Intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration

Abstract

The present invention belongs to the technical field of intelligent additive manufacturing quality control and process optimization, and in particular relates to an intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration. The method specifically comprises: obtaining print line samples and process parameters; obtaining the correlation between the print line samples and the process parameters based on the print line samples and process parameters; obtaining the morphological distribution intervals of the print line samples in different modes based on the correlation; labeling the print line samples based on the morphological distribution intervals to obtain labeled print line samples; constructing a print quality monitoring model, wherein the print quality monitoring model is trained using the labeled print line samples; and performing quality monitoring during the printing process based on the print line morphology collected in real time based on the print quality monitoring model. The present invention can improve printing quality.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent additive manufacturing quality control and process optimization, and in particular relates to an intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration. Background Art

[0002] Fused Deposition Modeling (FDM) plays a key role in the aerospace, biomedical, and automotive industries. As the most mature AM (Additive Manufacturing) method, FDM minimizes material waste and energy consumption, and demonstrates great potential in advanced manufacturing by combining technological diversity with sustainable innovation.

[0003] However, its industrial scalability potential depends critically on precise parameter optimization, particularly nozzle temperature, print speed, and material flow rate, as even minor deviations can lead to defects that compromise part functionality. For example, filament discontinuities caused by nozzle clogging or temperature instability can weaken interlayer adhesion, risking delamination of structural components. Extrusion at inappropriate flow rates or layer heights can create microvoids and voids, reducing tensile strength and fatigue resistance, while rapid cooling or mechanical vibration can exacerbate edge roughness and compromise the dimensional tolerances required for precision components such as seals and bearings. Similarly, uneven extrusion pressure or bed misalignment can lead to irregular line diffusion, distorting delicate geometries such as microfluidic channels or threaded interfaces. These defects collectively degrade surface finish, geometric accuracy, and mechanical integrity, posing obstacles to applications requiring reliability under load or tight tolerances. Therefore, comprehensive analysis of print quality and systematic optimization of FDM process parameters are essential to balance deposition efficiency and defect mitigation, ensuring reproducible production of functional parts that meet industrial standards for strength, accuracy, and surface quality.

[0004] Quality optimization in traditional FDM processes primarily relies on experience-based parameter control. This means that the setting of existing process parameters (nozzle temperature, layer thickness, printing speed, etc.) relies heavily on operator experience. Furthermore, ignoring the impact of system offsets during printing leads to a dynamic mismatch between the material flow characteristics and the molding environment (e.g., hysteresis in non-Newtonian fluid phase transitions). This makes the multi-physics coupling effects (thermal-mechanical-rheological) difficult to accurately model through trial-and-error methods, resulting in fluctuations in interlayer bond strength and residual stress accumulation that can cause warping. Furthermore, traditional offline quality inspection methods (coordinate measurement, CT scanning, etc.) are only applicable in the post-processing stage and cannot capture dynamic molding defects. Delayed defect detection results in material waste. Furthermore, numerical simulation methods, based on finite element models, have high computational complexity and difficulty in real-time coupling to environmental disturbances, making real-time performance impossible to guarantee. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes an intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration. This method organically combines machine learning and deep learning technologies with real-time computer vision to optimize FDM process parameters and realize in-situ anomaly detection.

[0006] The present invention discloses an intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration, which is characterized by comprising:

[0007] Obtain printing line samples and process parameters;

[0008] Based on the printed line samples and the process parameters, obtaining a correlation between the printed line samples and the process parameters;

[0009] Based on the correlation, obtaining the morphological distribution intervals of the printed line samples in different modes;

[0010] Based on the morphological distribution interval, the printed line samples are labeled to obtain labeled printed line samples;

[0011] Constructing a print quality monitoring model, wherein the print quality monitoring model is trained using labeled print line samples;

[0012] Based on the print quality monitoring model, quality monitoring is performed during the printing process according to the print line morphology collected in real time.

[0013] Optionally, obtaining the process parameters includes:

[0014] Get line width and line roughness;

[0015] The process parameters are obtained based on the line width and line roughness using variance analysis.

