspace based on the Latin hypercube sampling method. Figure 3 (b) in the figure is a schematic diagram of the experimental point design in a two-dimensional workspace based on the Latin hypercube sampling method;

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

[0035] Figure 5 This is a schematic diagram of the original working parameters window for aerosol printing based on SVM classification in an embodiment of the present invention;

[0036] Figure 6 This is a process model based on the relationship between GPR's main working parameters and print quality, according to an embodiment of the present invention. Figure 6 (a) in the diagram is a schematic diagram of the process model relating sheath gas velocity, carrier gas velocity, and printed line width. Figure 6(b) in the diagram is a schematic diagram of the process model relating printing speed, carrier gas flow rate, and printed line width. Figure 6 (c) in the figure is a schematic diagram of the process model relating printing speed, sheath gas flow rate and printed line thickness;

[0037] Figure 7 This is a schematic diagram illustrating feature extraction from different printed line samples using the DINOv2 unsupervised learning algorithm, according to an embodiment of the present invention. Figure 7 (a1-a6) in the diagram represent different forms of the original printed line sample. Figure 7 (b1-b6) in the figure is a schematic diagram of the line shape features extracted from the printed line samples corresponding to (a1-a6);

[0038] Figure 8 This is a schematic diagram illustrating the performance of a classification model based on the fusion of DINOv2 and ViT models and trained with a large number of automatically labeled samples, according to an embodiment of the present invention.

[0039] Figure 9 This is a comparison chart showing the defect recognition performance of the ViT classification model under small sample conditions according to embodiments of the present invention. Figure 9 (a) in the figure is a schematic diagram comparing the generalization ability of the ViT model with that of the YOLO and CNN models under different sample sizes. Figure 9 (b) in the figure is a schematic diagram comparing the robustness performance of the ViT model with that of the YOLO and CNN models under the condition of added noise interference. Figure 9 (c) in the figure is a schematic diagram of the samples obtained after applying different types of noise interference to the printed samples;

[0040] Figure 10 This diagram illustrates the analysis of whether a printing system experiences drift based on a trained ViT classification model. It also incorporates an attention mechanism and uses gradient information visualization to highlight line features, enhancing the interpretability of the model's judgments. Figure 10 (a1-a6) in the diagram represent different forms of the original printed line sample. Figure 10 (b1-b6) in the figure is a schematic diagram of the printed line sample features corresponding to (a1–a6) obtained after combining the attention mechanism and the gradient information visualization method;

[0041] Figure 11 This is a schematic diagram illustrating the causes and effects of drift during the aerosol printing process. Figure 11 (a) Schematic diagram illustrating the analysis of the causes of printing process drift. Figure 11 (b) in the diagram illustrates the effect of system drift caused by long-term printing on the cross-sectional shape of the printed lines. Figure 11(c) in the diagram illustrates the effect of system drift caused by prolonged printing on the printed line width. Figure 11 (d) in the figure 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 illustrating the selection of representative experimental points based on the incremental clustering algorithm. Figure 12 (a) in the diagram is a schematic diagram of selecting representative experimental points in the source model space of the printed line width. Figure 12 (b) in the figure is a schematic diagram of the selection of representative experimental points in the source model space of the printed line thickness;

[0043] Figure 13 This is a schematic diagram of the working window after rapid correction of the drifting system working window 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. Figure 14 (a) in the figure is a comparison diagram between the line width model obtained based on transfer learning and the actual measurement results. Figure 14 (b) in the figure is a comparison diagram between the line thickness model obtained based on transfer learning and the actual measurement results. Detailed Implementation

