An OCT and deep learning-based 3D printing defect identification method and system

Abstract

The application relates to the technical field of image recognition, and discloses a 3D printing defect identification method and system based on OCT and deep learning, the method comprising the following steps: performing coherence gating benchmarking calibration on an OCT acquisition device to obtain a standardized image acquisition benchmark; based on the standardized image acquisition benchmark, performing coherence volume data tomographic reconstruction on real-time acquired OCT interference signals to obtain a coherent tomographic volume image; performing feature contrast enhancement on the coherent tomographic volume image to obtain a feature enhanced image; performing multi-scale feature discrimination on the feature enhanced image to obtain a defect identification result; performing three-dimensional geometric reconstruction on the defect identification result and the coherent tomographic volume image to obtain three-dimensional geometric representation data, and performing spatial configuration analysis on the three-dimensional geometric representation data to obtain a spatial topological atlas; and performing defect root cause tracing on a printer to obtain a process parameter adjustment strategy; the application can improve the efficiency of a 3D printing defect identification method based on OCT and deep learning.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for identifying defects in 3D printing based on OCT and deep learning. Background Technology

[0002] Currently, quality control in metal 3D printing mainly relies on offline inspection or online monitoring technology based on two-dimensional images, which makes it difficult to achieve real-time, accurate identification and closed-loop control of internal defects during the manufacturing process. Offline inspection has a serious lag and cannot prevent the continuous generation of defects; while existing online methods are limited by inspection depth and resolution, often only able to capture surface morphology anomalies, and are insufficient in identifying key internal defects such as porosity, lack of fusion, and microcracks. Furthermore, they are difficult to distinguish defect types and quantify their three-dimensional geometric features, resulting in a lack of precise basis for process adjustments.

[0003] Existing deep learning-based defect identification methods often directly process raw OCT images, neglecting the interference of strong speckle noise, signal attenuation, and system drift on image quality during metal printing, thus limiting the accuracy and robustness of the model. Furthermore, current technologies typically stop at defect detection and classification, failing to further correlate the identification results with the three-dimensional morphology and spatial distribution of defects. They also lack a complete technical chain for tracing the root causes of defects and adaptive control based on multi-dimensional defect features, thus failing to truly achieve integrated intelligent quality assurance encompassing "detection-diagnosis-control." Therefore, improving the efficiency of 3D printing defect identification based on OCT and deep learning has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method and system for identifying defects in 3D printing based on OCT and deep learning, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a 3D printing defect identification method based on OCT and deep learning, comprising:

[0006] S1. Based on the reference OCT image of the pre-printed metal, perform coherent gating benchmark calibration on the OCT acquisition device to obtain the standardized image acquisition benchmark of the OCT acquisition device;

[0007] S2. Based on the standardized image acquisition benchmark, coherent volume data tomography reconstruction is performed on the OCT interference signal acquired in real time during the metal printing process to obtain the coherent tomographic image of the OCT interference signal.

[0008] S3. Perform feature contrast enhancement on the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal;

[0009] S4. Based on a preset deep learning defect recognition network, perform multi-scale feature discrimination on the feature-enhanced image to obtain the defect recognition result of the feature-enhanced image;

[0010] S5. Perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomography image to obtain three-dimensional geometric representation data of the identified defect, and perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect.

[0011] S6. Based on the three-dimensional geometric representation data and the spatial topology map, the root cause of defects in the printer is traced to obtain the process parameter adjustment strategy for the printer.

[0012] In a preferred embodiment, the OCT acquisition device is coherently gated and benchmarked based on a reference OCT image of pre-printed metal to obtain a standardized image acquisition benchmark for the OCT acquisition device, including:

[0013] Based on a preset zero optical path difference reference, an interferometric optical path analysis is performed on the reference OCT image of the pre-printed metal to obtain the optical path deviation result of the reference OCT image.

[0014] Based on the optical path deviation results, the coherent gating of the OCT acquisition device is precisely registered to obtain the optical configuration parameters of the OCT acquisition device.

[0015] Based on the optical property parameters of the pre-printed metal, the electronic parameters of the OCT acquisition device are adapted to obtain the electronic initialization parameters of the OCT acquisition device.

[0016] The optical configuration parameters and the electronic initialization parameters are encapsulated into an operational reference to obtain a standardized image acquisition reference for the OCT acquisition device.

[0017] In a preferred embodiment, the step of performing coherent volume tomography reconstruction on the OCT interferometer signal acquired in real time during the metal printing process based on the standardized image acquisition benchmark to obtain a coherent tomographic image of the OCT interferometer signal includes:

[0018] Based on the standardized image acquisition benchmark, the OCT interference signal acquired in real time during the metal printing process is demodulated to obtain the demodulated spectrum data of the OCT interference signal;

[0019] Based on the scanning position parameters in the standardized image acquisition reference, the demodulated spectrum data is spatially encoded to obtain the encoded spectrum data of the demodulated spectrum data.

[0020] The encoded spectrum data is reconstructed in the frequency domain to obtain the initial three-dimensional volume data of the OCT interference signal;

[0021] The initial three-dimensional volume data is optimized for signal dynamic range to obtain the coherent tomographic volume image of the OCT interferometric signal.

[0022] In a preferred embodiment, the step of enhancing the feature contrast of the coherent tomography image to obtain a feature-enhanced image of the OCT interferometric signal includes:

[0023] Directional diffusion filtering is applied to the coherent tomographic image to obtain a denoised tomographic image of the coherent tomographic image;

[0024] Based on the local gray-level statistical characteristics of the denoised tomographic image, adaptive contrast optimization is performed on the denoised tomographic image to obtain a contrast-enhanced image of the coherent tomographic volume image.

[0025] The contrast-enhanced image is reconstructed by grayscale distribution to obtain a standard grayscale image of the coherent tomography volume image;

[0026] Edge structure sharpening is performed on the standard grayscale image to obtain the feature-enhanced image of the OCT interference signal.

[0027] In a preferred embodiment, the method based on a preset deep learning defect recognition network performs multi-scale feature discrimination on the feature-enhanced image to obtain the defect recognition result of the feature-enhanced image, including:

[0028] The suspected defect regions in the feature-enhanced image are spatially located to obtain the defect candidate regions of the feature-enhanced image;

[0029] Based on a pre-defined deep learning defect recognition network, multi-scale convolutional feature extraction is performed on the defect candidate region to obtain a multi-scale feature vector set of the defect candidate region.

[0030] Based on the deep learning defect recognition network and the multi-scale feature vector set, the defect candidate region is judged for defect category to obtain the preliminary defect classification result of the defect candidate region.

[0031] Based on the decision rules in the deep learning defect recognition network, a comprehensive confidence decision is made on the preliminary defect classification results to obtain the defect recognition results of the feature-enhanced image.