[0016] Optionally, based on the print line sample and the process parameter, obtaining the correlation between the print line sample and the process parameter includes:

[0017] The printed line samples and process parameters are input into a response model to obtain the correlation between the printed line samples and the process parameters, wherein the response model is constructed by a first-order linear sub-model or a second-order nonlinear sub-model.

[0018] Optionally, the first-order linear sub-model is: ;

[0019] The second-order nonlinear sub-model is: ;

[0020] in, is the response of the first-order linear sub-model, is the response of the second-order nonlinear sub-model, is the intercept term, is the value of the i-th process parameter, is the value of the jth process parameter, is the interaction coefficient between different process parameters, is a random error term that follows a normal distribution.

[0021] Optionally, based on the correlation, obtaining the morphological distribution intervals of the printed line samples in different modes includes:

[0022] Based on the correlation, the contribution values of the process parameters are sorted using the SHAP method to obtain key process parameters;

[0023] A Gaussian mixture model is used to analyze the morphological distribution range of the printed line sample under the key process parameters.

[0024] Optionally, the method for sorting the contribution values of the process parameters using the SHAP method is:

[0025] ;

[0026] in, is the contribution value, N is the set of all features, S is the subset that does not contain feature k, is the predicted value of subset S, is the gain of feature k in set S, is the number of permutations of features within subset S, is the number of permutations of the remaining features, is the total number of permutations of all features.

[0027] Optionally, analyzing the morphological distribution interval of the printed line sample under the key process parameters using a Gaussian mixture model includes:

[0028] Calculating the probability density of the printed line samples;

[0029] Based on the probability density, using an expectation maximization algorithm, obtaining a posterior probability that the printed line sample belongs to a Gaussian cluster by iteratively optimizing a log-likelihood function;

[0030] Based on the posterior probability, the optimal number of clusters is identified using the Bayesian Information Criterion;

[0031] Based on the optimal number of clusters, the morphological distribution interval is obtained.

[0032] Optionally, before labeling the printed line samples, the method further includes: obtaining an optimal operating window using a support vector machine.

[0033] Optionally, the print quality monitoring model is constructed based on a YOLO model, and the YOLO model includes: a backbone network, a feature fusion module, and a detection head module;

[0034] The backbone network is used to extract key features from the image;

[0035] The feature fusion module is used to fuse the key features to obtain a feature map;

[0036] The detection head module is used to predict the feature map and output the category, position and confidence of each detection box.

[0037] Compared with the prior art, the present invention has the following advantages and technical effects:

[0038] 1. This invention uses machine learning to systematically optimize printing parameters before printing begins, when printing materials are changed, or when the working environment changes, thereby improving print quality.

[0039] 2. The present invention is based on online monitoring and correction of printing status based on deep learning and machine vision. At the same time, at the beginning of printing, it analyzes the quality of FDM printing lines in real time, evaluates the printing status, and performs online quality correction, thereby reducing the impact of system offset during the printing process and improving the reliability and robustness of printing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] 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:

[0041] Figure 1 This is a flow chart of a method for intelligent process optimization and real-time quality monitoring of fused deposition modeling based on multi-algorithm collaboration according to an embodiment of the present invention;

[0042] Figure 2 Figure 1 is a diagram of the main hardware configurations of an embodiment of the present invention. (a) is a diagram of the hardware configuration for systematically optimizing printing process parameters based on a coaxial camera before printing begins, when printing materials are changed, or when the working environment changes. (b) is a diagram of the hardware configuration for real-time evaluation of printed line quality and online quality correction based on a process camera during printing.

[0043] Figure 3 Schematic diagram of the CCD experimental design according to an embodiment of the present invention;

[0044] Figure 4 is a schematic diagram of print line quality analysis according to an embodiment of the present invention;

[0045] Figure 5It is a main working parameter print quality model based on RSM in an embodiment of the present invention;

[0046] Figure 6 Schematic diagram of sample type distribution based on GMM clustering according to an embodiment of the present invention;

[0047] Figure 7 1 is a schematic diagram of an optimal working parameter window based on SVM classification according to an embodiment of the present invention;

[0048] Figure 8 2 is a schematic diagram of online sample defect recognition and printing status evaluation based on the YOLO model according to an embodiment of the present invention. (a) is a line feature map obtained by classifying the quality of printed lines based on the trained YOLO model; (b) is an evaluation map of the printing status changing over time based on the trained YOLO model. DETAILED DESCRIPTION

[0049] 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.

