Automatic regulation and control method and system for photocuring 3D printing of ceramic core

Abstract

The invention relates to the field of 3D printing, in particular to an automatic regulation and control method and system for photocuring 3D printing of a ceramic core. The method comprises the following steps that an online detection platform for photocuring 3D printing of the ceramic core is constructed, and a ceramic core image is obtained according to the online detection platform; analyzing a pre-acquired historical defect image to obtain an influence weight of a defect type; building a defect detection model based on the influence weight, and detecting by utilizing the defect detection model according to the ceramic core image to obtain defect data; and constructing an equipment parameter adjustment model, and obtaining optimized equipment parameters by utilizing the equipment parameter adjustment model according to the defect data. The problems that in the prior art, photocuring 3D printing of the ceramic core excessively depends on manual detection and adjustment, and the efficiency and precision of 3D printing are low are solved, and through automatic execution of defect detection and equipment parameter regulation and control, the printing efficiency and the finished product precision of the ceramic core are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of 3D printing, specifically to an automatic control method and system for photopolymerization 3D printing of ceramic cores. Background Technology

[0002] Ceramic cores are indispensable core components in the aerospace field. Their structural integrity and dimensional accuracy directly determine their performance and reliability, thus placing extremely stringent requirements on their quality. With the iteration of additive manufacturing technology, 3D printing technology for ceramic cores has become a key means of preparing high-end ceramic cores due to its ability to directly form complex structures and significantly shorten the research and development and production cycle, providing important support for the precision manufacturing of aerospace components.

[0003] However, during the 3D printing of ceramic cores, defects such as bubbles and cracks are easily generated due to the characteristics of the ceramic powder and the equipment parameters. If these defects are not detected in a timely and accurate manner, they can not only lead to a decline in the mechanical properties and shorten the service life of core components such as turbine blades, but also potentially cause significant safety hazards in the operation of aero engines. Therefore, achieving efficient and accurate detection of defects in ceramic core 3D printing and reasonable control of equipment parameters are the core prerequisites and key constraints for this technology to move towards large-scale industrial application.

[0004] Traditional ceramic core defect detection primarily relies on manual visual inspection, with inspectors judging defects using the naked eye or simple optical equipment. This method has significant drawbacks: not only is the detection efficiency extremely low, making it difficult to adapt to the mass production pace of 3D printing, but it also heavily depends on the experience of the inspectors, resulting in a high degree of subjectivity. Different personnel have vastly different judgment standards, easily leading to false positives and false negatives. Furthermore, the manual adjustment of equipment parameters is inconsistent; different personnel adjust equipment parameters significantly differently for the same defect, resulting in poor ceramic core quality stability and making it difficult to guarantee the efficiency and consistency of the printing process. Therefore, there is an urgent need for a more scientific and efficient automatic control method for ceramic core 3D printing to improve the efficiency of ceramic core 3D printing and reduce the generation of defects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automatic control method and system for photopolymerization 3D printing of ceramic cores, which solves the problem of low efficiency caused by the reliance on manual experience to adjust equipment parameters in existing 3D printing technologies.

[0006] To achieve the above objectives, one aspect of the present invention provides an automatic control method for photopolymer 3D printing of ceramic cores, comprising: constructing an online detection platform for photopolymer 3D printing of ceramic cores, and acquiring ceramic core images based on the online detection platform; analyzing pre-acquired historical defect images to obtain the influence weights of defect types; constructing a defect detection model based on the influence weights, and performing detection using the defect detection model based on the ceramic core images to obtain defect data; constructing an equipment parameter adjustment model, and obtaining optimized equipment parameters based on the defect data using the equipment parameter adjustment model.

[0007] This invention acquires accurate ceramic core images by constructing an online detection platform, providing a high-quality data foundation for subsequent defect detection. This avoids the errors and inefficiencies of manual image acquisition. Historical defect image analysis determines the influence weight of defect types, improving the recognition performance of the defect detection model. By adjusting the equipment parameters based on the defect data, the model optimizes the equipment parameters, achieving automatic and precise adjustment of equipment parameters. This greatly improves the efficiency and accuracy of photopolymerization 3D printing of ceramic cores.

[0008] Optionally, obtaining the ceramic core image according to the online inspection platform includes: obtaining the original image according to the online inspection platform; and extracting the ceramic core image from the original image.

[0009] This invention acquires raw images containing the printing scene through an online inspection platform, and then extracts pure core images from them. This ensures that the image source is real-time and accurate, providing a reliable data foundation for subsequent defect detection, while also eliminating irrelevant background interference, thus improving the efficiency and accuracy of subsequent image processing.

[0010] Optionally, the step of analyzing the pre-acquired historical defect images to obtain the influence weights of defect types includes: dividing the pre-acquired historical defect images into areas of severe quality influence; extracting bubble defect area data and crack defect length data from the historical defect images; calculating the bubble area ratio of the severe quality influence area in each historical defect image using the bubble defect area data; calculating the crack length ratio of the severe quality influence area in each historical defect image using the crack defect length data; and determining the influence weights of bubble defects and crack defects based on the bubble area ratio and the crack length ratio.

