Laser powder bed fusion anomaly monitoring method based on molten pool time-lapse tomography
Patent Information
- Application Number
- CN202611365550.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-03
AI Technical Summary
Existing laser powder bed melting technology is difficult to achieve full coverage and high resolution monitoring of the melting process in scenarios involving multiple laser parallel scanning and large-size forming. Furthermore, existing equipment is complex to install, costly, and difficult to adapt to multi-laser systems.
Long-exposure acquisition was performed using an industrial area array camera mounted on a rangefinder. Combined with melt pool time-delay tomography, an anomaly feature database was established through automatic layering, layer-by-layer stitching, perspective transformation correction, and false color processing. Anomaly detection was then performed using the YOLOv3 model.
It achieves highly sensitive, comprehensive, and low-cost monitoring of abnormal melting processes in multi-laser, large-size forming cavities, supports rapid automatic identification and real-time intervention throughout the entire printing cycle, and reduces hardware modification costs.
Smart Images
Figure CN123033203A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of metal powder bed melting technology, and is a laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography. Background Technology
[0002] Laser Powder Bed Fusion (LPBF) is an additive manufacturing technology that uses a laser as a heat source to selectively melt metal powder layer by layer. It can integrally form parts with complex internal cavities, lattice structures, and gradient structures, and has broad application prospects in aerospace, biomedicine, and other fields. With the rapid development of equipment towards multi-laser and large-size applications, forming efficiency and unit value have significantly improved. However, the long printing cycle and high material and equipment operating costs mean that if anomalies occur during the printing process and are not detected in time, the entire part is often scrapped, resulting in huge economic losses. Therefore, online monitoring of the LPBF forming process to achieve early detection and intervention of anomalies is of great significance for ensuring forming quality and reducing failure losses.
[0003] Online monitoring of LPBF can be divided into two categories based on the timing of monitoring: powder bed monitoring and melting process monitoring. Powder bed monitoring mainly refers to monitoring after powder spreading and after forming. Its imaging conditions are controllable, the signal-to-noise ratio is high, and it is easy to quantitatively interpret. However, it cannot capture dynamic information at the moment of melting, and by the time an anomaly is detected, irreversible defects have often formed, making it difficult to guide real-time intervention. Melting process monitoring refers to in-situ acquisition during laser scanning and melting. It can directly reflect the behavior of the molten pool and the state of energy deposition, and has the potential for real-time feedback and closed-loop control. However, it requires high resolution and anti-interference capabilities from the acquisition system.
[0004] Existing methods for monitoring the melting process mainly include coaxial photodiodes, coaxial high-speed cameras, and optical tomography. Coaxial photodiodes and coaxial high-speed cameras acquire molten pool radiation signals through a scanning galvanometer optical path, offering high temporal resolution. However, they require the integration of a beam splitting and acquisition module into the laser optical path, resulting in complex installation, high modification costs, and the need for an independent coaxial channel for each laser path, leading to poor compatibility with multi-laser systems and difficulty in adapting to multi-laser forming scenarios. Optical tomography uses a scientific area array camera in conjunction with narrowband filters to perform long-exposure integral imaging of the forming area, obtaining the light distribution of the entire molten pool. However, existing solutions are limited by pixel size and field of view, resulting in limited spatial resolution and a relatively large minimum defect size that can be monitored, making it difficult to meet the high-sensitivity detection requirements for minute defects. Therefore, developing high-resolution, large-field-of-view, and low-cost area array imaging-based molten process monitoring technology has become an important direction.
[0005] Monitoring the melting process for industrial applications still faces multiple challenges. First, it requires blind-spot-free monitoring under conditions of multi-laser parallel scanning and overlapping stitching, and full field-of-view coverage of large-size forming cavities, while balancing global coverage and local high-resolution imaging. Second, it requires continuous and non-destructive data acquisition throughout the entire printing cycle, and the ability to identify anomalies quickly after they occur to support the development of real-time intervention technologies. At the same time, it is necessary to reduce hardware and computing costs while ensuring monitoring performance to adapt to large-scale industrial deployment.