[0050] 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.

[0051] This embodiment proposes an intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration, which specifically includes the following steps:

[0052] Obtain printing line samples and process parameters;

[0053] Based on the printed line samples and the process parameters, obtaining a correlation between the printed line samples and the process parameters;

[0054] Based on the correlation, the morphological distribution range of the printed line samples in different modes is obtained;

[0055] Based on the morphological distribution interval, the printed line samples are marked to obtain the marked printed line samples;

[0056] Constructing a print quality monitoring model, wherein the print quality monitoring model is trained using labeled print line samples;

[0057] Based on the print quality monitoring model, quality monitoring is performed during the printing process based on the real-time collected print line morphology.

[0058] Specifically, this embodiment is divided into two parts. The first part is the optimization of printing working parameters system based on machine learning:

[0059] (1) Using the statistical central composite design (CCD) method to maximize the design space coverage with minimal experimental data and achieve efficient exploration of the parameter space;

[0060] (2) The topographic features of the printed line samples generated under the CCD frame were quantified based on the coaxial camera system, and statistically significant process parameters were identified through analysis of variance (ANOVA);

[0061] (3) Use response surface methodology (RSM) to construct an input-output relationship model to guide process optimization;

[0062] (4) SHAP (SHapley Additive exPlanations) is used to quantitatively rank the contributions of significant working parameters and select key variables for subsequent optimization steps;

[0063] (5) The distribution of different patterns of printed lines is analyzed by Gaussian mixture model (GMM), and the optimal operating window is defined using support vector machine (SVM) to achieve robust classification of normal / abnormal lines.

[0064] Part 2: Online monitoring and correction of printing status based on deep learning and machine vision:

[0065] Based on the identified distribution of printed lines, print samples are sampled and automatically annotated. Then, based on these automatically annotated samples, a YOLO deep learning-based object detection model is trained. The trained YOLO object detection model can efficiently and accurately identify abnormal samples and their locations, enabling real-time online evaluation and correction of print status.

[0066] More specifically, the overall workflow is as follows Figure 1 The system includes a dual-camera configuration, system optimization of working parameters based on integrated machine learning and coaxial cameras, and print quality monitoring and real-time correction based on deep learning and process cameras. The specific steps are described as follows:

[0067] Step S1. FDM printing platform hardware setup:

[0068] like Figure 2 As shown, the printer integrates a dual imaging system: a coaxial camera and a process camera.

[0069] (1) Coaxial camera: refers to a camera whose optical path is coaxially aligned with the print head, used to capture a standard vertical angle image of the printed sample. The printed wire sample is immediately transferred to the mobile platform below the coaxial camera (the spatial coordination between the mobile platform and the coaxial camera can avoid the print head blocking the field of view). Then, the real-time digital image acquisition and integrated machine learning algorithm of the coaxial camera are used to achieve rapid autonomous process optimization.

[0070] (2) Process camera: It is deployed adjacent to the print head to monitor the quality of wire deposition under the nozzle in real time. It can capture transient defect characteristics such as subtle deformation, pores or interlayer dislocation of the extruded wire during the printing process and perform real-time in-situ abnormality detection.

[0071] Step S2. System optimization of working parameters based on integrated machine learning and coaxial camera:

[0072] In this part of this embodiment, the input and output of the specific steps of the integrated machine learning algorithm are shown in Table 1:

[0073] Table 1

[0074]

[0075] The specific execution steps of Table 1 above are as follows:

[0076] S2.1. Experimental design using the central composite design (CCD) method:

[0077] Central composite design (CCD) method based on statistics such as Figure 3 As shown, the experimental design points include the following:

[0078] (1) Determine the experimental factors and levels:

[0079] Select independent variables: Identify key factors (such as printing temperature, extrusion speed, layer height, etc.) that have a significant impact on the response variable (such as print edge roughness, line discontinuity, etc.).