[0011] This invention divides areas into zones of severe quality impact, focuses on key quality regions, and then specifically extracts data on bubbles and cracks, calculating their proportions to determine their influence weights. This accurately identifies regions that play a decisive role in quality, avoiding interference from irrelevant areas. By quantifying data, it distinguishes the degree of impact of two types of defects on quality, and the influence weights quantify the importance of different defect identifications, further improving the performance of the defect identification model.

[0012] Optionally, determining the influence weights of bubble defects and crack defects based on the bubble area ratio and the crack length ratio includes: constructing a first matrix using the bubble area ratio and the crack length ratio; normalizing the first matrix to obtain a second matrix; calculating the dispersion of the defect type index based on the second matrix; and calculating the influence weights of bubble defects and crack defects based on the dispersion.

[0013] This invention constructs a first matrix by using the proportion of bubble area and the proportion of crack length, and then normalizes it to obtain a second matrix to eliminate differences in data dimensions. Next, it calculates the dispersion and obtains the influence weights accordingly. The process is progressive, and the data is made more regular through matrixing and normalization. By using the dispersion measure to measure the distribution characteristics of defect indicators, it can accurately distinguish the degree of influence of two types of defects on quality, and improve the scientificity and accuracy of weight calculation.

[0014] Optionally, constructing a defect detection model based on the influence weights includes: introducing a genetic algorithm, using the influence weights to construct the objective function of the genetic algorithm to obtain an optimized genetic algorithm; constructing a defect detection model, and optimizing the hyperparameters of the defect detection model based on the optimized genetic algorithm.

[0015] This invention introduces a genetic algorithm and combines it with influence weights to construct an objective function, resulting in an optimized genetic algorithm. This optimized genetic algorithm is then used to optimize the hyperparameters of the defect detection model. This approach ensures that the algorithm's objective matches the actual impact of defects on quality, while also improving model performance through hyperparameter optimization. This leads to more accurate detection of critical defects, reduced false positives and false negatives, and ultimately, improved performance of the defect detection model.

[0016] Optionally, optimizing the hyperparameters of the defect detection model based on the optimized genetic algorithm includes: using the defect detection model to predict a pre-acquired validation set, and performing statistical analysis on the prediction results to obtain defect recognition performance data; using the defect recognition performance data to calculate the total defect classification loss, the total defect bounding box loss, and the total defect mask loss; using the total defect classification loss, the total defect bounding box loss, and the total defect mask loss to calculate a fitness value according to the objective function; and based on the iteration of the optimized genetic algorithm, determining the optimal hyperparameters of the defect detection model according to the fitness value.

[0017] This invention predicts the validation set and statistically analyzes performance data through a defect detection model, then calculates the sum of three types of losses, calculates the fitness value according to the objective function, and finally determines the optimal hyperparameters through a genetic algorithm, thereby improving the scientific nature and accuracy of hyperparameter calculation.

[0018] Optionally, the optimized genetic algorithm satisfies the following formula: in, To minimize the objective function , Assigning the weight of the impact of bubble defects, To verify the sum of classification losses for all bubble instances in the set, To verify the sum of bounding box losses for all bubble instances in the set, To verify the sum of mask losses for all bubble instances in the set, The influence weight of crack defects, To verify the sum of classification losses for all crack instances in the set, To verify the sum of the bounding box losses for all crack instances in the set, To verify the sum of mask losses for all crack instances in the set.

[0019] The objective function formula of this invention minimizes the objective function by setting the influence weights of bubble and crack defects, respectively, and weighting and summing the classification, bounding box, and mask losses for the two types of defects. This allows the algorithm to more accurately focus on the losses of different defect types, specifically optimize defect detection, and improve the accuracy of the defect recognition model.

[0020] Optionally, the sum of the defect bounding box losses satisfies the following formula: in, To verify all defects in the set The sum of bounding box losses for each instance. To verify all defects in the set Instance collection Index in For the first One defect The x-coordinate of the top left corner of the instance bounding box y-axis ,Width and high The index of the dimensional set that makes up the structure. For the first One defect Instances in dimensions The normalized offset of the predicted bounding box on the top, For the first One defect Instances in dimensions The normalized offset of the true bounding box on the surface.

[0021] The defect bounding box loss formula of this invention calculates the loss piecewise, employing different calculation methods when the difference between the predicted and actual bounding box offsets varies, thus more accurately measuring the bounding box prediction error. This effectively improves the accuracy of defect bounding box localization and enhances the scientific rigor of the defect bounding box loss calculation.

[0022] Optionally, the sum of the defect masking losses satisfies the following formula: in, To verify all defects in the set The sum of masking losses for the instances. To verify all defects in the set Instance collection Index in mask height and mask width The set of pixels that make up the composition For the first One defect In the example, the first The actual mask label of each pixel. For the first One defect In the example, the first The probability of a pixel being predicted as r.

[0023] The defect mask loss formula of this invention is based on cross-entropy loss, calculating the difference between the pixel-level mask prediction and the true label for each defect instance. This accurately measures the accuracy of the defect mask prediction, enabling the model to more finely distinguish between defective and non-defective regions at the pixel level, thus improving the accuracy of defect mask generation.