[0006] Therefore, there is an urgent need for a melting process monitoring method that is oriented towards the printing process, applicable to multiple lasers, large sizes, and low cost, in order to achieve highly sensitive, full-coverage, and full-cycle online monitoring of LPBF process anomalies, ensure the quality of formed parts and the stability of the forming process, and reduce printing failure losses. Summary of the Invention
[0007] The purpose of this invention is to provide a laser powder bed melting anomaly monitoring method based on melt pool time-delay tomography, in order to solve the problems of existing coaxial monitoring which is complicated to install and difficult to adapt to multi-laser forming scenarios, and the low spatial resolution of optical tomography due to the limitation of the pixel scale of scientific area array cameras, which makes it difficult to achieve both large-format full coverage and local high resolution. The invention aims to achieve highly sensitive, full-coverage, fast-response, and low-cost online monitoring of melting process anomalies for multi-laser, large-size forming cavities, and the entire printing cycle.
[0008] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides a method for monitoring melting anomalies in a laser powder bed based on molten pool time-delay tomography, comprising the following steps:
[0009] Step 1: During the entire laser printing process, long-exposure images of the melting process are continuously acquired using a melt pool time-delay tomography imaging device. After removing the non-printing background area from each long-exposure image, MTT signal extraction images are obtained, forming an MTT signal extraction image sequence.
[0010] Step 2: Perform automatic layering, layer-by-layer stitching, perspective transformation correction, and false color processing on the MTT signal extracted image in sequence to obtain the MTT false color stitched and corrected image of each printed layer.
[0011] Step 3: After registering the slice images of each printed layer with the corresponding MTT pseudo-color stitching correction images, the corresponding MTT pseudo-color stitching correction images are segmented based on the binary mask results of the printed areas of each slice image, and the two-dimensional features of each MTT pseudo-color stitching correction image are extracted.
[0012] Step 4: Perform 3D reconstruction using the MTT pseudo-color splicing and correction images of each printed layer, and extract 3D features from the 3D reconstruction results;
[0013] Step 5: Establish an MTT process anomaly feature database, train the YOLOv13 model, and use the trained YOLOv13 to monitor laser powder bed melting anomalies; the MTT process anomaly feature database consists of several data samples, each data sample includes an MTT pseudo-color splicing and correction image and its two-dimensional features, a three-dimensional reconstruction result and its three-dimensional features, and corresponding labels.
[0014] As a further improvement to one embodiment of the present invention, the automatic layering in step 2 is either automatic layering based on grayscale features or automatic layering based on the start and stop time points of the scraper, specifically including:
[0015] (1) Based on the full dark grayscale in the non-printing state, the MTT signal extraction images are sequentially identified in terms of grayscale range. The MTT signal extraction images that are equal to the full dark grayscale reference are marked as images without printing marks. Then, the MTT signal extraction image sequence located between two adjacent images without printing marks is taken as the set of MTT signal extraction images of the same printing layer, so as to realize automatic layering based on grayscale features.
[0016] (2) Match the scraper start time stamp with the start or end time stamp of MTT original image acquisition, and take the MTT signal extraction image sequence between two adjacent scraper starts as the MTT signal extraction image set of the same printing layer to realize automatic layering based on the scraper start and stop time points.
[0017] As a further improvement to one embodiment of the present invention, step 2, layer-by-layer splicing, includes,
[0018] Traverse the MTT signal extraction images within any printed layer, and perform the following stitching preprocessing on each MTT signal extraction image: Set a stitching grayscale threshold, and mark the positions above the stitching grayscale threshold as bright areas; Set a consecutive pixel number threshold, and remove the markings of bright areas with a consecutive pixel number below the consecutive pixel number threshold; Set an extension pixel number, and for each MTT signal extraction image, retain only the bright areas and the number of pixels that extend outward from the boundary of each bright area.
[0019] Images of MTT signals that have undergone splicing preprocessing within the same printing layer are extracted and superimposed to obtain grayscale spliced images of each printing layer. If the cumulative grayscale value of a pixel after superposition exceeds the set grayscale value upper limit, the grayscale value of that pixel after superposition is set as the maximum grayscale value.
[0020] As a further improvement to one embodiment of the present invention, step 3 includes:
[0021] The slice images of each printing layer are registered with the corresponding MTT pseudo-color splicing and correction images to achieve spatial alignment between the printing outline of the slice images and the actual printing outline of the MTT pseudo-color splicing and correction images.
[0022] A binary mask is generated based on the printing area of each slice image. The corresponding position of the mask in the corresponding MTT pseudo-color splicing and correction image is extracted. Each MTT pseudo-color splicing and correction image is segmented into the expected printing area MTT image and the unexpected printing area MTT image.