[0080] Set levels: Define two extreme levels for each factor (low: -1, high: +1) and determine the center point level (0).

[0081] (2) Constructing experimental points:

[0082] (a) Cubic point:

[0083] Definition: An experimental point at a vertex in the factor space, corresponding to a combination of a full factorial design or a fractional factorial design.

[0084] Coordinates: Each factor takes a level value of ±1, and the remaining factors are at the center level (0).

[0085] Number: For k factors, the number of cubic points is 2 k (full factor) or 2 k-p Fractional factorial design.

[0086] (b) Axial point:

[0087] Definition: Experimental points located on the factor axis and outside the cube are used to estimate the quadratic effect.

[0088] Coordinates: Extend along the axis of each factor in the form of (±α,0,0),(0,±α,0),(0,0,±α), where α determines the extension distance.

[0089] Quantity: Each factor corresponds to 2 axial points, and the total number is 2k (k is the number of factors).

[0090] (c) Center point:

[0091] Definition: All factors take the middle level (0), located at the geometric center of the cube.

[0092] Coordinates: (0,0,…,0).

[0093] Quantity: Usually repeated 3-6 times to estimate experimental error and verify model curvature.

[0094] Furthermore, obtaining process parameters includes:

[0095] Get line width and line roughness;

[0096] Based on the line width and line roughness, the process parameters are obtained using variance analysis method.

[0097] Specifically, S2.2, quantify the printed line quality based on the coaxial camera and identify statistically significant process parameters through analysis of variance (ANOVA);

[0098] The schematic diagram of print line quality analysis is as follows Figure 4 As shown, the details are as follows:

[0099] (1) Average value of scattered line width ( ), calculated by spatially averaging the pixel intensity data by column:

[0100] = (1);

[0101] in Represents the discretized line width of the sth column, and NL is the total number of columns.

[0102] (2) 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:

[0103] (2);

[0104] in is the discretized average grayscale intensity of the tth column, is the average background intensity (used to offset the interference of illumination on pixel intensity).

[0105] (3) Average line edge roughness ( ): Calculate the standard deviation of the actual edge fluctuation based on the reference line (average line):

[0106] = (3);

[0107] Among them are and Respectively represent the deviation of the upper and lower edges of the u-th column relative to the mean line.

[0108] (4) Line discontinuity ( ): Statistical average of the number of upper and lower edge detection failures:

[0109] = (4);

[0110] in and The number of failed upper and lower edge detections, respectively.

[0111] (5) Overall print quality ( ): Obtained by weighted summation of main line feature parameters, used for binary classification of sample quality:

[0112] (5);

[0113] in , and is the weight coefficient of line density, edge roughness and discontinuity.

[0114] 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, then reject the null hypothesis, thereby identifying a statistically significant process parameter:

[0115] (6);

[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] (7);

[0118] (8);

[0119] in, is the number of groups, is the sample size of group r, is the mean of group r, is the overall mean, is the f-th observation value of the r-th group, and TS is the total sample size.

[0120] Furthermore, based on the print line sample and the process parameters, obtaining the correlation between the print line sample and the process parameters includes:

[0121] The printed line samples and process parameters are input into the response model to obtain the correlation between the printed line samples and the process parameters, wherein the response model is constructed by a first-order linear sub-model and a second-order nonlinear sub-model.

[0122] Specifically, S2.3 uses the response surface methodology (RSM) to construct an input-output relationship model to guide the direction of process optimization:

[0123] Response surface methodology is an optimization technique that combines statistics and mathematics. It aims to study the effects of multiple independent variables (factors) on response variables (results) through experimental design, model construction, and data analysis. The core idea is to establish a polynomial regression model (such as a first-order or second-order model) between the independent variables and the response variables through systematic experimental design. The model is used to predict the response value, locate the optimal operating conditions, or analyze the interaction between factors. The model diagram based on the response surface methodology is as follows: Figure 5 shown.