[0024] In another aspect, the present invention provides an automatic control system for photopolymerization 3D printing of ceramic cores, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the automatic control method for photopolymerization 3D printing of ceramic cores as described in any of the preceding aspects of the present invention.

[0025] The present invention provides an automatic control system for photopolymerization 3D printing of ceramic cores, which is compact in structure, stable in performance, highly integrated and simple in composition. It can stably execute the automatic control method for photopolymerization 3D printing of ceramic cores provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0026] Figure 1This is a flowchart of an automatic control method for photopolymerization 3D printing of ceramic cores according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an automatic control system for photopolymerization 3D printing of ceramic cores according to an embodiment of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0028] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0029] Please see Figure 1 In one embodiment of the present invention, an automatic control method for photopolymerization 3D printing of ceramic cores is provided, which solves the problem of insufficient quality and stability in the prior art of photopolymerization 3D printing of ceramic cores. Figure 1 The method shown includes the following steps: Step S1: Construct an online inspection platform for photopolymerization 3D printing of ceramic cores, and obtain ceramic core images based on the online inspection platform.

[0030] In this embodiment, the hardware setup of the online inspection platform is fundamental to acquiring high-quality ceramic core image data. Regarding camera selection, industrial cameras with a resolution of at least 15 megapixels are preferred. Specifically, the Hikvision MV-CH150-31GM model camera can be selected, as it possesses excellent image capture capabilities, with a frame rate of up to 31fps, enabling timely and clear capture of the ceramic core's shape during rapid printing.

[0031] The configuration of the light source system is crucial. If the ceramic core surface is smooth and highly reflective, a ring-shaped diffused LED light source is an excellent choice. It can evenly illuminate the core from multiple angles, eliminating glare and shadows, resulting in uniform brightness in the image captured by the camera. For cores with rich surface textures and requiring detailed illumination, structured light sources can project specific patterns onto the core surface, allowing for more precise surface information obtained by analyzing pattern deformation. When setting up the light source, multiple experiments are necessary to adjust the distance and angle between the light source and the core to ensure that the light can fully and evenly cover the core surface.

[0032] The mechanical structure must possess high strength and high precision. An aluminum alloy or steel frame is used to ensure it will not shift due to vibration or external forces during long-term operation. High-precision linear guides and stepper motors are installed, with the linear guide straightness error controlled within ±0.05mm / m and the stepper motor positioning accuracy reaching ±0.01mm. This allows for precise control of the camera's movement in the X, Y, and Z axes, enabling accurate positioning and imaging of various parts of the ceramic core.

[0033] Next, software configuration is performed. Image acquisition software is developed using a professional image acquisition library, allowing for the setting and flexible adjustment of parameters such as exposure time and gain, while preprocessing such as grayscale conversion and noise reduction is conducted. Platform control software is developed to achieve unified control of the hardware devices, integrating them with the 3D printing equipment control system. This automatically triggers image acquisition tasks based on the printing progress, ensuring coordination between the inspection and printing processes.

[0034] Before image data acquisition, the online inspection platform must be fully initialized. The camera should be calibrated by capturing standard checkerboard images and using camera calibration algorithms to calculate the camera's intrinsic and extrinsic parameters, ensuring that the captured images accurately reflect the actual size and position of the ceramic core. Simultaneously, light source parameters, such as brightness and color temperature, should be adjusted to achieve optimal illumination. A stable communication connection with the 3D printing equipment should be established, using Ethernet, serial port, or other communication methods, to receive start, pause, and end signals for printing, preparing for image acquisition.

[0035] During the 3D printing process, real-time image acquisition is performed according to a preset detection frequency. Generally, an image is acquired after each layer is printed to achieve real-time monitoring of the printing process.

[0036] The specific steps for obtaining ceramic core images using the online detection platform include the following: Step S101: Obtain the original image according to the online detection platform.

[0037] In this embodiment, the automated control process of photopolymer 3D printing of ceramic cores relies on a pre-built online detection platform for photopolymer 3D printing. Following the rule of acquiring printing data once per layer, after each ceramic core layer is printed, the platform's image acquisition device captures real-time image information including the current ceramic core and its surrounding printing scene, thus obtaining the original image. This original image fully records the surface features of the ceramic core during the current printing process, while also including irrelevant background information.

[0038] Step S102: Extract the ceramic core image from the original image.

[0039] In this embodiment, the original image is first preprocessed by converting it to grayscale to unify the color dimension. Then, Gaussian filtering and other denoising algorithms are used to eliminate noise interference caused by light fluctuations and equipment reflections in the printing environment, laying the foundation for subsequent segmentation. Next, semantic segmentation technology is introduced. Combining the layered structure features of the ceramic core in the 3D printing process, the preset core outline size parameters, and the differences in grayscale values ​​and texture complexity between the ceramic core and the background (such as the printing platform, support structure, etc.) in the original image, a segmentation threshold is established or a dedicated segmentation model is trained to achieve pixel-level differentiation between the ceramic core area and the background area in the image. Subsequently, the preliminary segmentation results are post-processed by using morphological operations (such as dilation and erosion) to eliminate small background noise areas and correcting the segmentation boundary according to the complete geometric shape of the ceramic core, removing residual background pixels at the edges. Finally, a clean image containing only the ceramic core is output.