[0023] Based on a preset signal strength comparison standard library and part process parameters, two-dimensional statistical features and spatial features of each MTT pseudo-color splicing correction image are extracted; among them, the two-dimensional statistical features include the mean and variance of grayscale within the printing layer, and the spatial features include the grayscale distribution area and perimeter within the printing layer.
[0024] As a further improvement to one embodiment of the present invention, step 4 includes:
[0025] The printing layer thickness △z is obtained based on the part process parameters. The two-dimensional physical dimensions △x and △y of a single pixel are obtained based on the physical dimensions of the full-width visible light image. The voxel size is determined by △x, △y and △z. Combined with the printing layer order, voxel meshes are constructed for the MTT pseudo-color stitching correction image, the MTT image of the expected printing area, and the MTT image of the unexpected printing area for each printing layer. The MTT three-dimensional reconstruction results, the expected printing MTT three-dimensional reconstruction results, and the unexpected printing MTT three-dimensional reconstruction results are obtained.
[0026] Based on the signal strength comparison standard library and part process parameters, the three-dimensional statistical features and morphological features of the MTT three-dimensional reconstruction results are extracted. Among them, the three-dimensional statistical features include the mean and variance of grayscale between printing layers, and the morphological features include the volume and surface area of grayscale distribution between printing layers.
[0027] As a further improvement of one embodiment of the present invention, in step 1, the molten pool time-delay tomography device is mounted off-axis outside the forming chamber, so that the optical axis of the industrial area array camera and the scanning optical path are kept at an angle; the molten pool time-delay tomography device includes an industrial area array camera, a lens and an attenuator.
[0028] The resolution and field of view of the industrial area array camera are determined by the forming cavity size, installation distance and monitoring accuracy, so that the physical size of the single pixel of the long exposure image meets the defect detection requirements; the photosensitive range covers the visible light and part of the near-infrared region, so as to realize the synchronous acquisition of multi-spectral information of the target in visible light and near-infrared; the frame rate is not lower than the highest frame rate corresponding to the shortest long exposure time, to ensure the continuity of long exposure acquisition;
[0029] The choice of lens is determined by the installation distance and the field of view of the industrial area scan camera, so that it can cover the shaped area and meet the depth of field requirements, and focus on the acquisition plane;
[0030] The attenuator should meet the following requirements: when the industrial area scan camera is shooting the printing substrate area and the laser is not emitting light, the image is completely dark, that is, there is no effective signal output; at the same time, when the laser is emitting light for printing, the signal amplitude is completely within the signal sensitivity range of the industrial area scan camera.
[0031] As a further improvement of one embodiment of the present invention, during laser-emitting printing, the laser trajectory and signal intensity at each emission position are collected within a single long exposure, thereby forming a long-exposure image in which the background of the non-printing area is completely dark and the laser-printed area presents an effective monitoring signal.
[0032] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the program is executed on the processor, it implements the steps in the laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography as described above.
[0033] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps in the laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography as described above.
[0034] Compared with the prior art, the significant advantages of this invention are:
[0035] This invention adopts a long exposure acquisition scheme of a general industrial area array camera mounted on a rangefinder to construct a molten pool time-delay tomography imaging monitoring system that does not affect the laser optical path. It does not require an independent acquisition channel for each laser path, is naturally adapted to multi-laser parallel scanning and overlapping stitching scenarios, takes into account the full field of view coverage of large-size forming cavities and local high-resolution imaging, and significantly reduces hardware modification costs.
[0036] This invention establishes a signal intensity comparison standard library for different materials and process parameters, providing a unified benchmark for the quantitative interpretation of molten pool optical signals;
[0037] This invention develops a processing algorithm suitable for raw images of time-delay tomography of molten pools, achieving automatic layering, layer-by-layer stitching, perspective transformation correction, and false color processing. It registers feature points between the raw sliced data and single-layer time-delay tomographic images, using a sliced data mask to divide the image into expected and unexpected printing areas, and extracts two-dimensional features by combining signal intensity with a standard library. By constructing a voxel mesh, it performs three-dimensional reconstruction on complete information, expected printing areas, and unexpected printing areas respectively, extracting three-dimensional features. It establishes a time-delay tomographic imaging process anomaly feature database integrating two-dimensional and three-dimensional features, and trains a process anomaly detection model using labeled data to achieve rapid and automatic identification of process defects such as path filling anomalies and energy input anomalies, automatically generating structured anomaly reports containing anomaly type, spatial location, and geometric features. This supports highly sensitive, comprehensive, and fast-response online monitoring and real-time intervention throughout the entire printing cycle. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the hardware system and parietal setup for online monitoring of anomalies in the LPBF melt pool time-delay tomography process.