[0124] Specifically, in this embodiment, based on the identified statistically significant process parameters, the response surface methodology is used to establish the statistical correlation between the statistically significant process parameters and the main characteristics of the printed lines. This can be used as a guiding strategy for online correction and optimization of print quality. The specific model is as follows:

[0125] First-order linear model (9);

[0126] Second-order nonlinear model (10);

[0127] in, is the response of the first-order linear sub-model, is the response of the second-order nonlinear sub-model. In fact, the specific experimental statistical analysis results determine whether to use the first-order linear model or the second-order nonlinear model to describe the response. is the intercept term, is the value of the i-th process parameter, is the value of the jth process parameter, is the interaction coefficient between different process parameters, is a random error term that follows a normal distribution.

[0128] Furthermore, based on the correlation, the morphological distribution ranges of the printed line samples in different modes are obtained, including:

[0129] Based on the correlation, the SHAP method is used to sort the contribution values of process parameters and obtain the key process parameters;

[0130] The Gaussian mixture model is used to analyze the morphological distribution range of printed line samples under key process parameters.

[0131] Specifically, in S2.4, SHAP is used to quantitatively rank the contributions of significant working parameters and select key variables for subsequent optimization steps:

[0132] Although statistically significant process parameters can be identified based on ANOVA, if the process is still in a high-dimensional space, further simplification of the parameters is required in order to efficiently identify the optimal process parameter space in the next step.

[0133] SHAP is a model interpretation method based on game theory that quantifies the contribution of each feature to the model's predictions. Its core concept is derived from the Shapley value in cooperative game theory. Specifically, it considers model predictions as the result of multi-feature "cooperation," and the Shapley value fairly distributes the difference between the total predicted value and the baseline value to each feature.

[0134] (11);

[0135] in, is the contribution value, N is the set of all features, S is the subset that does not contain feature k, is the predicted value of subset S, is the gain of feature k in set S, is the number of permutations of features within subset S, is the number of permutations of the remaining features (excluding S and feature k in N), is the total number of permutations of all features.

[0136] Based on the above formula (11), the characteristic value of each individual sample is calculated The size indicates the contribution of the feature in the prediction of the current sample. Then the global contribution of all samples is calculated and sorted from large to small according to the global contribution value to determine the order of importance of the parameters.

[0137] After using SHAP to quantitatively rank the contributions of significant working parameters, key process parameters can be screened according to importance and enter the subsequent optimization stage.

[0138] Furthermore, the Gaussian mixture model is used to analyze the morphological distribution range of printed line samples under key process parameters, including:

[0139] Calculate the probability density of the printed line samples;

[0140] Based on the probability density, the expectation maximization algorithm is used to iteratively optimize the log-likelihood function to obtain the posterior probability that the printed line sample belongs to the Gaussian cluster;

[0141] Based on the posterior probability, the optimal number of clusters is identified using the Bayesian Information Criterion;

[0142] Based on the optimal number of clusters, the morphological distribution interval is obtained.

[0143] Specifically, S2.5 uses a Gaussian mixture model (GMM) to analyze the distribution of different patterns of printed lines, and uses a support vector machine (SVM) to define the optimal operating window to achieve robust classification of normal / abnormal lines:

[0144] As indicated Figure 6 As shown, in order to study the distribution of different forms of printed lines, the line features ( , , ) as the model input, a Gaussian mixture model is used to perform cluster analysis on the printed line morphology. The specific steps are as follows:

[0145] (1) Define the Gaussian mixture model. Assume that the line shape is a mixture of Gaussian distributions, and its probability density function is:

[0146] (12);

[0147] in, To print line sample features ( , , ), is the mixing weight of the qth Gaussian cluster, The mean is , the covariance is Gaussian distribution of sample points The probability density of .

[0148] (2) Using the expectation maximization (EM) algorithm, the log-likelihood function is optimized through iteration :

[0149] Calculation sample The posterior probability of belonging to the qth Gaussian cluster :

[0150] (13);

[0151] Then update the model parameters:

[0152] Blending weights:

[0153] (14);

[0154] Mean vector:

[0155] (15);

[0156] Covariance matrix:

[0157] (16);

[0158] in, is the total number of samples, is the number of samples belonging to cluster q.