[0040] Step S2: Analyze the previously acquired historical defect images to obtain the influence weight of the defect type.

[0041] The specific steps for analyzing pre-acquired historical defect images to obtain the influence weight of defect types include the following: Step S201: Divide the historical defect image obtained in advance into a region with severe quality impact.

[0042] In this embodiment, the severe quality impact zone is delineated based on the pre-acquired historical defect images of the ceramic core, taking into account the core's functional requirements, structural characteristics, and historical failure data. Priority is given to identifying areas that play a decisive role in key performance aspects such as core molding accuracy and structural strength. The scope is initially defined in conjunction with the core's application scenario, and then verified and optimized through historical defect failure cases (such as core scrapping due to defects or casting defects), ultimately determining the severe quality impact zone.

[0043] Step S202: Extract bubble defect area data and crack defect length data from the historical defect image.

[0044] In this embodiment, the bubble defect area data includes the overall bubble defect area data and the bubble defect area data in the heavily affected area, and the crack defect data includes the overall crack defect length data and the crack defect length data in the heavily affected area.

[0045] When extracting bubble and crack defect area data from historical defect images, it is necessary to base the data on the pre-annotated information (including bubble and crack category labels and pixel-level masks) obtained using instance segmentation and annotation tools such as LabelMe and CVAT. The process follows a logical sequence: first, extract the basic defect data from the entire image region; for bubbles, count the total number of bubble instances in the entire region, and calculate the area of ​​each bubble and the total bubble area in the entire region using pixel-level masks; for cracks, count the total number of crack instances in the entire region, and calculate the length of each crack and the total crack length in the entire region by combining the pixel-level masks with the true bounding boxes (including the coordinates of the top-left corner, width, and height). Then, based on the pre-defined heavily affected quality areas (with a clearly defined pixel coordinate range in the image), defect instances falling within these areas are selected from the entire region's defect data. The number of bubble instances, the area of ​​a single bubble, and the total bubble area within the heavily affected quality areas are counted. Simultaneously, the number of crack instances, the length of a single crack, and the total crack length within the affected quality areas are also counted. The final result includes data on bubble area and crack length for the entire region, as well as data on bubble area and crack length for the heavily affected area.

[0046] Step S203: Calculate the proportion of bubble area in the severely affected area of ​​each historical defect image using the bubble defect area data.

[0047] In this embodiment, when calculating the proportion of bubble area in the severely affected area of ​​each historical defect image using the bubble defect area data, it is necessary to use the bubble defect area data corresponding to a single historical defect image as the basis, extract the total bubble area of ​​the severely affected area of ​​the image as the numerator from the data, and extract the total bubble area of ​​the entire image as the denominator. By dividing the total bubble area of ​​the severely affected area by the total bubble area of ​​the entire image, the proportion of bubble area in the severely affected area of ​​a single historical defect image is obtained.

[0048] Step S204: Calculate the proportion of crack length in the heavily affected area of ​​each historical defect image using the crack defect length data.

[0049] In this embodiment, when calculating the proportion of crack length in the heavily affected area of ​​each historical defect image using the crack length data, the crack length data corresponding to a single historical defect image is used as the basis. The total crack length in the heavily affected area of ​​the image is extracted from the data as the numerator, and the total crack length of the entire image is extracted as the denominator. The proportion of crack length in the heavily affected area of ​​a single historical defect image is obtained by dividing the total crack length in the heavily affected area by the total crack length of the entire image.

[0050] Step S205: Determine the influence weights of bubble defects and crack defects based on the bubble area ratio and the crack length ratio.

[0051] The determination of the influence weights of bubble defects and crack defects based on the bubble area ratio and the crack length ratio specifically includes the following steps: Step S20501: Construct a first matrix using the bubble area ratio and the crack length ratio.

[0052] In this embodiment, using the bubble area ratio and crack length ratio of the severely affected area calculated for each historical defect image as described above, a first matrix of m rows and 2 columns is constructed, with the total number of historical defect images (i.e., the number of historical samples) as the number of rows and the key ratios corresponding to the two types of defects as the number of columns. Specifically, the element in the i-th row and 1-th column of the first matrix represents the bubble area ratio of the i-th historical defect image, and the element in the i-th row and 2-th column represents the crack length ratio of the i-th historical defect image.

[0053] Step S20502: Normalize the first matrix to obtain the second matrix.

[0054] In this embodiment, when normalizing the first matrix to obtain the second matrix, normalization operations are performed on the two types of defect proportion data based on the structure of the first matrix: First, all elements in the first column of the first matrix, i.e., the bubble area proportion, are extracted. Min-max normalization is used to eliminate the numerical range differences of different image proportions within this column, and the normalized value of the corresponding column is calculated. Then, the same method is used to normalize all elements in the second column of the first matrix, i.e., the crack length proportion column. The resulting second matrix still maintains an m-row, 2-column structure. This normalization process can eliminate the dimensional differences caused by the different fluctuation ranges of the original values ​​of the two types of defect proportions.