[0039] Figure 2 A flowchart is provided for establishing a standard library for the calibration and signal intensity comparison of molten pool time-delay tomography systems.
[0040] Figure 3 This diagram illustrates the automatic layering, layer-by-layer stitching, perspective transformation correction, and false color processing of time-delay tomography images of molten pools.
[0041] Figure 4 The following is a comparison of image stitching algorithms for melt pool delay tomography, wherein (a) is the image stitching algorithm of the present invention; (b) is the maximum value stitching algorithm; and (c) is the time integral stitching algorithm.
[0042] Figure 5 This is a flowchart of a process anomaly detection method based on melt pool time-delay tomography images.
[0043] Figure 6 The images are delayed tomographic images of the molten pool with abnormal scanning path, where (a) is a delayed tomographic image of the molten pool and (b) is a slice image of the printed layer.
[0044] Figure 7 The images are time-lapse tomographic images of molten pools with abnormal energy input, where (a) is a time-lapse tomographic image of a molten pool with abnormal energy and (b) is a time-lapse tomographic image of a molten pool with normal printing.
[0045] Figure 8 This is a flowchart of the method of the present invention. Detailed Implementation
[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0047] First, it should be noted that the molten-pool time-lapse tomography (MTT) system involved in this invention is a hardware and software system for real-time monitoring of the thermal characteristics of the molten pool during the laser powder bed melting process. Its hardware consists of a commercially available standard industrial area array camera, lens, and attenuator. Its image acquisition software acquires images of the printing process through continuous long exposures. The image acquired in a single long exposure is recorded as the original MTT image. During laser emission, the printed area of the original MTT image is highlighted, while the non-printed area is completely dark; during non-laser emission, the original MTT image is completely dark. Its image processing includes: grayscale preprocessing, which removes non-printing background areas from the original MTT image, and the processed image is recorded as the MTT signal extraction image; automatic layering, which divides the MTT signal extraction image into sets of MTT signal extraction images for each printing layer through grayscale recognition or scraper start / stop recognition; layer-by-layer stitching, which stitches the sets of MTT signal extraction images for each printing layer, resulting in one stitched image for each layer, recorded as the MTT grayscale stitched image; perspective transformation correction, which eliminates the spatial perspective angle of each layer of the MTT grayscale stitched image, and the processed image is recorded as the MTT grayscale stitched corrected image; and pseudo-color processing, which applies grayscale coloring to the MTT corrected grayscale stitched image, and the processed image is recorded as the MTT pseudo-color stitched corrected image.
[0048] like Figure 8 As shown, the present invention provides a laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography, the specific steps of which are as follows:
[0049] S1: Use Figure 1 The diagram shows the hardware system and off-axis setup for online monitoring of LPBF melt pool anomalies during time-lapse tomography. The LPBF melt pool time-lapse tomography monitoring equipment includes a laser generator, a powder spreading device, and a monitoring device. The laser generator includes a laser and a galvanometer; the powder spreading device includes a doctor blade, a substrate, and a powder bed; and the monitoring device includes an industrial area array camera, an adapter lens, an attenuator, and a computer. The melt pool time-lapse tomography system is set up off-axis outside the LPBF forming chamber to acquire long-exposure images of the melting process. The specific steps are as follows:
[0050] S11: Based on the laser powder bed melting equipment and the expected monitoring accuracy, select the industrial area array camera, lens, and attenuator: Determine the resolution and field of view of the industrial area array camera by considering the forming cavity size, installation distance, and monitoring accuracy, ensuring that the physical size of a single pixel meets the defect detection requirements; Select the lens based on the installation distance and field of view, ensuring it covers the forming area with sufficient depth of field and focuses on the acquisition plane; Select a suitable attenuator based on the molten pool radiation intensity to prevent overexposure and saturation, ensuring stable and clear signals under splashing dust, providing reliable images for quantitative analysis. The specific steps are as follows:
[0051] S111: The industrial area scan camera involved in this invention uses a commercially available standard industrial area scan camera as its hardware foundation, requiring no customized modifications. The improvement of this invention lies not in the industrial area scan camera hardware itself, but in the design of the optical system, acquisition method, and image processing algorithm, enabling the standard commercial industrial area scan camera to meet monitoring requirements. Using commercially available standard components ensures a stable supply chain, consistency across multiple batches, complete certification compliance, and controllable costs, significantly improving the engineering feasibility and scalability of the solution.