[0045] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

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

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

[0048] To address the problems existing in current technologies, this invention proposes a method for real-time quality monitoring and intelligent system drift correction in aerosol 3D printing. This method organically combines machine learning, deep learning, and transfer learning with real-time computer vision to optimize AJP process parameters while simultaneously achieving online monitoring and rapid correction of system drift. Specifically, this method first employs an integrated machine learning strategy to optimize the main printing parameters, identifying the initial optimal workspace and process model to ensure systematic optimization of the parameters at the start of the initial printing operation. Subsequently, to reduce the impact of system drift on printing defects and dimensional inaccuracies, as well as the functionality and reliability of the final product, caused by system drift during AJP operations, the invention deploys a deep learning model based on the fusion of Self-Distilled Vision Transformer 2 (DINOv2) and Vision Transformer (ViT) for system drift monitoring and real-time analysis of the AJP printing system status. Furthermore, after detecting system drift, a transfer learning algorithm is used to effectively fuse the knowledge of the original working window and process model with the newly sampled samples, thereby achieving rapid correction of the drifting system and effectively improving the stability and repeatability of aerosol printing.

[0049] This invention combines advanced machine learning, deep learning, transfer learning, and machine vision technologies to perform fine-grained process parameter optimization, real-time monitoring, and rapid transfer correction on aerosol printing systems. This can effectively improve the stability and repeatability of aerosol printing, ensuring that the produced functional components meet stringent requirements for mechanical performance, geometric accuracy, and surface quality, thereby promoting its application in high-precision fields such as aerospace and biomedicine.

[0050] The technical process of the above technical solution is described as follows:

[0051] To address the aforementioned technical problems, the basic concept of this invention is to propose a method for real-time monitoring of aerosol 3D printing quality and intelligent correction of system drift. This method can systematically optimize the printing quality of AJP during the initial printing job; simultaneously, it analyzes the AJP printing system status in real time during the printing process; furthermore, after detecting system drift, it employs a transfer learning algorithm to effectively fuse the knowledge of the original working window and process model with newly sampled samples, achieving rapid correction of the drifting system and effectively improving the stability and repeatability of aerosol printing.

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

[0053] S1: Initial working window and printing process model recognition 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) The morphological features of the printed line samples generated under the LHS experimental design framework are quantified, and the initial optimal operation window is defined using support vector machine (SVM) to achieve robust classification of normal / abnormal lines.

[0057] (S1.4) Using the Gaussian process regression method, an initial mapping model of aerosol printing working parameters-line feature process is established to guide autonomous decision-making and robust adjustment of quality optimization during the printing process.

[0058] The initial working window and process model identified in the first part enable systematic optimization of print quality in aerosol jet 3D printing during the initial print job.

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

[0060] (S2.1) In the workspace, a large number of unlabeled samples are randomly printed using an aerosol printer, and images of the samples are collected.

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

[0062] (S2.3) The DINOv2 model is fused with the ViT model and trained using 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 evaluate whether the printing status has drifted in real time.

[0064] S3: Rapid drift correction in aerosol printing systems based on transfer learning:

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

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

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

[0068] (S3.3) Using the transfer learning method, the printing physical knowledge in the initial working window and process model identified in the S1 stage is extracted, and combined with new sample data, the target working window and process model are transferred and updated 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, operations, and functions:

[0070] Coaxial camera: Settings as follows Figure 2 As shown in (a), Figure 2 (a) is a hardware configuration diagram for systematically optimizing printing process parameters based on a coaxial camera before the start of aerosol printing / when changing printing materials / changing the working environment. The printing sample placed on the platform is transported to the vertical position directly below the lens of the coaxial camera by using a mobile printing platform. Then, the coaxial camera is used to capture images of the printing sample, and the printing quality of the sample is quantitatively analyzed based on formulas (1)-(6).

[0071] Process camera: Settings as follows Figure 2 As shown in (b) in the figure, Figure 2 Figure (b) shows the hardware configuration for real-time evaluation and online quality correction of printed lines during aerosol printing, based on a process camera. This camera is installed near the nozzle of the aerosol 3D printer, maintaining no relative movement to the nozzle during printing, and simultaneously acquiring line images of the printed sample at an angle of approximately 45-60 degrees. Subsequently, the developed quality monitoring model is used to assess the real-time printing status based on the acquired sample images.

[0072] During the printing process, images of the main features of the printed sample are captured at a certain angle to make printing quality decisions.

[0073] Latin Hypercube Sampling Experimental Design (LHS): An efficient experimental design method with key advantages including efficient coverage of the input space, reduced sample size, avoidance of sample concentration, applicability to high-dimensional problems, and global exploration and optimization of the AJP multifactor working parameter space with minimal experimental effort.