[0055] Step S20503: Calculate the dispersion of the defect type index based on the second matrix.

[0056] The dispersion satisfies the following formula: in, For the first The dispersion of various defect type indices Let be the row number of the second matrix. For the second matrix, the first Line number Normalized values ​​of various defect type indicators.

[0057] The above formula quantifies the normalized data of defect indicators. The degree of dispersion in the distribution of historical defect image samples is first calculated for each sample. The Defect normalization value This accounts for the total normalized values ​​of all samples in this class. The proportion (considered as the sample proportion probability) is then summed over all samples using the logarithm of this proportion. The logarithm reflects the uncertainty corresponding to the probability. Finally, through... Scaling, to reduce dispersion The size should reasonably reflect the distribution characteristics: The larger it is, the more likely it is to be the first The more dispersed the distribution of the normalized index of a defect type is among the samples, the more significant the differences in the index among the samples. The smaller the value, the more concentrated the distribution of the indicators.

[0058] Step S20504: Calculate the influence weights of bubble defects and crack defects based on the dispersion.

[0059] The influence weights satisfy the following formula: in, For the first The influence weight of the defect type index For the first The dispersion of various defect type indices For the first The dispersion of various defect type indices.

[0060] The above formula is used to calculate the respective influence weights of bubble defects and crack defects. Its essence is the dispersion of defect type indicators. This transforms the concentration of indicator distribution into a quantitative allocation of influencing weights. Characterizing the first The degree of concentration of various defect indicators The smaller the value, the more concentrated the distribution of indicators. The larger the value, the more consistent and typical the manifestation of this type of defect. Dividing the concentration of a certain type of defect by the sum of the concentrations of the two types of defects yields... This refers to the weight of the impact of this type of defect on quality; the more concentrated the distribution, the higher the weight. The larger the defect type, the higher its proportion in the total weight, which means that its impact on quality is more significant. This is used to quantify and distinguish the relative importance of the two types of defects on quality.

[0061] Step S3: Construct a defect detection model based on the influence weights, and use the defect detection model to detect defects based on the ceramic core image to obtain defect data.

[0062] The specific steps for constructing a defect detection model based on the influence weights are as follows: Step S301: Introduce a genetic algorithm and use the influence weights to construct the objective function of the genetic algorithm to obtain an optimized genetic algorithm.

[0063] The optimized genetic algorithm satisfies the following formula: in, To minimize the objective function , Assigning the weight of the impact of bubble defects, To verify the sum of classification losses for all bubble instances in the set, To verify the sum of bounding box losses for all bubble instances in the set, To verify the sum of mask losses for all bubble instances in the set, The influence weight of crack defects, To verify the sum of classification losses for all crack instances in the set, To verify the sum of the bounding box losses for all crack instances in the set, To verify the sum of mask losses for all crack instances in the set.

[0064] The above formula uses a genetic algorithm to evaluate the objective function. Minimization optimization is performed to balance the performance of bubble defects and crack defects in the detection task, enabling the model to more accurately match the actual impact of the two types of defects on quality. The objective function incorporates the influence weight of bubble defects. Weighting of crack defects The classification loss (measuring the accuracy of category judgment), bounding box loss (measuring the accuracy of target localization), and mask loss (measuring the fineness of instance segmentation) for the two types of defects in the validation set are weighted and summed. and This reflects the relative importance of the two types of defects on quality, therefore, a genetic algorithm is used to minimize... This allows the model to allocate resources based on the impact weight of defects during the optimization process, focusing more on improving the detection accuracy of defect types that have a greater impact on quality, and ultimately achieving a match between overall detection performance and quality requirements.

[0065] The validation set is a subset of the sample data used to train the defect detection model.

[0066] Step S302: Construct a defect detection model and optimize the hyperparameters of the defect detection model based on the optimized genetic algorithm.

[0067] In this embodiment, Mask R-CNN is first selected as the basic architecture of the defect detection model. This is because this framework adds a dedicated mask branch to the core logic of Faster R-CNN, which can simultaneously and efficiently complete three key tasks: target classification, accurate bounding box localization, and instance-level segmentation. The detection requirements for bubbles and cracks perfectly match these three task dimensions. It is necessary not only to accurately determine the defect category (distinguishing between bubbles and cracks), but also to clarify the spatial location range of the defect in the image through bounding boxes, and to rely on pixel-level masks to finely depict the contour morphology of the defect (irregular clusters of bubbles, linear shapes of cracks, etc.). Therefore, the multi-task capability of Mask R-CNN is naturally adapted to the detection scenarios of these two types of defects.