[0052] S112: The light-sensing range of the industrial area array camera should cover the 400-900nm wavelength band, i.e., the visible light and part of the near-infrared region, to achieve simultaneous acquisition of multi-spectral information of the target in both visible light and near-infrared, balancing detection capability and engineering feasibility; the resolution of the industrial area array camera should be matched with the monitoring area, and the physical monitoring size of a single pixel should not be less than 100μm, to ensure that the spatial resolution of the monitoring coverage area meets the detection requirements; since the exposure time of the industrial area array camera is set to long exposure mode, the frame rate of the industrial area array camera should not be lower than the highest frame rate corresponding to the shortest long exposure time, to ensure the continuity of long exposure acquisition.
[0053] S113: After setting up the industrial area scan camera and adapter lens, select the attenuator. The selected attenuator should meet two constraints simultaneously: when the industrial area scan camera is shooting the printing substrate area and the laser is not emitting light, the image should be completely dark, i.e., there is no effective signal output; at the same time, when the laser is emitting light normally for printing, the signal amplitude should fall entirely within the signal sensitive range of the industrial area scan camera (i.e., the full-scale range from no signal to the highest signal), and should be as far away as possible from the two edges of the signal sensitive range to retain sufficient dynamic margin and avoid the signal touching the lower limit no signal area or the upper limit overexposure area, which would cause information loss.
[0054] S114: Compared to the optical tomography (OT) scheme with a narrowband filter, the MTT of the industrial area array camera plus a broadband attenuator of this invention has three advantages: First, it provides richer information. The attenuator only attenuates the light intensity without cutting off the spectral band, and a single frame carries full-spectrum radiation information, which can still capture transient anomalous signals outside the narrowband passband. Second, it offers higher resolution. Standard high-pixel industrial area array camera models are readily available, and the attenuator is angle-insensitive and does not degrade the optical transfer function. Third, it is more cost-effective. Industrial area array cameras are cheaper for the same resolution specifications, and the attenuator is cheaper per unit than a narrowband filter and has strong versatility, which is conducive to large-scale deployment.
[0055] S12: The MTT online monitoring hardware system consists of an industrial area scan camera, lens and attenuator. It is mounted on an adjustable bracket and off-axis outside the forming chamber to keep the optical axis of the industrial area scan camera at an angle to the scanning optical path.
[0056] S13: Based on the image acquisition software corresponding to a commercially available industrial area scan camera, and combined with the frame rate setting of the industrial area scan camera, single long exposure parameters are set, and the camera is triggered cyclically at fixed long exposure intervals to continuously acquire images throughout the entire printing cycle. The original images are stored in chronological order. The specific steps are as follows:
[0057] S131: The long exposure time must be set in conjunction with the frame rate of the selected industrial area scan camera. The single exposure time should be higher than the shortest exposure time corresponding to the highest frame rate. During laser-based printing, the laser trajectory and signal intensity at each position are collected within a single long exposure, thus forming a long exposure image in which the background of the non-printing area is completely dark and the laser-printed area shows an effective monitoring signal. This image is the MTT raw image.
[0058] S132: Set the image acquisition mode of the industrial area array camera to continuous acquisition, and sequentially perform long exposure acquisition to ensure that the laser printing signal is not lost between two long exposures, so as to realize uninterrupted monitoring of the entire laser printing process. Each acquired image should include a timestamp of the start or end of the acquisition in its name.
[0059] S14: Using the grayscale thresholding method, remove the non-printing background area and retain only the light-printing area to obtain the MTT signal extraction image, thus achieving initial data cleaning of non-interest areas.
[0060] S2: Figure 2 The flowchart for establishing a calibration and signal intensity comparison standard library for the molten pool time-delay tomography system of the present invention is used to complete the perspective transformation correction calibration and pixel physical size calibration of the molten pool time-delay tomography system, and to establish a signal intensity comparison standard library under different materials and process parameters. The specific steps are as follows:
[0061] S21: Set up a high-intensity light source to illuminate the printing chamber, so that the calibration board has uniform brightness and clear texture. Extract corner features and use homography matrix estimation to perform perspective transformation correction. The rectangular substrate corresponds to the edge of the full-width visible light image which is parallel, and the circular substrate corresponds to the edge of the full-width visible light image which is tangent. The transformation matrix is uniformly encoded and stored for subsequent use in perspective transformation correction of layer-by-layer MTT grayscale stitched images.