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

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

[0076] Because the established GPR model can not only generate predicted values ​​for the output response across the entire workspace, but also provide quantitative information on the uncertainty of the corresponding predictions, the parameter scheme with the lowest uncertainty can be selected from multiple candidate schemes during parameter tuning. This enables autonomous decision-making for quality optimization and effectively improves the corresponding robustness.

[0077] Self-Distilled Vision Transformer 2 (DINOv2) is an unsupervised learning feature extraction model that primarily 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 self-attention mechanism of the Transformer, which gives ViT an advantage in handling long-distance dependencies and capturing global features.

[0079] Incremental clustering algorithms select a few of the most representative points from a large number of experimental points in the design space. This effectively covers the diversity of the design space and represents different potential regions or regional characteristics. The reason for selecting these points is to reduce the number of experiments or computational load, while ensuring that the selected points can represent the characteristics of the entire design space.

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

[0081] The above technical solution will be described in detail with reference to the accompanying drawings:

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

[0083] Step S1. Initial working window and printing process model identification 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 main 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 no need to make assumptions about the distribution of input variables. LHS ensures that each dimension is adequately represented in the input space by uniformly dividing the input variables into intervals and randomly sampling from them, thus providing more accurate results with fewer samples. Its superior sample distribution characteristics make LHS widely used in computer experiments, optimization problems, and uncertainty analysis, especially when dealing with complex models, where it can significantly improve computational efficiency and accuracy.

[0086] This invention uses Latin hypercube sampling to perform reasonable sampling in the process parameter space, where the process parameters include sheath gas flow rate, carrier gas flow rate, and printing speed.

[0087] like Figure 3 As shown, the left figure is the LHS sampling experiment design in a three-dimensional workspace; the right figure is 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 equally probable intervals, and then randomly select a sample point from each interval. This ensures that the value range of each input variable is uniformly covered and can better avoid the concentration of samples, providing a more comprehensive sample distribution. In multidimensional cases, 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 the range of variables: Assume that there are Sk input variables (or factors) in the problem, and each variable has a defined range. For example, the possible range of values ​​for the variable SHGFR (sheath flow rate) is [a1, b1], the possible range of values ​​for the variable CGFR (carrier flow rate) is [a2, b2], the possible range of values ​​for the printing speed is [a3, b3], and so on.

[0090] (2) Divide the range of values ​​of each input variable into equal intervals.

[0091] (3) Random sampling: For each variable, a sample point is randomly selected from each corresponding interval, i.e., a value is selected. In order to ensure uniform coverage, the sample positions in the intervals 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 sample, the sampling points of different values ​​of the same process parameter in each column are randomly arranged. This can avoid the repetition of certain variables in all samples, thereby enhancing the diversity of the sample.

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

[0095] A schematic diagram of print line quality analysis is shown below. Figure 4 As shown, the specific analysis of the printed line quality is as follows:

[0096] (1) Average divergence linewidth ( The result is obtained by spatial averaging the pixel intensity data column by column:

[0097] = (1)

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

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

[0100] (2)

[0101] (3) Average linear density ( Based on the identified printed line edges, calculate the average grayscale intensity of the pixels within their enclosed area, and subtract the influence of background light intensity.

[0102] (3)

[0103] in Let be the discretized average gray intensity of the i-th column. This is the average background intensity (used to counteract the interference of lighting on pixel intensity).

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

[0105] = (4)

[0106] Among them respectively and These represent the deviations of the upper and lower edges of the i-th column relative to the average line, respectively.

[0107] (5) Line discontinuity ): The average number of failed top and bottom edge detections.

[0108] = (5)

[0109] in and These represent the number of failed detections for the top and bottom edges, respectively.

[0110] (6) Overall print quality This is derived by weighted summation of the principal line feature parameters and is used for binary classification of sample quality.

[0111] (6)

[0112] in , and These are the weighting coefficients for line density, edge roughness, and discontinuity.