[0159] (3) Evaluate models with different K values using the Bayesian Information Criterion (BIC)

[0160] (17);

[0161] in, is the number of model parameters, is the maximum likelihood value of the model.

[0162] By minimizing the BIC function, the optimal number of clusters K of the printed line morphology distribution is identified, and the cluster distribution of the printed line morphology is obtained, as shown in the figure. Figure 6 As shown in the figure, lines of different shapes are distributed in different positions in the workspace, which is conducive to visually analyzing the impact of process parameters on line shape and optimizing the robustness of printing.

[0163] Furthermore, before labeling the printed line samples, the method further includes: obtaining an optimal operating window using a support vector machine.

[0164] Specifically, in addition, based on the comprehensive print quality ( ), the print line quality is divided into normal and abnormal categories, and the support vector machine (SVM) and kernel function method are used to identify the optimal decision boundary of classification statistics and use it as the printing working window, as shown in the figure Figure 7 shown.

[0165] 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 separation between data of different categories, thereby improving the model's generalization capabilities. In this invention patent, since normal print samples and abnormal samples are not linearly separable in the operating space, this embodiment uses a Gaussian kernel function to map the training features into a high-dimensional space:

[0166] (18);

[0167] in, and Two samples from the training dataset, is the Gaussian kernel variance, is the kernel function that maps the data in the low-dimensional space to the high-dimensional space.

[0168] 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.

[0169] (19);

[0170] (20);

[0171] Among them, st means that the following connection is a constraint 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 . Is a custom penalty coefficient used to control the training error. This embodiment uses a uniform grid set as the initial parameter to search for the optimal hyperparameter and The schematic diagram of the identified optimal parameter workspace is as follows: Figure 7As shown, the working parameters within the optimal working space can achieve the best printing quality, while the parameters outside the printing working space will cause abnormal printing quality.

[0172] Furthermore, the print quality monitoring model is constructed based on the YOLO model, which includes: a backbone network, a feature fusion module, and a detection head module;

[0173] Backbone network, used to extract key features in the image, such as edges, textures, shapes and other information;

[0174] The feature fusion module is used to fuse features from different levels or scales to obtain a richer image representation, helping the model better identify objects of different sizes and improve detection accuracy;

[0175] The detection head module is used to predict the feature maps obtained from the backbone network and feature fusion module, output the category, location (bounding box) and confidence of each detection box, and complete the target detection task.

[0176] Specifically, step S3. Print quality monitoring and real-time correction based on deep learning and process camera:

[0177] S3.1. Based on the identified line shape cluster distribution space, collect and automatically label samples

[0178] like Figure 6 As shown in Figure 2, based on the precise division of printed line morphology distribution areas using Gaussian mixture model (GMM) clustering in step S2.5, large-scale random sampling can be performed for different feature areas, and sample labels can be quickly generated using automated annotation technology. This method significantly improves the efficiency of constructing YOLO model training datasets while ensuring data distribution diversity and annotation consistency.

[0179] S3.2. Development of anomaly detection model based on YOLO:

[0180] Convolutional neural networks (CNNs) are at the core of deep learning frameworks for computer vision tasks, with YOLO (Yolo) becoming the preferred architecture due to its exceptional performance in object detection and defect identification. Unlike traditional classification-oriented CNNs, YOLO integrates a backbone network, feature fusion layers, and a detection head module to simultaneously perform defect localization (based on bounding box regression) and defect classification, which is crucial for preventing the propagation of 3D printing faults.