[0032] In a preferred embodiment, the step of performing a comprehensive confidence decision on the preliminary defect classification result based on the decision rules in the deep learning defect recognition network to obtain the defect recognition result of the feature-enhanced image includes:

[0033] The confidence level of the preliminary defect classification results is evaluated to obtain the confidence level evaluation results of the defect candidate regions;

[0034] Based on the spatial location relationship of the defect candidate regions, the spatial distribution logic verification of the preliminary defect classification result is performed to obtain the classification consistency verification result of the defect candidate regions.

[0035] Based on the confidence assessment results and the classification consistency verification results, the preliminary defect classification results are optimized to obtain a credible defect classification result for the defect candidate region.

[0036] The credible defect classification result and the confidence assessment result are structurally encapsulated to obtain the defect recognition result of the feature-enhanced image.

[0037] In a preferred embodiment, the step of performing three-dimensional geometric reconstruction of the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and performing spatial configuration analysis on the three-dimensional geometric representation data to obtain a spatial topological map of the identified defect, includes:

[0038] Based on the spatial location information in the defect identification result, three-dimensional mask data is generated on the coherent tomography image to obtain the defect three-dimensional data block of the defect identification result;

[0039] The boundary contour of the defect three-dimensional data block is extracted to obtain the defect three-dimensional boundary data of the defect identification result;

[0040] Based on the three-dimensional boundary data of the defect, spatial measurement analysis is performed on the identified defect in the defect identification result to obtain the three-dimensional geometric representation data of the identified defect.

[0041] Based on the three-dimensional geometric representation data and spatial location information of the identified defect, spatial connectivity analysis is performed on the identified defect to obtain a spatial topological map of the identified defect.

[0042] In a preferred embodiment, the step of performing spatial connectivity analysis on the identified defect based on its three-dimensional geometric representation data and spatial location information to obtain a spatial topological map of the identified defect includes:

[0043] Based on the three-dimensional geometric representation data and spatial location information of the identified defect, a neighborhood correlation analysis is performed on the identified defect to obtain the adjacency relationship matrix of the identified defect.

[0044] Based on the adjacency matrix, a spatial topology network is constructed for the identified defects to obtain a spatial topology network diagram of the identified defects.

[0045] Graph theory feature analysis is performed on the spatial topology network graph to obtain the topological feature parameters of the spatial topology network graph;

[0046] The topological feature parameters and the spatial topological network graph are integrated into a graph to obtain the spatial topological graph of the identified defect.

[0047] In a preferred embodiment, the step of tracing the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map to obtain a process parameter adjustment strategy for the printer includes:

[0048] Based on the three-dimensional geometric representation data and the defect identification results, the morphological characteristics of the identified defects are analyzed for process correlation to obtain the potential process deviation types associated with the identified defects.

[0049] Based on the spatial distribution of defects in the spatial topology map, the clustering characteristics of the identified defects are used to trace the source of process anomalies, thereby obtaining the printing process anomaly pattern corresponding to the identified defects.

[0050] Based on the potential process deviation type and the abnormal printing process pattern, a rule mapping retrieval is performed on the preset process parameter rule base to obtain process parameter optimization suggestions for the identified defects.

[0051] The process parameter optimization suggestions and the defect identification results are structurally integrated to obtain the process parameter adjustment strategy for the printer.

[0052] To address the aforementioned problems, this invention also provides a 3D printing defect recognition system based on OCT and deep learning, the system comprising:

[0053] The system calibration module is used to perform coherent gating benchmark calibration on the OCT acquisition device based on the reference OCT image of the pre-printed metal, so as to obtain the standardized image acquisition benchmark of the OCT acquisition device.

[0054] The three-dimensional imaging module is used to perform coherent volume data tomography reconstruction on the OCT interference signal acquired in real time during the metal printing process based on the standardized image acquisition benchmark, so as to obtain the coherent tomographic image of the OCT interference signal.

[0055] The image enhancement module is used to enhance the feature contrast of the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal;

[0056] The intelligent recognition module is used to perform multi-scale feature discrimination on the feature-enhanced image based on a preset deep learning defect recognition network, so as to obtain the defect recognition result of the feature-enhanced image;

[0057] The three-dimensional analysis module is used to perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and to perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect.

[0058] The process decision module is used to trace the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map, so as to obtain the process parameter adjustment strategy for the printer.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention significantly improves the imaging quality and signal consistency of the OCT system in a metal printing environment through systematic coherent gating benchmark calibration and image enhancement processing, providing a stable and reliable data foundation for subsequent high-precision identification. The coherent tomographic images generated by this method have high contrast and low noise characteristics, allowing the subtle features of internal defects to be clearly presented. This ensures that the deep learning network can extract more discriminative multi-scale features, greatly improving the accuracy and reliability of defect category discrimination.

[0061] 2. This invention achieves a complete closed loop from defect identification to process decision-making. Through 3D geometric reconstruction and spatial configuration analysis, not only are precise 3D quantitative data of defects obtained, but their spatial distribution and connectivity patterns are also revealed, forming a spatial topological map rich in process information. Based on this, combined with a process knowledge base for root cause tracing, targeted process parameter adjustment strategies can be automatically derived, thereby directly converting online detection results into executable process optimization instructions. This enables real-time quality control and autonomous optimization of the manufacturing process, effectively improving the yield and process stability of metal 3D printing. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a 3D printing defect identification method based on OCT and deep learning, provided in an embodiment of the present invention.

[0063] Figure 2 A functional block diagram of a 3D printing defect recognition system based on OCT and deep learning provided in an embodiment of the present invention;

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0066] This application provides a 3D printing defect identification method based on OCT and deep learning. The execution entity of this OCT and deep learning-based 3D printing defect identification method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the OCT and deep learning-based 3D printing defect identification method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0067] Reference Figure 1 The diagram shown is a flowchart illustrating a 3D printing defect identification method based on OCT and deep learning according to an embodiment of the present invention. In this embodiment, the 3D printing defect identification method based on OCT and deep learning includes:

[0068] S1. Based on the reference OCT image of the pre-printed metal, perform coherent gating benchmark calibration on the OCT acquisition device to obtain the standardized image acquisition benchmark of the OCT acquisition device;

[0069] In this embodiment of the invention, the reference OCT image based on pre-printed metal is used to perform coherent gating benchmark calibration on the OCT acquisition device to obtain a standardized image acquisition benchmark for the OCT acquisition device, including:

[0070] Based on a preset zero optical path difference reference, an interferometric optical path analysis is performed on the reference OCT image of the pre-printed metal to obtain the optical path deviation result of the reference OCT image.

[0071] Based on the optical path deviation results, the coherent gating of the OCT acquisition device is precisely registered to obtain the optical configuration parameters of the OCT acquisition device.

[0072] Based on the optical property parameters of the pre-printed metal, the electronic parameters of the OCT acquisition device are adapted to obtain the electronic initialization parameters of the OCT acquisition device.