[0113] Subsequently, since line width and roughness are the main morphological characteristics of printed lines, the line width is quantized ( ) and roughness ( The F-value is calculated using analysis of variance (ANOVA) and compared 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, thus simplifying the model.

[0114] The p-value is a probability value in statistical hypothesis testing, used to measure the probability that the observed data will yield the current or more extreme results given that the null hypothesis is true. The alpha value is a pre-set threshold used to decide whether to reject the null hypothesis. It is usually set to 0.05 (i.e., 5%), meaning there is a 5% probability that the null hypothesis can be rejected incorrectly. If the p-value is less than the alpha value, the result is considered statistically significant, and the null hypothesis is rejected.

[0115] (7)

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

[0117] (8)

[0118] (9)

[0119] in, For the number of groups, Let i be the sample size of the i-th group. Let be the mean of the i-th group. This is the overall mean. Let be the j-th observation in the i-th group, and N be the total sample size.

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

[0121] The overall print quality of the lines is calculated based on formula (6). In formula (6), the weight coefficients can be allocated according to the quality requirements, and the sum of the weights is 1. Then, based on the set judgment threshold of the overall quality of the printed lines, the quality of the printed line samples is divided into two categories: normal and abnormal. Then, the optimal decision boundary for classification is identified using the method of support vector machine (SVM) and kernel function, and is used as the printing working window, as shown in the figure. Figure 5 As shown.

[0122] In this content, the input of the support vector machine 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 that belong to the normal classification can be determined, which serves as the printing working window.

[0123] Specifically, Support Vector Machines (SVMs) are a supervised learning algorithm primarily used for classification and regression tasks. Their core idea is to find an optimal hyperplane that maximizes the margin between different classes of data, thereby improving the model's generalization ability. In the above discussion, since normal and abnormal printed samples are not linearly separable in the operation space, a Gaussian kernel function K is used to map the training features to a high-dimensional space.

[0124] (10)

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

[0126] Simultaneously, slack variables are introduced into the objective function to relax the constraints, thereby making the training data approximately linearly separable in the new space and identifying the nonlinear decision boundary. 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] Where st represents the condition. Let be the normal vector of the decision hyperplane. For bias terms, Let x be the mapping function of the sample in the high-dimensional space. Indicates the distance between the interval and the data points: if the sample If correctly classified by the hyperplane, then ;like If it is located on the wrong side of the hyperplane, then c is a user-defined penalty coefficient used to control training error. This invention uses a uniform grid set as initial parameters and searches for the optimal hyperparameters c and . A schematic diagram of the optimal parameter working space (the optimal working window enclosed by the decision boundary) identified by the source model (Support Vector Machine) is shown below. Figure 5 As shown, the working parameters within the optimal workspace can achieve the best print quality and ensure the printing of normal line samples. Parameters outside the print workspace will cause abnormal print quality.

[0130] The Support Vector Machine (SVM) classifies printed line quality by finding an optimal hyperplane in its working space. First, the goal of SVM is to find a hyperplane that maximizes the distance between the two classes of data points, ensuring that the distance from the nearest point (i.e., support vector) to the hyperplane is maximized while guaranteeing that each data point is correctly classified. For linearly separable data, SVM finds the optimal hyperplane using a simple linear function. However, for non-linearly separable data, SVM uses a kernel function to map the data to a higher-dimensional feature space, making the data linearly separable in this new higher-dimensional space. In this new space, SVM can separate the data by finding a linear hyperplane, even though the data distribution in the original space is non-linear. Common kernel functions include the multinomial kernel and the Gaussian radial basis function (RBF), which help SVM find the optimal separating boundary in the higher-dimensional space. In this way, SVM can handle complex non-linear problems while ensuring classification accuracy.