[0062] S22: Set up a high-intensity light source to illuminate the printing chamber, place the standard measurement block on the substrate, and after uniform illumination, acquire a full-width visible light image and detect its edges. Count the number of pixels occupied by the measurement block, divide the actual size of the measurement block by the number of pixels occupied by the measurement block to obtain the physical size of a single pixel, and then obtain the pixel physical size and image physical size from the total number of pixels in the full-width visible light image. Record and store the calibration results to provide a spatial reference for subsequent defect location and geometric dimension quantitative analysis.
[0063] S23: For the same geometric test piece, select typical materials and process parameters such as laser power and scanning speed, continuously acquire time-delay tomography images, extract the intensity of the molten pool optical signal, and statistically analyze its proportion according to intensity level to establish a signal intensity comparison standard library.
[0064] S3: Figure 3 This diagram illustrates the automatic layering, layer-by-layer stitching, perspective transformation correction, and false-color processing of molten pool time-lapse tomography images according to the present invention. It is used for the preliminary processing of molten pool time-lapse tomography images, and the specific steps are as follows:
[0065] S31: Develop an automatic image layering algorithm for MTT signal extraction, set layer grayscale thresholds, determine the start and end of each layer's forming by identifying image grayscale features or the start and stop times of the scraper, and divide continuously acquired images according to printing layers. The specific steps are as follows:
[0066] S311: Using the full dark grayscale of the image in the non-printing state as a reference, the grayscale range of the MTT signal extraction images is identified sequentially. Images equal to the full dark grayscale reference are marked as images without printing marks. The sequence of MTT signal extraction images located between two adjacent images without printing marks is taken as the set of MTT signal extraction images for the same printing layer, realizing automatic layering based on grayscale features.
[0067] S312: Match the scraper start time stamp with the start or end time stamp of MTT original image acquisition, and take the MTT signal extraction image sequence between two adjacent scraper starts as the MTT signal extraction image set of the same printing layer to achieve automatic layering based on the scraper start and stop time points.
[0068] S32: Develop an MTT single-layer image stitching algorithm, set a stitching grayscale threshold, extract connected regions with grayscale values greater than the stitching grayscale threshold, perform grayscale integration and superposition on the extracted single-layer MTT signal image, and obtain the MTT grayscale stitched image of each printing layer. The specific steps are as follows:
[0069] S321: Traverse the MTT signal extraction images within any printed layer, and perform the following stitching preprocessing on each MTT signal extraction image: Set a stitching grayscale threshold, and mark the positions above the stitching grayscale threshold as bright areas; Set a consecutive pixel number threshold, and for the marked bright areas, eliminate bright areas with a consecutive pixel number lower than the consecutive pixel number threshold; Set an extension pixel number, and extend the bright area outward by the extension pixel number and retain the original grayscale value of the extension part;
[0070] S322: Retain the spatial positions of the highlight areas and extensions of each MTT signal extracted image within a single layer, and superimpose the images within the layer. That is, sequentially accumulate the gray values of the same pixel position in multiple images within a single layer to obtain an MTT gray-scale stitched image; wherein, if the accumulated value exceeds the upper limit of gray values, the gray value at that pixel position is taken as the maximum gray value. Compared with the maximum value stitching algorithm and the time integration algorithm, the stitching algorithm developed in this invention preprocesses each MTT signal extracted image within the layer, so that the stitched image eliminates stitching marks while eliminating bad pixel noise in non-printed industrial area array cameras, such as... Figure 4 As shown.
[0071] S33: Based on the perspective transformation correction data of the MTT system, the MTT grayscale stitched images of each layer are corrected to obtain the MTT grayscale stitched corrected image;
[0072] S34: Perform pseudo-color processing on the MTT grayscale stitched and corrected image, and color each layer of the image according to the grayscale value to obtain the MTT pseudo-color stitched and corrected image of each layer.
[0073] S4: Perform data registration between the printed layer slice image and the single-layer MTT pseudo-color stitching and correction image, and extract two-dimensional features from each layer of the MTT pseudo-color stitching and correction image. The specific steps are as follows:
[0074] S41: Extract feature points from the original slice data and the MTT pseudo-color stitching correction image, match the two sets of feature points and remove mismatches to achieve spatial alignment between the printed contour in the slice data and the actual printed contour in the MTT pseudo-color stitching correction image.