[0073] The optical configuration parameters and the electronic initialization parameters are encapsulated into an operational reference to obtain a standardized image acquisition reference for the OCT acquisition device.

[0074] Optical coherence tomography (OCT) is used to obtain a reference image of the pre-printed metal sample. This reference image is a baseline interference signal image acquired under zero optical path difference conditions. Then, the interference optical path data corresponding to each pixel in the reference image is analyzed. Specifically, the phase information of the interference signal is calculated and compared with the ideal zero optical path difference position to obtain the optical path deviation at each position on the entire reference image. These optical path deviations constitute the optical path deviation result.

[0075] The optical system of the optical coherence tomography (OCT) acquisition device is adjusted based on the obtained optical path deviation results. Specifically, the length of the interferometer reference arm and the position of the reflector are adjusted so that the device can compensate for the measured optical path deviation during actual scanning. This adjustment process generates a set of optical system settings, including the reference arm displacement, scan start position, and scan range. These settings are recorded as optical configuration parameters.

[0076] The electronic system of the acquisition device is set according to the optical property parameters of the pre-printed metal material, including reflectivity, scattering coefficient and absorption coefficient. Specifically, the appropriate detector gain, analog front-end amplification factor and digital sampling rate are calculated based on the optical properties of the material, and these electronic adjustment parameters are combined into electronic initialization parameters.

[0077] The optical configuration parameters and electronic initialization parameters are integrated and encapsulated. Specifically, the two sets of parameters are merged according to a preset data structure to generate a standardized configuration file containing complete device initialization information. This file serves as a standardized image acquisition benchmark for device initialization in all subsequent image acquisition processes.

[0078] The beneficial effects include precise calibration of the optical system using a pre-printed metal reference image, eliminating inherent optical path deviations and improving measurement accuracy. Optimization of electronic parameters based on material optical properties ensures optimal matching between signal acquisition sensitivity and signal-to-noise ratio. Encapsulating the optical and electronic configurations into a standard reference file enables rapid and accurate device initialization, guaranteeing consistency and reliability in subsequent image acquisition.

[0079] S2. Based on the standardized image acquisition benchmark, coherent volume data tomography reconstruction is performed on the OCT interference signal acquired in real time during the metal printing process to obtain the coherent tomographic image of the OCT interference signal.

[0080] In this embodiment of the invention, the step of performing coherent volume tomography reconstruction on the OCT interferometric signal acquired in real time during the metal printing process based on the standardized image acquisition benchmark to obtain a coherent tomographic image of the OCT interferometric signal includes:

[0081] Based on the standardized image acquisition benchmark, the OCT interference signal acquired in real time during the metal printing process is demodulated to obtain the demodulated spectrum data of the OCT interference signal;

[0082] Based on the scanning position parameters in the standardized image acquisition reference, the demodulated spectrum data is spatially encoded to obtain the encoded spectrum data of the demodulated spectrum data.

[0083] The encoded spectrum data is reconstructed in the frequency domain to obtain the initial three-dimensional volume data of the OCT interference signal;

[0084] The initial three-dimensional volume data is optimized for signal dynamic range to obtain the coherent tomographic volume image of the OCT interferometric signal.

[0085] The raw interference signals acquired in real time are processed using the center wavelength and bandwidth parameters of the light source set in the standardized image acquisition benchmark. These time-domain interference signals are loaded into the computing unit and transformed from the time dimension to the frequency dimension by performing discrete Fourier transform operations. This transformation process maps the interference intensity information of each sampling point to the amplitude and phase information of different frequency components, and finally generates a structured frequency domain dataset, which is the demodulated spectrum data.

[0086] The demodulated spectrum data is processed based on the scan start point coordinates and scan step size parameters recorded in the standardized image acquisition reference file. A three-dimensional spatial coordinate label is assigned to the signal corresponding to each frequency point in the demodulated spectrum data. This coordinate label is calculated by the scan order and reference position parameters, so that each unit in the spectrum data clearly corresponds to a physical location in the printing area. The data processed by this mapping relationship forms the coded spectrum data.

[0087] Image reconstruction processing is performed on the encoded spectrum data. All frequency point data with spatial coordinate labels are arranged and combined in space according to their corresponding coordinate positions. The spectrum information at each position is converted back to the intensity distribution in the depth direction through inverse Fourier transform. This transformation is performed synchronously on all scan points, and finally a complete three-dimensional matrix is ​​generated. This matrix is ​​the initial three-dimensional volume data, and each element value represents the signal intensity of the corresponding spatial point.

[0088] The initial 3D volume data undergoes signal optimization processing. The entire 3D matrix is ​​scanned to find its maximum and minimum signal intensity values, thereby determining the dynamic range of the original signal. Then, a logarithmic compression function is used to perform nonlinear scaling on all signal intensity values. This processing converts the original large dynamic range signal into a smaller range suitable for display and human observation, while preserving the detail visibility of weak signals. The 3D signal matrix after this optimization processing finally forms a coherent tomographic image that can be directly analyzed and displayed.

[0089] The beneficial effects include the accurate extraction of depth information from interference signals through standardized spectral demodulation; the use of reference position parameters to achieve precise correspondence between spectral data and physical space, ensuring the accuracy of 3D reconstruction; and the reconstruction of complete 3D volumetric data reflecting the internal structure through inverse transformation. Dynamic range optimization is employed to enhance image contrast and detail visibility, ultimately yielding clear and reliable tomographic images.

[0090] S3. Perform feature contrast enhancement on the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal;

[0091] In this embodiment of the invention, the step of enhancing the feature contrast of the coherent tomography image to obtain a feature-enhanced image of the OCT interferometric signal includes:

[0092] Directional diffusion filtering is applied to the coherent tomographic image to obtain a denoised tomographic image of the coherent tomographic image;

[0093] Based on the local gray-level statistical characteristics of the denoised tomographic image, adaptive contrast optimization is performed on the denoised tomographic image to obtain a contrast-enhanced image of the coherent tomographic volume image.

[0094] The contrast-enhanced image is reconstructed by grayscale distribution to obtain a standard grayscale image of the coherent tomography volume image;

[0095] Edge structure sharpening is performed on the standard grayscale image to obtain the feature-enhanced image of the OCT interference signal.

[0096] Directional diffusion filtering (DDP) selectively smooths each pixel in a coherent tomographic image based on the direction of the gray-level gradient in its surrounding area. It calculates the gradient vector for each pixel, indicating the direction of the most dramatic gray-level change. Then, a weighted average of the gray values ​​is performed along a path perpendicular to the gradient direction. This averaging operation is only performed in this vertical direction, thus smoothing noise while avoiding blurring across image edges. The entire process iterates through all pixels in the image, progressively eliminating speckle noise and random noise by performing the directional smoothing operation. The final result is an image that is smooth in uniform areas while remaining sharp at edges; this is the denoised tomographic image.