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

[0132] Gaussian Process Regression (GPR) is a nonparametric regression method based on Bayesian theory. Unlike traditional parametric regression models, it does not build a model by fitting parameters, but rather directly models by defining a latent function space. GPR updates the model through Bayesian inference, that is, it calculates the posterior distribution of the latent function using known data. Because it is a nonparametric model, GPR does not require pre-specifying a specific distribution of the data. It can adapt to complex and unknown relationships and is well-suited for handling high-dimensional and nonlinear data. Given a set of observed data, GPR calculates the conditional probability distribution, predicts the function value at new input points, and provides the uncertainty of the prediction. The specific modeling conclusions are as follows:

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

[0134] (13)

[0135] For the prediction function, It is Gaussian noise. For training the response, X represents all the input data points in the training data, then the predicted point is... With training data They follow a joint Gaussian distribution:

[0136] (14)

[0137] Based on Bayesian inference, the posterior prediction distribution of line features in aerosol printing. Let p represent the probability distribution function:

[0138] (15)

[0139] in, This represents the prediction function given the observed data y. The joint distribution of the training function f. This represents the marginal distribution of the observed data y. Represents the prediction function The joint prior distribution of the training function f. Let f denote the distribution of the observed response y given a training function f.

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

[0141] (16)

[0142] Predictive variance for:

[0143] (17)

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

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

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

[0147] The system drift online monitoring is based on the fusion of Self-Distilled Vision Transformer 2 (DINOv2) and VisionTransformer (ViT) models. The intelligent printing 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 adopts the self-supervised learning DINOv2 model, which optimizes the 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 from different levels or different sources;

[0151] The classification head module is used for fully connected layer connections to transform feature mappings into the final classification result, thus completing the final classification task.

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

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

[0154] (2) The aerosol printer randomly prints according to the designed experimental points, generating a large number of different unlabeled sample data, which include a variety of possible experimental results and variable combinations.

[0155] (3) Collect these generated samples. 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 annotation for subsequent supervised model training.

[0157] The above methods improve the efficiency of constructing the training dataset for the ViT model while ensuring the diversity of data distribution.

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

[0159] DINOv2 is an unsupervised feature extraction model that primarily utilizes self-distillation techniques for unsupervised learning and exhibits impressive performance in visual tasks. The specific training principle of DINOv2 is based on the self-distillation concept, where training is conducted through mutual guidance between a "teacher model" and a "student model." Specifically, it employs a "contrastive learning" method. Unlike traditional contrastive learning, DINOv2's teacher and student are the same model at different time steps, without any manual labels between them. The main principles are as follows:

[0160] (1) Self-distillation process: During training, the student network of the model learns the output features of the teacher network. The parameters of the teacher network are the moving average of the student network. 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 contrastive 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, the images undergo 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 performing self-comparison on each image sample. In this way, it can not only capture global information, but also capture local details, thereby extracting more discriminative features.

[0163] The specific training process for DINOv2 is as follows:

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

[0165] (2) Data augmentation: Two different random data augmentation methods (such as random cropping, color perturbation, random flipping, etc.) are applied to each input image. This allows the model to learn by comparison between different transformations of the same image.

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

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

[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 using a moving average of the student network.

[0169] Student network: The student network learns by comparing its output with that of the teacher network, with the aim 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 student network are trained in parallel. The teacher network is updated using a moving average, while the student network is optimized using a contrastive loss function. The goal of the student network is to replicate the output feature representation of the teacher network as closely as possible. The parameter update formula for the teacher network is:

[0171] (18)

[0172] in, The parameters representing the teacher network, Parameters representing the student network, It is the moving average coefficient, which is usually close to 1.

[0173] (6) Feature Contrast Learning: This method minimizes the distance between different augmented views of the same image and maximizes the distance between different images using contrastive loss. Each pair of views is used to calculate their similarity to train the student network. Contrast Loss The calculation formula is:

[0174] (19)

[0175] in These are student network outputs augmented with different data. The cosine similarity between two feature vectors, where Id represents the index of the sample.

[0176] (7) The training objective is to minimize the total loss function. The loss function includes contrastive loss. and distillation loss The total loss function can be written as:

[0177] (20)

[0178] in , These are the weighting coefficients for comparison loss and distillation loss, and these coefficients are usually adjusted according to the experimental setup.

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

[0180] In summary, DINOv2 learns efficient and discriminative visual features through unsupervised learning and self-distillation, 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 teacher and student networks.