[0075] S42: Generate a binary mask based on the printed area of the printed layer slice image, with the printed area as the foreground and the non-printed area as the background. Extract the corresponding position of the mask in the MTT pseudo-color stitching and correction image of each layer, and segment the MTT pseudo-color stitching and correction image of each layer into the expected printed area MTT image and the unexpected printed area MTT image.
[0076] S43: Based on the preset signal strength comparison standard library and part process parameters, extract the two-dimensional statistical features and spatial features of each layer of MTT pseudo-color splicing correction image; among them, the statistical features include the gray mean and variance within the printing layer, and the spatial features include the gray distribution area and perimeter within the printing layer.
[0077] S5: Perform 3D reconstruction by stitching together and correcting images from each layer of MTT pseudo-color. The specific steps are as follows:
[0078] S51: Obtain the printing layer thickness △z based on the part process parameters, obtain the two-dimensional physical dimensions △x and △y of a single pixel based on the physical dimensions of the full-width visible light image, determine the voxel size based on △x, △y and △z, and construct voxel meshes for the MTT pseudo-color stitching correction image, the expected printing area MTT image, and the unexpected printing area MTT image for each printing layer, in combination with the printing layer sequence, to obtain the MTT 3D reconstruction results, the expected printing MTT 3D reconstruction results, and the unexpected printing MTT 3D reconstruction results.
[0079] S52: Based on the signal strength comparison standard library and part process parameters, extract the three-dimensional statistical and morphological features of each layer of MTT. The statistical features include the mean and variance of grayscale between printing layers, and the morphological features include the volume and surface area of grayscale distribution between printing layers.
[0080] S6: Establish an MTT process anomaly feature database. The database consists of several data samples. Each data sample includes an MTT pseudo-color stitching corrected image and its 2D features, a 3D reconstruction result and its 3D features, and a corresponding label. Use the trained YOLOv13 process anomaly detection model to detect anomalies in the laser powder bed melting process, such as... Figure 5 As shown, the time-lapse tomography image of the molten pool with an abnormal scanning path is as follows: Figure 6 As shown, the time-delay tomography image of the molten pool with abnormal energy input is as follows: Figure 7 As shown, the specific steps are as follows:
[0081] S61: The two-dimensional statistical features, spatial features, and complete information of the MTT pseudo-color splicing and correction image, as well as the three-dimensional statistical features and morphological features of the MTT, are associated and integrated according to the printing layer sequence. Normal and abnormal samples of each layer and each three-dimensional region are uniformly encoded and stored to establish a process abnormality feature database.
[0082] S62: Using the Labellmg data annotation software, abnormal areas are selected on the MTT pseudo-color stitching and correction image, and categories such as path filling abnormalities and energy input abnormalities are labeled. The annotation results are checked and corrected one by one, and a labeled melting process abnormal dataset is generated after unifying the storage format.
[0083] S63: Train the YOLOv13 process anomaly detection model using the melting process anomaly dataset. Divide the melting process anomaly dataset into training and validation sets, set appropriate learning rates, batch sizes, and iteration rounds, input MTT pseudo-color stitched and corrected images for end-to-end training, monitor indicators such as the loss curve and average accuracy during training, and adjust hyperparameters as needed to prevent overfitting;
[0084] S64: Using the trained YOLOv13 process anomaly detection model, process anomalies in preprocessed MTT pseudo-color stitching corrected images are quickly detected. The anomaly category and confidence level are output, and the inference results are associated with image coordinate information to achieve online fast detection covering the entire printing cycle.
[0085] S65: Generate a process anomaly report. The process anomaly information report includes the type of anomaly, the spatial location of the anomaly area, and the geometric characteristics of the anomaly area.
[0086] In summary, this invention proposes a laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography, comprising: setting up a paraxial acquisition end, unifying the spatial reference through perspective transformation correction and pixel physical size calibration, and establishing interpretation criteria by comparing signal intensity with a standard library; then completing preprocessing through automatic layering, layer-by-layer stitching, and pseudo-color rendering; dividing the expected and unexpected printing areas through layer feature point registration and printing area masking, extracting two-dimensional features such as gray-level distribution area and gray-level gradient, constructing a voxel mesh partitioned three-dimensional reconstruction, and obtaining three-dimensional features such as gray-level distribution volume and gray-level gradient change; establishing a two- / three-dimensional feature database of melting process anomalies, training a YOLOv13 model to identify path filling anomalies and energy input anomalies, and generating a report covering type, spatial location, and geometric features, providing a comprehensive, fast-response, and low-cost monitoring means for forming quality assurance and process intervention.