[0181] In the fused deposition modeling (FDM) process monitoring scenario, YOLOv5 demonstrates unique advantages with its optimized balance of inference speed, detection accuracy, and embedded system compatibility. Its architecture uses CSPDarknet53 as the backbone network for feature extraction, implements multi-scale feature fusion through PANet, and performs bounding box prediction based on the anchor box detection head. Key features include: data augmentation to improve model generalization capabilities, adaptive anchor box tuning tailored to FDM defect size, and self-learning bounding box priors to reduce the need for manual calibration. In this embodiment, the total loss function of YOLO is Integrates the three core components required for defect detection: Bounding Box Loss Quantify the difference between the predicted box and the real box coordinates, target existence loss Evaluate the confidence of the target in the prediction box, classification loss Complete defect classification while addressing class imbalance.

[0182] (twenty one);

[0183] (twenty two);

[0184] (twenty three);

[0185] (twenty four);

[0186] in, Integrated target box prediction ( ) and the ground truth box ( ), center distance (ρ), bounding box diagonal (c), and aspect ratio consistency (v), and are weighted by the hyperparameter α; Calculated in S² grid cells, the real existence ( ) and predicted probability ( ) binary cross entropy between ; Using cross entropy on C classes, compare the true labels ( ) and predicted probability ( ). In addition, λ box ,λ object ,λ class is a hyperparameter used to achieve a balance between high detection accuracy and fast inference to suit the FDM real-time object detection task.

[0187] The specific YOLOv5 model training process includes:

[0188] (1) Data preparation and preprocessing:

[0189] Dataset construction: Collect image data containing the target object and divide it into training set, validation set and test set in proportion (the typical ratio is 8:1:1).

[0190] At the same time, the data set is adapted for the FDM process, including printing defects (such as layer misalignment, porosity, etc.), ensuring that the data covers different lighting conditions, viewing angles and defect scales.

[0191] It also uses data augmentation strategies, such as random rotation (±45°), scaling (0.5-1.5x), and color space transformation (HSV adjustment), to simulate complex backgrounds and improve small target detection capabilities.

[0192] (2) Model configuration and initialization:

[0193] Architecture selection: Select the model size (YOLOv5n / s / m / l / x) based on hardware conditions. Smaller models (such as YOLOv5n) are suitable for embedded FDM real-time monitoring.

[0194] Pre-trained weight loading: Use the COCO dataset pre-trained weights to accelerate convergence and freeze the backbone network for transfer learning fine-tuning.

[0195] Hyperparameter setting: setting of key hyperparameters such as initial learning rate, final learning rate decay coefficient, SGD momentum, and weight decay.

[0196] (3) Training process optimization:

[0197] Multi-scale training: Randomly adjust the input image size (e.g., from 320×320 to 640×640) every 10 batches to enhance the scale robustness of the model.

[0198] Loss function dynamic weighting: through the hyperparameter λ box ,λ object ,λ class (Formula 24) sets the task of balancing localization and classification.

[0199] Gradient accumulation and automatic batching: Dynamically adjust the batch size based on GPU memory and support gradient accumulation to stabilize small-batch training.

[0200] (4) Training monitoring and tuning:

[0201] Visualize key metrics: Use TensorBoard or the built-in logging system to monitor metrics such as training loss, validation mAP@0.5, and recall.

[0202] Early stopping mechanism: If the validation set mAP does not improve for 10 consecutive epochs, the training is terminated and rolled back to the optimal weights.

[0203] Model EMA smoothing: Use exponential moving average to update model parameters to reduce the impact of training fluctuations on the final model.

[0204] (5) Model verification and deployment:

[0205] Performance evaluation: mAP@0.5:0.95 (mean average precision at multiple IoU thresholds) and FPS (frame rate) were calculated on the test set to ensure that the FDM real-time monitoring requirements were met.

[0206] Model export: Convert to ONNX or TensorRT format to optimize inference speed for embedded devices.

[0207] S3.3. Online anomaly monitoring based on the YOLO deep learning model:

[0208] CAM-Grad (Class Activation Mapping with Gradient-based Localization) is a visualization method that combines gradient information to enhance model interpretability. Fusion of a trained YOLO model with CAM-Grad for online in-situ inspection facilitates online evaluation of FDM printing, suppresses background noise, enhances the characteristic response of the target area, accurately locates defects such as layer misalignment and insufficient extrusion, and analyzes the causes of defects (such as temperature fluctuations or nozzle clogging) through thermal maps. (See the diagram for details.) Figure 8 shown.