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

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

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

[0184] In this invention, since DINOv2 has already been trained as the backbone module of ViT, and we are ready to continue training the ViT model based on a small number of labeled samples, it is not necessary to retrain the entire ViT model. We only need to freeze the parameters of DINOv2, use its trained features for ViT, avoid training the feature extraction part from scratch, and then perform further supervised learning optimization on the entire model. Specifically, the training steps include the following:

[0185] (1) Freeze 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 supervised learning training. 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: Automatically collected and labeled samples will be used for the supervised learning task. The image features extracted by the DINOv2 model will be used 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 Transformer layers and the classification head. The output of the Transformer is used to make the final classification prediction through a set of classification heads (fully connected layers) to classify the quality of aerosol printed lines and determine whether the printing quality is normal or abnormal. Among them, the line shape of abnormal printing quality is as follows: Figure 7 As shown in (a1-a5), these represent: discontinuous lines, oversprayed lines, low line thickness, high edge roughness, and high diffusion, respectively. The line shape for normal print quality is as follows: Figure 7 As shown in (a6).

[0188] (4) Loss Function and Optimization: Cross-entropy loss is used as the loss function. Specific loss function... The expression is:

[0189] (twenty one)

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

[0191] Optimization algorithms (such as Adam or AdamW) can be used during training to update the weights of the ViT part in order to minimize the loss function and complete the training of the ViT model.

[0192] The pre-trained DINOv2 model serves as the backbone for ViT, and the ViT model is then trained further based on labeled samples, combining the advantages of self-supervised learning and transfer learning. This allows the model to achieve good performance with less labeled data, while reducing computational resources and training time, effectively improving the model's task adaptability and performance. The training process and performance of the ViT model are illustrated in the diagram. Figure 8 As shown. Furthermore, further experiments (see...) Figure 9 The performance of modeling on a small sample dataset was evaluated. Figure 9 (a) in the figure is a comparison of the classification performance of the ViT classification model with the YOLO model and the CNN model in the case of small sample size. Figure 9 (b) shows the robustness of classification performance of the ViT classification model compared with the YOLO and CNN models in the case of small samples. Figure 9(c) in the figure represents the image type used in the robustness comparison, including processing methods such as contrast, blur, saturation, rotation, and noise. Comparative experiments show that the ViT model also exhibits good classification performance and robustness in small sample scenarios. Different forms of noise (such as brightness, contrast, saturation, hue, mirroring, and rotation) were introduced, and the classification performance of various models was compared. Experimental results show that the ViT model trained based on this method outperforms classic CNN convolutional neural networks and YOLO object detection models in small sample generalization ability, and demonstrates superior noise robustness, exhibiting stronger anti-interference capabilities.

[0193] Furthermore, attention heatmaps are visualization tools used to show how a model allocates attention to different parts of the input data when processing a specific task. They use color changes to identify the attention weight of each part of the input. Attention heatmaps help understand which parts of the input the model "focuses" on when making decisions. The heatmap clearly shows which input features or locations the model relies on in determining whether the printing system has shifted. This is crucial for improving the model's interpretability. Integrating the trained ViT model with attention heatmaps for online monitoring helps the AJP printing online evaluation system suppress background noise, enhance the feature response of target areas, accurately locate defects such as low print line density and overspray, and robustly determine whether the printing system has shifted through heatmap analysis, as illustrated in the diagram. Figure 10 As shown, (a1)-(a6) are the original images of the printed line samples, and (b1)-(b6) are the heat maps of the printed line samples corresponding to the original images.

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

[0195] Modal detection and threshold triggering algorithms ensure system stability by monitoring and continuously analyzing the prediction frequency of anomalies in consecutive printed layers in real time.

[0196] First, a statistical modal threshold is calculated based on historical data. This threshold serves as a criterion for identifying whether the system has deviated. Specifically, an anomaly threshold can be set by analyzing the normal fluctuation range of historical data (e.g., 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 historical data, it is determined to be a systematic deviation. When the number of anomalies detected by the system in consecutive judgments exceeds the set modal threshold, 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 it to a normal state. Through this precise anomaly detection and correction triggering mechanism, printing quality problems or production errors caused by system drift can be effectively avoided, thereby ensuring the final accuracy and quality. At the same time, this mechanism reduces human intervention while also improving production efficiency and the system's self-healing capabilities.