[0097] The adaptive contrast optimization process is based on the local gray-level statistical characteristics of the denoised tomographic image. The image is divided into multiple overlapping small regions. Within each small region, the mean and standard deviation of the gray-level values ​​are calculated. Then, the gray-level mapping curve for each pixel within that region is dynamically adjusted based on the local mean and standard deviation. For regions with low local contrast, a steeper mapping curve is used to stretch the gray-level differences, while for regions with already high local contrast, a relatively gentler mapping curve is used to avoid over-enhancement. This adjustment is achieved by constructing a gray-level transformation function for each pixel based on the statistical characteristics of its local region. The specific shape of the transformation function depends on the gray-level distribution characteristics of that local region. After this region-by-region adaptive mapping transformation, the contrast of different regions is evenly enhanced, ultimately generating an image with good visual contrast both overall and locally—this is the contrast-enhanced image.

[0098] The grayscale distribution reconstruction process involves readjusting the global grayscale distribution of the contrast-enhanced image. This involves statistically analyzing the grayscale value distribution of all pixels in the entire contrast-enhanced image, generating a grayscale histogram, and analyzing its distribution pattern to identify the main concentrated areas and the distribution characteristics at both ends. Then, based on a preset standard grayscale distribution template—which defines the ideal distribution pattern of image grayscale values—the grayscale value of each pixel is adjusted by establishing a mapping relationship from the current grayscale distribution to the standard grayscale distribution. This mapping relationship is achieved through a grayscale value redistribution function, which redistributes the original grayscale values ​​to new grayscale levels according to the requirements of the standard distribution. After this global grayscale remapping, the entire image exhibits standardized grayscale distribution characteristics, ultimately generating an image with a normalized grayscale distribution, which is the standard grayscale image.

[0099] Edge sharpening enhances edge features in a standard grayscale image. Specifically, it identifies edge locations by detecting regions with significant grayscale changes. A differential operator calculates the rate of grayscale change for each pixel within its neighborhood, determining the edge's intensity and direction based on the magnitude and direction of this rate. For identified edge regions, the grayscale difference between the two sides is enhanced along the edge's normal direction by applying opposite grayscale adjustments, making one side brighter and the other darker. Meanwhile, non-edge regions are kept relatively unchanged in grayscale to avoid introducing noise. This edge enhancement operation iterates through all identified edge regions in the image, adjusting the grayscale values ​​of pixels on either side of the edge to enhance its visual sharpness. After this process, structural boundaries and detailed features in the image are significantly highlighted, resulting in a sharp, detailed image—the feature-enhanced image.

[0100] The beneficial effects include effectively suppressing image noise while preserving important edge structures through directional filtering, significantly improving the image signal-to-noise ratio. Adaptive enhancement of contrast in different regions is achieved using local statistical properties, ensuring clear visibility of image details under varying lighting conditions. Standardization and reconstruction of the entire image's grayscale distribution result in consistent and standardized brightness levels. Edge sharpening enhances structural contours and detailed features, ultimately yielding a feature-enhanced image with rich details and sharp boundaries.

[0101] S4. Based on a preset deep learning defect recognition network, perform multi-scale feature discrimination on the feature-enhanced image to obtain the defect recognition result of the feature-enhanced image;

[0102] In this embodiment of the invention, the method of performing multi-scale feature discrimination on the feature-enhanced image based on a preset deep learning defect recognition network to obtain the defect recognition result of the feature-enhanced image includes:

[0103] The suspected defect regions in the feature-enhanced image are spatially located to obtain the defect candidate regions of the feature-enhanced image;

[0104] Based on a pre-defined deep learning defect recognition network, multi-scale convolutional feature extraction is performed on the defect candidate region to obtain a multi-scale feature vector set of the defect candidate region.

[0105] Based on the deep learning defect recognition network and the multi-scale feature vector set, the defect candidate region is judged for defect category to obtain the preliminary defect classification result of the defect candidate region.

[0106] Based on the decision rules in the deep learning defect recognition network, a comprehensive confidence decision is made on the preliminary defect classification results to obtain the defect recognition results of the feature-enhanced image.

[0107] The step of performing a comprehensive confidence decision on the preliminary defect classification result based on the decision rules in the deep learning defect recognition network to obtain the defect recognition result of the feature-enhanced image includes:

[0108] The confidence level of the preliminary defect classification results is evaluated to obtain the confidence level evaluation result of the defect candidate region, wherein the confidence level is calculated using the following formula:

[0109] ;

[0110] In the formula, This indicates that the defect candidate region is determined to be the first... Confidence score for class defects This indicates that the deep learning defect recognition network is effective for the first... The original probability of the class defect output. Preset feature consistency weight coefficient, The multi-scale feature vector representing the defect candidate region is in the first... Projection vector along the defect direction, This indicates all items initially classified as number one. The mean of the regional feature vectors of the class defect, Indicates the first Standard deviation of the distribution of class defect characteristics This represents the spatial saliency score of the candidate defect region. This represents the preset volume ratio enhancement coefficient. This represents the volume estimate of the candidate defect region. This represents the sum of the volume estimates of the candidate defect regions within the current analysis field of view. Indicates category, Represents an exponential function. Represents the maximum value function;

[0111] Based on the spatial location relationship of the defect candidate regions, the spatial distribution logic verification of the preliminary defect classification result is performed to obtain the classification consistency verification result of the defect candidate regions.

[0112] Based on the confidence assessment results and the classification consistency verification results, the preliminary defect classification results are optimized to obtain a credible defect classification result for the defect candidate region.

[0113] The credible defect classification result and the confidence assessment result are structurally encapsulated to obtain the defect recognition result of the feature-enhanced image.

[0114] In feature-enhanced images, local regions with significant differences from the background are identified using a gray-level thresholding method. A gray-level difference threshold is pre-set. The gray-level values ​​of all consecutive pixels in the image are compared with the average gray-level value of the surrounding background region. When the difference exceeds the set threshold, the pixel is marked as a suspected point. Subsequently, all spatially adjacent suspected points are merged into connected regions, each representing an independent suspected defect region. The minimum bounding rectangle of these regions is calculated, which defines the specific location and extent of the suspected defect in the image. Finally, a series of rectangular regions with position coordinates and size information are generated; these rectangular regions are the defect candidate regions.

[0115] Image data for each defect candidate region is input into a pre-trained deep learning defect recognition network. This network contains multiple parallel feature extraction pathways with convolutional kernels of different sizes. Each pathway performs convolution operations on the input region image to capture feature patterns at different spatial scales. Small-scale convolutional kernel pathways extract subtle textures and point features, medium-scale convolutional kernel pathways extract edge and local shape features, and large-scale convolutional kernel pathways extract overall contour and structural features. The feature maps extracted by each pathway are reduced in dimensionality through pooling operations, flattened, and concatenated into a comprehensive feature description sequence. Each defect candidate region generates a fixed-dimensional numerical vector after this series of processing. The set of all these vectors constitutes a multi-scale feature vector set.