[0068] Historical defect images covering diverse production scenarios were collected to construct a training sample set. The samples needed to include defects generated during different production periods, with varying process parameters, and under different equipment conditions. This ensured that bubble defects encompassed both small, dense types and large, sparse types, while crack defects included narrow, short cracks and wide, long, extending cracks, providing sufficient diversity in the sample set. Next, professional annotation tools were used to perform detailed annotation on each image. For bubble defects, the category label "bubble" was assigned, and the smallest rectangular bounding box surrounding the defect was drawn (recording its position and approximate size). The mask outline of the bubble region was then drawn pixel by pixel to accurately capture the fine morphology of the defect. Similarly, for crack defects, the category label "crack," bounding box, and pixel-level mask were assigned. After completing the annotation of all images, the sample data was obtained. A small portion of the sample data was allocated as a validation set for hyperparameter selection.

[0069] The training data is input into Mask R-CNN for training. Through the iterative process of generating prediction results through forward propagation, calculating the difference between the prediction and the label (loss), and optimizing network parameters (such as convolutional layer weights, activation function parameters, etc.) through backpropagation, the model gradually learns the visual feature patterns of bubbles and cracks, thus obtaining a defect detection model.

[0070] The optimization of the hyperparameters of the defect detection model based on the aforementioned optimized genetic algorithm specifically includes the following steps: Step S30201: The defect detection model is used to predict the pre-acquired validation set, and the prediction results are statistically analyzed to obtain defect identification performance data.

[0071] In this embodiment, the defect detection model is configured with initialized hyperparameters. Each image labeled with a real bubble or crack defect from a pre-divided validation set is input into the defect detection model. For each bubble or crack instance in each image, the model outputs three core prediction results and performs statistical analysis: At the classification level, it outputs the probability that each instance is predicted to be the corresponding defect category; at the bounding box level, it outputs the top-left corner coordinates, width, and height of the predicted bounding box for each instance, and calculates the normalized offsets of these dimensions relative to the real bounding box, while retaining the normalized offsets of the real bounding box; at the mask level, it outputs the probability that each pixel in each instance is predicted to be the corresponding defect category, and extracts the binary label of each pixel in the real mask (1 for defect, 0 otherwise). Finally, it iterates through all instances in the validation set according to bubble and crack categories, summarizing the classification probability of each instance, the normalized offsets of the bounding boxes in each dimension (predicted vs. real), the mask prediction probability of each pixel, and the real label, ultimately forming the defect recognition performance data.

[0072] Step S30202: Calculate the total defect classification loss, the total defect bounding box loss, and the total defect mask loss using the defect identification performance data.

[0073] The total loss of the defect bounding box satisfies the following formula: in, To verify all defects in the set The sum of bounding box losses for each instance. To verify all defects in the set Instance collection Index in For the first One defect The x-coordinate of the top left corner of the instance bounding box y-axis ,Width and high The index of the dimensional set that makes up the structure. For the first One defect Instances in dimensions The normalized offset of the predicted bounding box on the top, For the first One defect Instances in dimensions The normalized offset of the true bounding box on the surface.

[0074] The above formula quantifies the deviation between the defect identification model's predicted defect bounding box and the actual bounding box, guiding the defect identification model to optimize the bounding box localization accuracy, specifically for defects in the validation set. Each instance (belongs to the instance collection) ), Traverse the x-coordinate of the top left corner of its bounding box y-axis ,Width ,high Corresponding dimensions Based on the predicted bounding box normalized offset Normalized offset from the true bounding box absolute difference Whether the deviation is ≤1, the loss is calculated using a piecewise function: when the deviation is small Using quadratic functions This makes the loss in the small deviation region more sensitive to deviation, facilitating fine-grained optimization; when the deviation is large... Using linear functions To avoid excessive loss growth under large deviations, gradient balancing is used to stabilize training. Finally, the losses of each dimension for all defect r instances are summed to obtain the sum of the bounding box losses for this type of defect, reflecting the overall error of the model in the bounding box localization task.

[0075] It is the first in the verification set Class defects ( Corresponding to bubble defects, The set of all instances of the corresponding crack defect is derived from all instances labeled as type r defects selected from the validation set. These instances together constitute ,For example To verify the set of all bubble defect instances, This is a set of all crack defect instances in the verification set.

[0076] The total loss of the defect mask satisfies the following formula: in, To verify all defects in the set The sum of masking losses for the instances. To verify all defects in the set Instance collection Index in mask height and mask width The set of pixels that make up the composition For the first One defect In the example, the first The actual mask label of each pixel. For the first One defect In the example, the first The probability of a pixel being predicted as r.

[0077] The above formula measures the degree of matching between the predicted mask and the real mask when the defect recognition model performs pixel-level segmentation of defects. This is used to evaluate and optimize the model's accuracy in characterizing the defect contour at the pixel level, specifically for defects in the validation set. Each instance (From a defect) Instance collection ), traversing every pixel contained in its mask. (common The nth pixel (determined by the mask height H and width W) is used to apply binary cross-entropy loss to the nth pixel. One defect In the example, the first The real mask label of each pixel (If the pixel is a defect) (If the value is 1, otherwise it is 0) and the model predicts that the pixel belongs to a defect. probability Substitution Calculate the loss for each individual pixel, and finally apply this loss to all pixels and all defects. The defects are obtained by summing the losses of the instances. The sum of mask losses is used to quantify the overall error of the model on pixel-level segmentation tasks.