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

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

[0199] Figure 11 This diagram illustrates the causes and effects of system drift. To quickly correct the drifting system, the selected experimental points should be able to capture 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) experiments on the target process. Although LHS ensures uniform coverage of the target input feature space, this method may be suboptimal for nonlinear target processes with sufficient original datasets. Therefore, assuming process similarity between the two domains, two rules are used to select representative experimental points for the target process:

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

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

[0202] By jointly analyzing the working parameters of aerosol printing—line characteristics—each point in the printing space... The importance P of the representative experimental point is calculated as follows:

[0203] (twenty two)

[0204] in , These are the dimensions of the input and output spaces, respectively. This represents the point with the largest P-value at the i-th coordinate in different directions of the printing space. The selected representative experimental point is then used to select the next representative experimental point using an iterative method.

[0205] (twenty three)

[0206] in This involves adjusting the parameters, which are set to 1.6 in this invention. This update-selection process will iterate continuously until the required number of representative experimental points are selected. Figure 12 A schematic diagram of the selected representative experimental points.

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

[0208] Because the inherent deposition mechanism of the AJP process typically remains unchanged, certain key printing characteristics and process behaviors will remain consistent even during system drift. System drift usually refers to deviations in the printing process performance under changes in materials, equipment, environment, etc. However, since these changes may not completely destroy the underlying deposition mechanism, certain features will still exhibit similar patterns and trends. Therefore, transfer learning methods can be used to address this challenge. Transfer learning can utilize existing source domain model knowledge, combined with a small number of newly sampled samples, to quickly correct the drifted system. Specifically, the source domain model provides a stable process pattern before drift, and through transfer learning, its knowledge can be transferred to the drifted system. Only a small amount of data from the target domain is needed to adjust the model parameters to adapt to the new operating conditions. This method effectively reduces the need for new data, lowers 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 this invention, the proportion of representative experimental points can be set to 10%-30% of the initial model's experimental points, and the specific proportion can be flexibly adjusted according to actual results and system requirements. The selection of representative experimental points aims to maximize the potential of existing data while avoiding excessive data collection and computational overhead. By accurately selecting representative experimental points, system characteristics and their changing trends can be captured more effectively, providing more accurate data support for subsequent model updates and optimizations. After feature extraction, the newly collected samples can be effectively fused with the initial model. The fused model combines information from the new samples with knowledge from the source model, thereby enabling rapid correction of the drifted system. This process not only accelerates the model's adaptation speed but also improves its accuracy and robustness. Through this method, continuous optimization of complex systems can be effectively achieved using a small amount of new data, allowing the model to maintain good performance and stability even when facing system or environmental changes.

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

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

[0212] This paper employs the Instance Transfer learning method for rapid transfer and correction of the original target working window. The main principle is to adapt the source task dataset to the target task by adjusting the instance weights or directly modifying the source task data. The core idea is to "align" the source and target task data, reducing their differences through certain techniques, thereby allowing the training of the source model to be directly transferred to the target task.

[0213] The main steps of migration:

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

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

[0216] Weighted source task instances: Each source task instance is assigned a weight, so that more important instances (those more closely related to the target task data distribution) have a greater impact on model training. The specific weight allocation formula is as follows:

[0217] (twenty four)

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

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

[0220] The model trained on the source task is then fine-tuned on the target task. By fine-tuning the target task data, the source model can be further adjusted to adapt to the target task. During fine-tuning, the parameters of the source task model are typically optimized using the target task dataset.

[0221] (25)

[0222] in, It is the predicted output of the target model. It is a loss function. It is the true label of the target task. These are the parameters of the original model. This represents the total number of target tasks. After fine-tuning, these parameters will be adapted to the target tasks.