[0116] The classification module of a deep learning defect recognition network is used to classify each candidate defect region. This module receives multi-scale feature vectors from the feature extraction stage, performs non-linear combinations and transformations on these features through fully connected layers, and outputs the probability value of the region belonging to each predefined defect category through a classifier. During training, the network has learned the discrimination boundaries of different defect categories in the feature space. The feature vector of the region to be tested is then mapped to this feature space, and the most likely defect type is determined based on the discrimination region it falls into. Simultaneously, the probability value corresponding to this type is output as a confidence score. This process generates a classification result containing a category label and the original confidence score for each candidate defect region. The set of these results forms the preliminary defect classification result.

[0117] The preliminary defect classification results are validated and integrated using a pre-defined decision rule in a deep learning defect recognition network. This rule comprehensively considers multiple factors, including the original classification confidence of each candidate region, the similarity between its feature vector and the feature distribution of similar samples, the saliency intensity within the region, and the relative volume proportion of the region. A comprehensive score for each candidate region based on these factors is calculated and compared with a pre-defined acceptance threshold. Regions with scores exceeding the threshold retain their classification results, while regions with scores below the threshold are excluded or marked for further review. Simultaneously, multiple spatially adjacent regions with consistent classification results are merged, ultimately generating a list containing complete information such as defect type, precise location, and confidence score. This list represents the defect recognition results of the feature-enhanced image.

[0118] The confidence assessment process involves calculating a comprehensive confidence score for each candidate defect region based on its initial defect classification results. This score integrates four aspects: the region's original classification probability, the similarity between its multi-scale feature vector and the average feature vector of similar regions, the region's spatial saliency score, and its relative volume proportion. Similarity is quantified by calculating the exponential decay function of the Euclidean distance between the region's feature vector and the mean feature vector of similar regions; closer distances result in higher similarity. The spatial saliency score is determined based on the ratio of the region's internal gradient range to its mean. The final comprehensive score is obtained by multiplying the original probability, the similarity weighting term, the saliency normalization term, and the volume enhancement term. Each region generates a numerical score between zero and one for its assigned category, and these scores for all regions constitute the confidence assessment result.

[0119] The parameters in the formula have a clear processing foundation. The raw probability output by the deep learning defect recognition network for each candidate region directly originates from the category probability distribution generated by the network classification module after calculating the multi-scale feature vector set. The projection vector of the multi-scale feature vector onto the specific category direction is obtained by performing a dot product operation between the complete feature vector of the region and the discriminative direction of the category in the feature space. The mean of the feature vectors of all regions initially classified into the same category and the standard deviation of the feature distribution of that category are obtained by online statistical analysis of the feature vectors of all regions of that category in the current analysis batch and calculating their arithmetic mean and dispersion. The spatial saliency score is a proportion obtained by performing specific algebraic operations on the maximum, minimum, and average gradient values ​​extracted from the feature-enhanced image of the candidate region. The volume estimate is calculated based on the size information of the region's bounding box and the integral result of the optical coherence tomography signal intensity along the depth direction.

[0120] The significance of this formula lies in constructing a comprehensive decision-making mechanism to evaluate the final credibility of classifying each candidate region as a specific defect category. This mechanism uses the original classification probability output by the deep learning network as the base credibility, while introducing a feature consistency verification term to evaluate the degree of agreement between the region's features and the overall distribution of features in similar regions; a higher degree of agreement results in a greater positive adjustment contribution. A spatial saliency term is introduced to reflect the structural prominence of the region in the image; higher saliency results in a greater positive adjustment contribution. A volume proportion term is introduced to consider the relative importance of the defect's physical size; a larger volume proportion results in a smaller positive enhancement contribution. Finally, the original probability is calibrated by weighting and integrating these adjustment factors to generate a more stable final confidence score that conforms to actual physical meaning. This score is used for subsequent decision filtering and result integration.

[0121] The spatial distribution logic verification process involves analyzing whether there are any contradictions between the physical spatial relationships of all candidate defect regions and their initial classification labels. It examines multiple regions that are adjacent to each other in three-dimensional space. If these regions are initially classified as different defect types, a judgment is made based on a pre-defined defect spatial distribution knowledge base. This knowledge base defines the reasonableness rules for different defect types to appear simultaneously in adjacent positions in actual printed parts. For adjacent region combinations that violate these rules, the system marks them as having spatial logic conflicts. After traversing all regions, a list is generated recording whether each region has conflicts and information about its conflicting adjacent regions. This list represents the classification consistency verification result.

[0122] The confidence optimization process involves revising and confirming the initial classification results by combining the confidence assessment results with the classification consistency verification results. An acceptance threshold is set based on the confidence assessment score; regions with scores below this threshold are temporarily shelved. For regions with spatial logical conflicts, the classification result of the region with the highest confidence assessment score within each group is retained, while the classification results of other conflicting regions within the group are either modified to match the region's category or marked as pending. All shelved low-confidence regions are re-examined; if they are spatially adjacent to a confirmed high-confidence region and belong to the same category, their category is confirmed as the same. After this series of iterative adjustments based on confidence and spatial logic, each defect candidate region obtains an optimized and confirmed defect type label. This set of labels constitutes the credible defect classification result.

[0123] The structured encapsulation process involves organizing the detailed information from the credible defect classification results and confidence assessment results into a standard-format output file. A data entry is created for each defect candidate region, containing the region's three-dimensional spatial coordinates, optimized defect type label, corresponding comprehensive confidence score, region volume estimate, and spatial saliency score. All data entries are sorted according to their spatial location or confidence score, and a header is added indicating metadata such as the total analysis field of view, total number of defects, and statistical summaries of various defect types. The resulting data file, containing complete defect identification information, is the defect identification result of the feature-enhanced image, which can be directly used for subsequent quality report generation or process feedback control.

[0124] The beneficial effects include: calculating a comprehensive confidence score by fusing multi-dimensional features, providing a quantitative basis for the reliability of each defect determination; using spatial distribution logic rules to verify the preliminary classification results, effectively identifying and eliminating physically unreasonable contradictions in defect coexistence; combining confidence scores and spatial logic to collaboratively optimize the classification results, improving the accuracy and consistency of the final defect type determination; and encapsulating the optimized classification results and confidence data into a standard structure file to form a complete, reliable, and directly applicable defect identification report.

[0125] S5. Perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomography image to obtain three-dimensional geometric representation data of the identified defect, and perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect.