[0078] The total loss from defect classification satisfies the following formula: To verify centralized defects The sum of classification losses for instances. To verify all defects in the set Instance collection Index in For the first One defect Instance prediction is The probability of.

[0079] The above formula measures the difference between the predicted probability and the true category (belonging to) when the model classifies defect r. The degree of fit of the class is used to assess classification accuracy and address defects in the validation set. Each instance (From a defect) Instance collection ), extract the model prediction for this instance as class probability The probability is transformed into loss using a logarithmic function (the closer the probability is to 1, the smaller the loss, and vice versa). Finally, the loss of all instances is negative and summed to obtain the total classification loss of defect r, thereby quantifying the overall error of the model in the category judgment task.

[0080] Step S30203: Calculate the fitness value based on the objective function using the total defect classification loss, the total defect bounding box loss, and the total defect mask loss.

[0081] In this embodiment, the sum of the defect classification loss, the sum of the defect bounding box loss, and the sum of the defect mask loss for the two types of defects are substituted into the objective function. Since the core objective of optimizing the genetic algorithm is to minimize... (That is, to reduce the overall detection loss of the two types of defects), therefore the objective function value is... Converted to fitness values, usually in the form of The reciprocal or the same The negative correlation values ​​are used as fitness values, so that the fitness value directly reflects the detection performance of the model under the current hyperparameters. The smaller the value, the greater the fitness value, and the better the performance of the defect identification model.

[0082] Step S30204: Based on the iteration of the optimized genetic algorithm, determine the optimal hyperparameters of the defect detection model according to the fitness value.

[0083] In this embodiment, the range of hyperparameters to be optimized is first defined, and reasonable value ranges for key hyperparameters such as learning rate, batch size, anchor box scale, and number of convolutional layer channels are preset. Multiple sets of hyperparameter combinations are randomly generated as the initial population of the genetic algorithm. Each set of combinations corresponds to a set of model hyperparameter configurations. Then, an iterative loop is entered. In each iteration, each set of hyperparameter configurations in the population is substituted into the defect detection model, and the fitness value is calculated through the validation set. Then, based on the fitness value, the core operations of the genetic algorithm, such as selection, crossover, and mutation, are performed to generate a new generation of population. The process of substituting hyperparameters into the model, calculating fitness values, and performing genetic operations is repeated until the iteration stopping condition is met (the preset number of iterations reaches a threshold, or the change in the optimal fitness value of the population for three consecutive generations is less than a set threshold, indicating that the performance tends to be stable). Finally, the set of hyperparameters with the highest fitness value is extracted from all the hyperparameter combinations of all iterations, which is the optimal hyperparameter of the defect detection model.

[0084] The ceramic core image is input into the defect detection model for identification to obtain defect data. The defect data includes the size and location of each defect, including the category label of various defects (to distinguish defect types such as bubbles and cracks), the bounding box coordinates of each defect instance (to locate the position and size of the defect in the image), and pixel-level mask (to accurately depict the shape and range of the defect).

[0085] Step S4: Construct an equipment parameter adjustment model, and use the equipment parameter adjustment model to obtain optimized equipment parameters based on the defect data.

[0086] In this embodiment, when constructing training samples for the equipment parameter adjustment model, historical images of the ceramic core photopolymer 3D printing process are first collected through an online detection platform, and the corresponding equipment parameters (including printing speed, light intensity, resin viscosity, etc.) are recorded. For each historical image, instance segmentation and annotation tools such as LabelMe and CVAT are used to annotate each bubble and crack instance in the image, and the location (record the horizontal and vertical coordinates, width, and height of the upper left corner through a bounding box) and size (based on the number of pixels with a pixel-level mask statistical value of 1, combined with the image resolution to convert to the actual size) of each defect are extracted to form historical defect data corresponding to the historical defect image. Then, for the extracted single set of defect data, based on the knowledge of ceramic core photopolymer 3D printing process and expert experience, the optimal equipment parameters that minimize and uniformly distribute the overall defects are determined to form a parameter adjustment sample set.

[0087] A device parameter adjustment model is obtained by training a fully connected neural network using a parameter tuning sample set. The loss function of the device parameter adjustment model satisfies the following formula: in, For loss function, This is the mean square error loss term. These are constraint terms.

[0088] After substituting into the formula, the loss function satisfies the following formula: in, For the sample size, For the first In the nth sample Real device parameters, For the first In the nth sample Predicted device parameters For the first In the nth sample Real device parameters, For the first The maximum change in each device parameter For the first The maximum adjustment amount of each device parameter.

[0089] The maximum change in equipment parameters is obtained from the statistics of historical adjustment data. The maximum change in each parameter during the normal printing process is the maximum change in equipment parameters. The maximum adjustment of equipment parameters is based on the performance of the parameter adjustment equipment. That is, the predicted value of the parameter is not allowed to be greater than the parameter corresponding to the maximum performance of the equipment, so as to ensure the safety of automatic parameter adjustment.