[0209] S3.4. Online correction based on the developed print quality model:

[0210] (1) Modal detection and threshold triggering: as shown Figure 8 As shown, the algorithm continuously analyzes the predicted frequency of abnormal categories in consecutive printing layers. When a certain type of defect exceeds the statistical modal threshold (calculated using historical data), it is determined to be a systematic process deviation (such as abnormal material shrinkage) rather than random noise, triggering a compensation mechanism.

[0211] (2) Parameter adjustment and optimization based on response surface model (RSM):

[0212] Based on the parameter-response relationship (RSM) model established in step S2.3, the parameter importance ranking and interaction priority table from step S2.4, a real-time adjustment strategy is generated, and parameter adjustment amounts are set. This allows for real-time optimization and targeted compensation of defects that occur during the printing process, while also evaluating the effectiveness of the compensation. This method seamlessly combines model guidance with human experience, effectively shortening defect response time while ensuring process safety.

[0213] 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. An intelligent process optimization and real-time quality monitoring method for fused deposition modeling based on multi-algorithm collaboration, characterized by: include: Obtain printed line samples and process parameters; Obtaining the process parameters includes: Get line width and line roughness; Identifying statistically significant process parameters based on quantifying the line width and line roughness using analysis of variance; Inputting the printed line samples and process parameters into a response model to obtain a correlation between the printed line samples and the process parameters; Based on the correlation, obtaining the print line morphology distribution intervals of the print line samples in different modes; Based on the correlation, obtaining the print line morphology distribution intervals of the print line samples in different modes includes: Based on the correlation, the contribution values of the process parameters are sorted using the SHAP method to obtain key process parameters; Analyzing the morphological distribution interval of the printed line sample under the key process parameters using a Gaussian mixture model; Based on the morphological distribution interval, the printed line samples are labeled to obtain labeled printed line samples; Constructing a print quality monitoring model, wherein the print quality monitoring model is trained using labeled print line samples; Based on the print quality monitoring model, quality monitoring is performed during the printing process according to the print line morphology collected in real time.

2. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 1 is characterized in that: The response model is constructed by a first-order linear sub-model or a second-order nonlinear sub-model.

3. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 2 is characterized in that: The first-order linear sub-model is: ; The second-order nonlinear sub-model is: ; in, is the response of the first-order linear sub-model, is the response of the second-order nonlinear sub-model, is the intercept term, is the value of the i-th process parameter, is the value of the jth process parameter, is the interaction coefficient between different process parameters, is a random error term that follows a normal distribution.

4. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 1 is characterized in that: The method for ranking the contribution values of the process parameters using the SHAP method is as follows: ； in, is the contribution value, N is the set of all features, S is the subset that does not contain feature k, is the predicted value of subset S, is the gain of feature k in set S, is the number of permutations of features within subset S, is the number of permutations of the remaining features, is the total number of permutations of all features.

5. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 1 is characterized in that: Analyzing the morphological distribution range of the printed line sample under the key process parameters using a Gaussian mixture model includes: Calculating the probability density of the printed line samples; Based on the probability density, using an expectation maximization algorithm, obtaining a posterior probability that the printed line sample belongs to a Gaussian cluster by iteratively optimizing a log-likelihood function; Based on the posterior probability, the optimal number of clusters of the printed line morphology distribution is identified using the Bayesian Information Criterion; Based on the optimal number of clusters, the morphological distribution interval is obtained.

6. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 1 is characterized in that: Before marking the printed line samples, the method further includes: obtaining an optimal operating window using a support vector machine.

7. The intelligent process optimization and real-time quality monitoring method of fused deposition modeling based on multi-algorithm collaboration according to claim 1 is characterized in that: The print quality monitoring model is constructed based on the YOLO model, which includes: a backbone network, a feature fusion module and a detection head module; The backbone network is used to extract key features from the image; The feature fusion module is used to fuse the key features to obtain a feature map; The detection head module is used to predict the feature map and output the category, position and confidence of each detection box.

Images