[0126] In this embodiment of the invention, the step of performing three-dimensional geometric reconstruction on the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and performing spatial configuration analysis on the three-dimensional geometric representation data to obtain a spatial topological map of the identified defect, includes:

[0127] Based on the spatial location information in the defect identification result, three-dimensional mask data is generated on the coherent tomography image to obtain the defect three-dimensional data block of the defect identification result;

[0128] The boundary contour of the defect three-dimensional data block is extracted to obtain the defect three-dimensional boundary data of the defect identification result;

[0129] Based on the three-dimensional boundary data of the defect, spatial measurement analysis is performed on the identified defect in the defect identification result to obtain the three-dimensional geometric representation data of the identified defect.

[0130] Based on the three-dimensional geometric representation data and spatial location information of the identified defect, spatial connectivity analysis is performed on the identified defect to obtain a spatial topological map of the identified defect.

[0131] The step of performing spatial connectivity analysis on the identified defect based on its three-dimensional geometric representation data and spatial location information to obtain a spatial topological map of the identified defect includes:

[0132] Based on the three-dimensional geometric representation data and spatial location information of the identified defect, a neighborhood correlation analysis is performed on the identified defect to obtain the adjacency relationship matrix of the identified defect.

[0133] Based on the adjacency matrix, a spatial topology network is constructed for the identified defects to obtain a spatial topology network diagram of the identified defects.

[0134] Graph theory feature analysis is performed on the spatial topology network graph to obtain the topological feature parameters of the spatial topology network graph;

[0135] The topological feature parameters and the spatial topological network graph are integrated into a graph to obtain the spatial topological graph of the identified defect.

[0136] Using the spatial coordinates of each confirmed defect in the defect identification results, these locations are marked in the 3D data matrix corresponding to the coherent tomography volumetric image. Specifically, the 3D bounding box coordinates of each defect region are mapped to the index of the volumetric data matrix, and all voxels within these index ranges are extracted. To ensure integrity, a small-scale morphological dilation is performed to cover blurred boundary areas. The set of all extracted voxels forms an independent 3D binary matrix, where defect location values ​​are true and background location values ​​are false. This 3D binary matrix is ​​the defect 3D data block, which completely preserves the 3D morphology and spatial distribution of each defect in the original volumetric data.

[0137] A 3D edge detection operation is performed on the defect 3D data block to obtain its surface contour. For each voxel marked as true, the process iterates through the 3D data block, checking if all six of its directly adjacent voxels are true. If at least one adjacent voxel is false, the voxel is determined to be a boundary point. The 3D coordinates of all determined boundary points are collected and grouped according to their respective independent connected regions, with each connected region corresponding to an independent defect entity. These grouped 3D coordinate sets accurately describe the spatial position of each defect surface, forming the defect 3D boundary data.

[0138] Based on the 3D boundary data of the defects, a series of geometric measurement calculations are performed. For each independent defect entity, the dimensions of its smallest circumscribed cuboid enclosing all boundary points are calculated, namely its length, width, and height. The arithmetic mean of the coordinates of all boundary points of the defect entity is calculated to obtain its centroid position. The distribution of boundary points relative to the centroid is calculated to estimate its approximate volume. Simultaneously, the extent of the boundary point set along the three principal axes is analyzed to determine its shape orientation. The set of measurement results, including dimensions, position, volume, and orientation, generated by these calculations constitutes the 3D geometric representation data.

[0139] For each defect entity, a spatial neighborhood check is performed to establish associations. The minimum Euclidean distance between the 3D boundary datasets of any two defect entities is calculated. If this distance is less than a preset neighborhood threshold, the two defects are considered spatially adjacent. All defect entity pairs are traversed, and defect pairs with adjacency relationships are recorded. A matrix structure is created to record these relationships. In this matrix, rows and columns correspond to each defect entity. If two defects are adjacent, the intersection of the corresponding row and column is marked as true; otherwise, it is marked as false. This matrix is ​​the adjacency matrix, which clearly expresses the spatial proximity relationships between defects.

[0140] A network graph with defective entities as nodes is constructed using an adjacency matrix. Each defective entity is abstracted as a node in the graph, and its attributes include its three-dimensional geometric representation data. According to the adjacency matrix, if the matrix positions corresponding to two defective entities are true, an undirected edge is added between the two nodes representing them. This generates a network structure graph consisting of nodes and edges, which visually displays the spatial connections between all identified defects. This network structure is called a spatial topology network graph.

[0141] Graph theory metrics are calculated for spatial topological network graphs. The overall structure of the graph is analyzed, and the number of connected components (i.e., the number of unconnected subgraphs) and the average path length (i.e., the average number of edges traversed in the shortest path between any two nodes) are calculated. The clustering coefficient is calculated, which measures the average degree of edge connectivity within a node's neighborhood. The degree of each node (i.e., the number of edges directly connected to it) is also calculated. These calculated numerical values ​​reflecting the overall structure and local connectivity characteristics of the graph constitute the topological feature parameters.

[0142] Integrating and encapsulating the structural information and topological feature parameters of a spatial topological network graph involves spatially arranging the node positions of the network graph according to the actual three-dimensional coordinates of the defects, and attaching the calculated topological feature parameters as property annotations to the graph. The final result is a composite data object containing a list of nodes, an edge list, node geometric attributes, overall graph feature parameters, and visual layout information. This composite object, which fully describes the spatial distribution and connectivity of defects, is the spatial topological graph.

[0143] The beneficial effects include generating independent 3D data blocks by extracting defect regions, accurately separating and preserving the complete 3D morphological information of each defect. Boundary extraction technology is used to obtain precise surface contours of defects, laying the foundation for subsequent geometric measurements. Systematic spatial metric analysis yields key geometric characterization data such as defect size, location, volume, and orientation. A spatial topological network is constructed based on adjacency relationships, and its graph theory characteristics are analyzed. Finally, these are integrated to form a comprehensive topological map describing the spatial distribution and connectivity of defects, achieving in-depth analysis of defect cluster structures.

[0144] S6. Based on the three-dimensional geometric representation data and the spatial topology map, the root cause of defects in the printer is traced to obtain the process parameter adjustment strategy for the printer.

[0145] In this embodiment of the invention, the step of tracing the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map to obtain a process parameter adjustment strategy for the printer includes:

[0146] Based on the three-dimensional geometric representation data and the defect identification results, process correlation analysis is performed on the morphological characteristics of the identified defects to obtain the potential process deviation types associated with the characteristics of the identified defects.

[0147] Based on the spatial distribution of defects in the spatial topology map, the clustering characteristics of the identified defects are used to trace the source of process anomalies, thereby obtaining the printing process anomaly pattern corresponding to the identified defects.

[0148] Based on the potential process deviation type and the abnormal printing process pattern, a rule mapping retrieval is performed on the preset process parameter rule base to obtain process parameter optimization suggestions for the identified defects.

[0149] The process parameter optimization suggestions and the defect identification results are structurally integrated to obtain the process parameter adjustment strategy for the printer.