[0090] The defect data obtained from the current detection is input into the equipment parameter adjustment model, and the equipment parameter adjustment model outputs the corresponding equipment parameters, that is, the equipment parameters are optimized.

[0091] like Figure 2 As shown, in another aspect, the present invention also provides an automatic control system for photopolymerization 3D printing of ceramic cores, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the relevant steps of the relevant embodiments of the automatic control method for photopolymerization 3D printing of ceramic cores of the present invention.

[0092] This invention provides an automated control system for photopolymerization 3D printing of ceramic cores. The functional components can be integrated into a single processing unit, or each component can exist independently, or two or more components can be integrated into one unit. The integrated components can be implemented in hardware or software.

[0093] In summary, this invention solves the problem of poor detection accuracy caused by the inability of traditional machine learning to fully train the detection of defect features. It allows the defect recognition model to better identify defects that have a significant impact on performance based on weights, and thus allows for more targeted parameter adjustments.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An automatic control method for photopolymerization 3D printing of ceramic cores, characterized in that, The method includes: An online inspection platform for photopolymer 3D printing of ceramic cores was constructed, and images of the ceramic cores were acquired based on the online inspection platform. The influence weight of defect type is obtained by analyzing the pre-acquired historical defect images; A defect detection model is constructed based on the influence weights, and defect data is obtained by using the defect detection model to detect defects based on the ceramic core image. A device parameter adjustment model is constructed, and optimized device parameters are obtained using the device parameter adjustment model based on the defect data.

2. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 1, characterized in that, The step of obtaining the ceramic core image based on the online detection platform includes: The original image is obtained using the online detection platform; Extract the ceramic core image from the original image.

3. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 1, characterized in that, The influence weights of defect types obtained by analyzing pre-acquired historical defect images include: The historical defect images acquired in advance are divided into areas of severe quality impact; Extract bubble defect area data and crack defect length data from the historical defect images; The bubble area percentage of the heavily affected area in each of the historical defect images is calculated using the bubble defect area data. The crack length percentage of the heavily affected area in each of the historical defect images is calculated using the crack defect length data. The influence weights of bubble defects and crack defects are determined based on the proportion of bubble area and the proportion of crack length.

4. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 3, characterized in that, The determination of the influence weights of bubble defects and crack defects based on the bubble area ratio and the crack length ratio includes: A first matrix is ​​constructed using the bubble area ratio and the crack length ratio; The first matrix is ​​normalized to obtain the second matrix; The dispersion of the defect type index is calculated based on the second matrix; The influence weights of bubble defects and crack defects are calculated based on the aforementioned dispersion.

5. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 1, characterized in that, The defect detection model constructed based on the influence weights includes: A genetic algorithm is introduced, and the objective function of the genetic algorithm is constructed using the influence weights to obtain an optimized genetic algorithm. A defect detection model is constructed, and the hyperparameters of the defect detection model are optimized based on the optimized genetic algorithm.

6. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 5, characterized in that, The optimization of the hyperparameters of the defect detection model based on the optimized genetic algorithm includes: The defect detection model is used to predict the pre-acquired validation set, and the prediction results are statistically analyzed to obtain defect identification performance data. The defect identification performance data are used to calculate the total defect classification loss, the total defect bounding box loss, and the total defect mask loss, respectively. The fitness value is calculated based on the objective function using the sum of the defect classification loss, the sum of the defect bounding box loss, and the sum of the defect mask loss. Based on the iteration of the optimized genetic algorithm, the optimal hyperparameters of the defect detection model are determined according to the fitness value.

7. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 5, characterized in that, The optimized genetic algorithm satisfies the following formula: in, To minimize the objective function , Assigning the weight of the impact of bubble defects, To verify the sum of classification losses for all bubble instances in the set, To verify the sum of bounding box losses for all bubble instances in the set, To verify the sum of mask losses for all bubble instances in the set, The influence weight of crack defects, To verify the sum of classification losses for all crack instances in the set, To verify the sum of the bounding box losses for all crack instances in the set, To verify the sum of mask losses for all crack instances in the set.

8. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 6, characterized in that, The total loss of the defect bounding box satisfies the following formula: in, To verify all defects in the set The sum of bounding box losses for each instance. To verify all defects in the set Instance collection Index in For the first One defect The x-coordinate of the top left corner of the instance bounding box y-axis ,Width and high The index of the dimensional set that makes up the structure. For the first One defect Instances in dimensions The normalized offset of the predicted bounding box on the top, For the first One defect Instances in dimensions The normalized offset of the true bounding box on the surface.

9. The automatic control method for photopolymerization 3D printing of ceramic cores according to claim 6, characterized in that, The sum of the defect masking losses satisfies the following formula: in, To verify all defects in the set The sum of masking losses for the instances. To verify all defects in the set Instance collection Index in mask height and mask width The set of pixels that make up the composition For the first One defect In the example, the first The actual mask label of each pixel. For the first One defect In the example, the first The probability of a pixel being predicted as r.

10. An automatic control system for photopolymerization 3D printing of ceramic cores, characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute an automatic control method for photopolymerization 3D printing of ceramic cores as described in any one of claims 1 to 9.

Images