[0223] Instance Transfer aligns the data from the original task with the data from the target task by assigning appropriate weights to instances in the original task. Effective transfer of the original model can be achieved by training on the original model, weighting the source data, and then fine-tuning. See the illustration for the new working window after transfer correction. Figure 13 .

[0224] (2) Fast transfer correction of the original Gaussian process regression model:

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

[0226] (26)

[0227] in and Let these represent the translation vector and scaling vector in the source neighborhood response space, respectively. and These represent the initial model and the target model, respectively. and These represent the input datasets for the initial model and the target model, respectively.

[0228] Furthermore, considering the similarity of the inputs, such as translational similarity and scaling similarity, an input slope / bias correction is used to further adjust for differences in the working range between the input feature spaces.

[0229] (27)

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

[0231] (28)

[0232] Finally, the following equation was optimized and combined with a representative dataset collected from the target domain. The correction vector can be estimated. :

[0233] (29)

[0234] in This represents the error between the actual observed values ​​and the predicted output of the target model. The correction vector is obtained through optimization using a genetic algorithm. See the illustration for the specific new process model after transfer correction. Figure 14 .

[0235] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time quality monitoring and intelligent system drift correction in aerosol 3D printing, characterized in that, include: Numerical sampling of process parameters for aerosol 3D printing is performed to obtain process parameter data. Based on the process parameter data, a printing working window is obtained through a machine learning model, which includes the optimal range of process parameter values. The relationship between process parameters and line quality characteristics is fitted using Gaussian process regression to obtain a process model. Based on the printing working window, corresponding process parameter control data is generated through the process model to control the printing process. During the printing process, image data of the aerosol printing results are acquired, and the image data is evaluated using a drift evaluation model to obtain the drift result of the printing system; wherein the drift evaluation model adopts the ViT model by replacing the backbone network with the DINOv2 model. When the printing system experiences drift, the printing work window and process model are corrected through transfer learning, and the printing process is controlled using the corrected printing work 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 employs a support vector machine (SVM). It acquires the overall print quality corresponding to the process parameter data, and performs supervised learning on the SVM using the process parameter data and the corresponding overall print quality to obtain the optimal hyperplane of the SVM as the printing working window. The overall print quality is the quantitative analysis result of the print results corresponding to the process parameter data.

4. The method according to claim 1, characterized in that, The process of generating the process model includes: The relationship between process parameters and line quality characteristics is modeled using Gaussian process regression. The process parameters include sheath gas velocity, carrier gas velocity, and printing speed, while the line quality characteristics include line width and line thickness. A process model is obtained, in which the process model is fitted to obtain the corresponding process parameters according to the required line quality characteristics. The fitted process parameters are then used to control the printing process.

5. The method according to claim 1, characterized in that, Before evaluating image data using a drift evaluation model, the following steps are also required: Unsupervised learning is performed on the DINOv2 model, and the backbone network of the ViT model is replaced by the unsupervised DINOv2 model. The ViT model is then trained with few-sample supervised learning to obtain the drift evaluation model.

6. The method according to claim 1, characterized in that, The process of evaluating image data includes: Feature extraction is performed on image data using the DINOv2 model to obtain key features. The remaining part of the ViT model is then used to fuse and identify the key features to obtain the final classification result. The final classification result is evaluated to determine whether the printing system has drift.

7. The method according to claim 5, characterized in that, The process of unsupervised learning of the DINOv2 model includes: Acquire sample image data of aerosol printing results, wherein the sample image data is unlabeled; preprocess 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 moving average based on 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, the weighted sum of the loss of contrastive learning and the loss of the teacher network is used as the loss function; thus, the optimized DINOv2 model is obtained.

8. The method according to claim 1, characterized in that, The process of training the replaced ViT model using few-shot supervised learning includes: The model parameters of the DINOv2 model after unsupervised learning are frozen, labeled sample images are obtained, labeled sample images are extracted using the unsupervised DINOv2 model, identification is performed using the remaining part of the ViT model, and the remaining part of the ViT model is optimized to obtain the drift evaluation model.

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

10. A real-time quality monitoring and intelligent system drift correction system for aerosol 3D printing, characterized in that, Used to perform the method described in any one of claims 1-9.

Images