[0150] The process of process correlation analysis involves jointly analyzing the three-dimensional geometric representation data of each identified defect with its corresponding defect type label in the defect identification results. Based on a pre-defined defect morphology feature library, which defines the defect morphology rules typically caused by different process deviations (e.g., elongated voids in a specific direction may be associated with uneven powder spreading, while spherical pores may be associated with molten pool instability), the system matches the actual measured size, shape proportions, and orientation information of each defect against the rules in the feature library, calculating the matching degree between each defect and the description of various potential process deviation morphologies. When the geometric features of a defect reach a pre-defined matching threshold with the typical morphology rule of a certain type of process deviation, that process deviation type is associated with the defect. After traversing all defects, a list is generated that lists one or more potential process deviation types that each defect may be associated with; this list represents the potential process deviation types.

[0151] The analysis for tracing process anomalies focuses on the distribution patterns of defect nodes in a spatial topology map. It analyzes the spatial clustering of defect nodes in the map to identify defect groups that are tightly clustered in space. For each identified defect cluster, its spatial distribution shape is analyzed (e.g., linear arrangement, planar distribution, or random scattering), combined with its three-dimensional spatial location in the printed part (e.g., near the edge, between layers, or concentrated at a specific height). These spatial clustering characteristics are compared with a known database of printing process anomaly patterns, which records typical defect spatial distribution characteristics caused by different equipment failures or process fluctuations. By matching, the most likely cause of the printing process anomaly corresponding to each defect cluster is determined. Finally, a report describing each defect cluster area and its corresponding process anomaly pattern is generated; this report is the printing process anomaly pattern report.

[0152] The rule mapping retrieval process takes identified potential process deviation types and printing process anomaly patterns as input, and queries a pre-defined process parameter rule base. This rule base is a structured knowledge base where each rule clearly defines the correspondence and adjustment direction between one or more process deviations or anomaly patterns and adjustable printing process parameters. The system uses the analyzed deviation types and anomaly patterns as query keywords, performs full-text matching and logical association searches in the rule base, and retrieves all related process parameter adjustment rules. All retrieved relevant rules are summarized and categorized according to their corresponding process systems to form a structured suggestion list, which serves as the process parameter optimization suggestion.

[0153] The structured integration process merges and organizes process parameter optimization suggestions with the original defect identification results to generate a final strategy document that directly guides equipment adjustments. This creates a structured document with multiple fields, fully replicating the key information from the defect identification results, including defect location, type, size, and confidence level. Following each defect entry or defect cluster entry, the process parameter optimization suggestions retrieved in the previous step are appended, clearly specifying the suggested parameter names (e.g., laser power, scanning speed, toner thickness), the adjustment direction (e.g., increase or decrease), and the suggested adjustment magnitude or range. The document summarizes all suggestions and groups them according to the process system, forming groups such as laser system adjustment group, toner delivery system adjustment group, etc., along with an explanation of the overall adjustment priority. This final, complete document containing specific defect information and clear adjustment instructions constitutes the printer's process parameter adjustment strategy.

[0154] The beneficial effects include accurately identifying the root causes of potential process deviations leading to various defects through matching analysis between defect morphology and the process rule base. Defect spatial clustering features are used to trace the source to specific abnormal patterns in the printing process, clarifying the stage and type of failure. Based on the accurately identified deviations and abnormal patterns, the process parameter rule base is searched to generate targeted and actionable parameter optimization suggestions. Finally, defect details and optimization suggestions are integrated into a structured adjustment strategy, providing printers with direct, clear, and comprehensive guidance for process parameter correction.

[0155] like Figure 2 The diagram shown is a functional block diagram of a 3D printing defect recognition system based on OCT and deep learning provided in an embodiment of the present invention.

[0156] The 3D printing defect recognition system 100 based on OCT and deep learning described in this invention can be installed in an electronic device. Depending on the functions implemented, the 3D printing defect recognition system 100 may include a system calibration module 101, a 3D imaging module 102, an image enhancement module 103, an intelligent recognition module 104, a 3D analysis module 105, and a process decision module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0157] In this embodiment, the functions of each module / unit are as follows:

[0158] The system calibration module 101 is used to perform coherent gating benchmark calibration on the OCT acquisition device based on the reference OCT image of the pre-printed metal, so as to obtain the standardized image acquisition benchmark of the OCT acquisition device.

[0159] The three-dimensional imaging module 102 is used to perform coherent volume data tomography reconstruction on the OCT interference signal acquired in real time during the metal printing process based on the standardized image acquisition benchmark, so as to obtain the coherent tomographic image of the OCT interference signal.

[0160] The image enhancement module 103 is used to enhance the feature contrast of the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal.

[0161] The intelligent recognition module 104 is used to perform multi-scale feature discrimination on the feature-enhanced image based on a preset deep learning defect recognition network, so as to obtain the defect recognition result of the feature-enhanced image.

[0162] The three-dimensional analysis module 105 is used to perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and to perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect.

[0163] The process decision module 106 is used to trace the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map, so as to obtain the process parameter adjustment strategy for the printer.

[0164] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0165] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0167] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0168] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying defects in 3D printing based on OCT and deep learning, characterized in that, The method includes: S1. Based on the reference OCT image of the pre-printed metal, perform coherent gating benchmark calibration on the OCT acquisition device to obtain the standardized image acquisition benchmark of the OCT acquisition device; S2. Based on the standardized image acquisition benchmark, coherent volume data tomography reconstruction is performed on the OCT interference signal acquired in real time during the metal printing process to obtain the coherent tomographic image of the OCT interference signal. S3. Perform feature contrast enhancement on the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal; S4. Based on a preset deep learning defect recognition network, perform multi-scale feature discrimination on the feature-enhanced image to obtain the defect recognition result of the feature-enhanced image; S5. Perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomography image to obtain three-dimensional geometric representation data of the identified defect, and perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect. S6. Based on the three-dimensional geometric representation data and the spatial topology map, the root cause of defects in the printer is traced to obtain the process parameter adjustment strategy for the printer.

2. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, The reference OCT image based on the pre-printed metal is used to perform coherent gating benchmark calibration on the OCT acquisition device to obtain a standardized image acquisition benchmark for the OCT acquisition device, including: Based on a preset zero optical path difference reference, an interferometric optical path analysis is performed on the reference OCT image of the pre-printed metal to obtain the optical path deviation result of the reference OCT image. Based on the optical path deviation results, the coherent gating of the OCT acquisition device is precisely registered to obtain the optical configuration parameters of the OCT acquisition device. Based on the optical property parameters of the pre-printed metal, the electronic parameters of the OCT acquisition device are adapted to obtain the electronic initialization parameters of the OCT acquisition device. The optical configuration parameters and the electronic initialization parameters are encapsulated into an operational reference to obtain a standardized image acquisition reference for the OCT acquisition device.

3. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, Based on the standardized image acquisition benchmark, the OCT interferometer signal acquired in real time during the metal printing process is reconstructed using coherent volume tomography to obtain a coherent tomographic image of the OCT interferometer signal, including: Based on the standardized image acquisition benchmark, the OCT interference signal acquired in real time during the metal printing process is demodulated to obtain the demodulated spectrum data of the OCT interference signal; Based on the scanning position parameters in the standardized image acquisition reference, the demodulated spectrum data is spatially encoded to obtain the encoded spectrum data of the demodulated spectrum data. The encoded spectrum data is reconstructed in the frequency domain to obtain the initial three-dimensional volume data of the OCT interference signal; The initial three-dimensional volume data is optimized for signal dynamic range to obtain the coherent tomographic volume image of the OCT interferometric signal.

4. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, The step of enhancing the feature contrast of the coherent tomography volume image to obtain the feature-enhanced image of the OCT interferometric signal includes: Directional diffusion filtering is applied to the coherent tomographic image to obtain a denoised tomographic image of the coherent tomographic image; Based on the local gray-level statistical characteristics of the denoised tomographic image, adaptive contrast optimization is performed on the denoised tomographic image to obtain a contrast-enhanced image of the coherent tomographic volume image. The contrast-enhanced image is reconstructed by grayscale distribution to obtain a standard grayscale image of the coherent tomography volume image; Edge structure sharpening is performed on the standard grayscale image to obtain the feature-enhanced image of the OCT interference signal.

5. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, The pre-defined deep learning defect recognition network performs multi-scale feature discrimination on the feature-enhanced image to obtain the defect recognition result of the feature-enhanced image, including: The suspected defect regions in the feature-enhanced image are spatially located to obtain the defect candidate regions of the feature-enhanced image; Based on a pre-defined deep learning defect recognition network, multi-scale convolutional feature extraction is performed on the defect candidate region to obtain a multi-scale feature vector set of the defect candidate region. Based on the deep learning defect recognition network and the multi-scale feature vector set, the defect candidate region is judged for defect category to obtain the preliminary defect classification result of the defect candidate region. Based on the decision rules in the deep learning defect recognition network, a comprehensive confidence decision is made on the preliminary defect classification results to obtain the defect recognition results of the feature-enhanced image.

6. The 3D printing defect identification method based on OCT and deep learning as described in claim 5, characterized in that, The step of performing a comprehensive confidence decision on the preliminary defect classification result based on the decision rules in the deep learning defect recognition network to obtain the defect recognition result of the feature-enhanced image includes: The confidence level of the preliminary defect classification results is evaluated to obtain the confidence level evaluation results of the defect candidate regions; Based on the spatial location relationship of the defect candidate regions, the spatial distribution logic verification of the preliminary defect classification result is performed to obtain the classification consistency verification result of the defect candidate regions. Based on the confidence assessment results and the classification consistency verification results, the preliminary defect classification results are optimized to obtain a credible defect classification result for the defect candidate region. The credible defect classification result and the confidence assessment result are structurally encapsulated to obtain the defect recognition result of the feature-enhanced image.

7. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, The step of performing three-dimensional geometric reconstruction of the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and performing spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topological map of the identified defect, includes: Based on the spatial location information in the defect identification result, three-dimensional mask data is generated on the coherent tomography image to obtain the defect three-dimensional data block of the defect identification result; The boundary contour of the defect three-dimensional data block is extracted to obtain the defect three-dimensional boundary data of the defect identification result; Based on the three-dimensional boundary data of the defect, spatial measurement analysis is performed on the identified defect in the defect identification result to obtain the three-dimensional geometric representation data of the identified defect. Based on the three-dimensional geometric representation data and spatial location information of the identified defect, spatial connectivity analysis is performed on the identified defect to obtain a spatial topological map of the identified defect.

8. The 3D printing defect identification method based on OCT and deep learning as described in claim 7, characterized in that, The step of performing spatial connectivity analysis on the identified defect based on its three-dimensional geometric representation data and spatial location information to obtain a spatial topological map of the identified defect includes: Based on the three-dimensional geometric representation data and spatial location information of the identified defect, a neighborhood correlation analysis is performed on the identified defect to obtain the adjacency relationship matrix of the identified defect. Based on the adjacency matrix, a spatial topology network is constructed for the identified defects to obtain a spatial topology network diagram of the identified defects. Graph theory feature analysis is performed on the spatial topology network graph to obtain the topological feature parameters of the spatial topology network graph; The topological feature parameters and the spatial topological network graph are integrated into a graph to obtain the spatial topological graph of the identified defect.

9. The 3D printing defect identification method based on OCT and deep learning as described in claim 1, characterized in that, The method of tracing the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map to obtain a process parameter adjustment strategy for the printer includes: Based on the three-dimensional geometric representation data and the defect identification results, the morphological characteristics of the identified defects are analyzed for process correlation to obtain the potential process deviation types associated with the identified defects. Based on the spatial distribution of defects in the spatial topology map, the clustering characteristics of the identified defects are used to trace the source of process anomalies, thereby obtaining the printing process anomaly pattern corresponding to the identified defects. Based on the potential process deviation type and the abnormal printing process pattern, a rule mapping retrieval is performed on the preset process parameter rule base to obtain process parameter optimization suggestions for the identified defects. The process parameter optimization suggestions and the defect identification results are structurally integrated to obtain the process parameter adjustment strategy for the printer.

10. A 3D printing defect recognition system based on OCT and deep learning, characterized in that, The system for implementing the 3D printing defect identification method based on OCT and deep learning as described in claim 1 includes: The system calibration module is used to perform coherent gating benchmark calibration on the OCT acquisition device based on the reference OCT image of the pre-printed metal, so as to obtain the standardized image acquisition benchmark of the OCT acquisition device. The three-dimensional imaging module is used to perform coherent volume data tomography reconstruction on the OCT interference signal acquired in real time during the metal printing process based on the standardized image acquisition benchmark, so as to obtain the coherent tomographic image of the OCT interference signal. The image enhancement module is used to enhance the feature contrast of the coherent tomography image to obtain a feature-enhanced image of the OCT interference signal; The intelligent recognition module is used to perform multi-scale feature discrimination on the feature-enhanced image based on a preset deep learning defect recognition network, so as to obtain the defect recognition result of the feature-enhanced image; The three-dimensional analysis module is used to perform three-dimensional geometric reconstruction on the defect identification result and the coherent tomographic image to obtain three-dimensional geometric representation data of the identified defect, and to perform spatial configuration analysis on the three-dimensional geometric representation data to obtain the spatial topology map of the identified defect. The process decision module is used to trace the root causes of defects in the printer based on the three-dimensional geometric representation data and the spatial topology map, so as to obtain the process parameter adjustment strategy for the printer.

Images