[0087] This invention designs a molten pool time-delay tomography imaging system composed of a general-purpose industrial area array camera, lens, and attenuator, which can handle both large-format panoramic and locally clear imaging. It designs a geometric correction and size conversion process under strong illumination and establishes a signal intensity comparison library that varies with material and process parameters. It designs an automatic layering, single-layer multi-image segmentation and overlay stitching, and pseudo-color rendering process based on grayscale features and scraper start / stop points to achieve automatic image reconstruction and intuitive visualization of each layer. It designs a geometric alignment and region separation system based on the layer contour, extracting two-dimensional and three-dimensional features for expected and unexpected regions respectively, achieving zonal quantitative characterization. Simultaneously, it designs an anomaly feature library integrating two-dimensional and three-dimensional information and a YOLOv13 process anomaly detection mechanism, supporting anomaly category discrimination, spatial positioning, and size calculation, automatically generating standardized detection results, and providing a complete and feasible solution for process quality control, real-time intervention, and continuous improvement.
[0088] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps in the laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography as described above.
[0089] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps in the laser powder bed melting anomaly monitoring method based on molten pool time-delay tomography as described above.
[0090] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the modules described above can be referred to the corresponding process in the aforementioned method implementation, and will not be repeated here.
[0091] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0092] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.
[0093] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer system (which may be a personal computer, server, or network system, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for monitoring melting anomalies in a laser powder bed based on time-delay tomography of the molten pool, characterized in that, The method includes: Step 1: During the entire laser printing process, long-exposure images of the melting process are continuously acquired using a melt pool time-delay tomography imaging device. After removing the non-printing background area from each long-exposure image, MTT signal extraction images are obtained, forming an MTT signal extraction image sequence. Step 2: Perform automatic layering, layer-by-layer stitching, perspective transformation correction, and false color processing on the extracted MTT signal images in sequence to obtain the MTT false color stitching and correction images of each printed layer; Step 3: After registering the slice images of each printed layer with the corresponding MTT pseudo-color stitching correction images, the corresponding MTT pseudo-color stitching correction images are segmented based on the binary mask results of the printed areas of each slice image, and the two-dimensional features of each MTT pseudo-color stitching correction image are extracted. Step 4: Perform 3D reconstruction using the MTT pseudo-color splicing and correction images of each printed layer, and extract 3D features from the 3D reconstruction results; Step 5: Establish an MTT process anomaly feature database, train the YOLOv13 model, and use the trained YOLOv13 to monitor laser powder bed melting anomalies; the MTT process anomaly feature database consists of several data samples, each data sample includes an MTT pseudo-color splicing and correction image and its two-dimensional features, a three-dimensional reconstruction result and its three-dimensional features, and corresponding labels.
2. The method as described in claim 1, characterized in that, The automatic layering in step 2 is either automatic layering based on grayscale features or automatic layering based on the start and stop time points of the scraper, specifically including: (1) Based on the full dark grayscale in the non-printing state, the grayscale range of the MTT signal extraction images is identified in sequence. The MTT signal extraction images that are equal to the full dark grayscale reference are marked as images without printing marks. Then, the sequence of MTT signal extraction images located between two adjacent images without printing marks is taken as the set of MTT signal extraction images of the same printing layer, so as to realize automatic layering based on grayscale features. (2) Match the scraper start time stamp with the start or end time stamp of MTT original image acquisition, and take the MTT signal extraction image sequence between two adjacent scraper starts as the MTT signal extraction image set of the same printing layer to realize automatic layering based on the scraper start and stop time points.
3. The method as described in claim 2, characterized in that, Step 2, layer-by-layer splicing, includes, Iterate through the MTT signal extraction images within any printed layer, and perform the following stitching preprocessing on each MTT signal extraction image: set a stitching grayscale threshold, and mark the positions above the stitching grayscale threshold as bright areas; Set a threshold for the number of consecutive pixels, and remove the markers of bright areas where the number of consecutive pixels is less than the threshold; set an extension pixel number, and extract images from each MTT signal, retaining only the bright areas and the number of pixels that extend outward from the boundary of each bright area. Images of MTT signals that have undergone splicing preprocessing within the same printing layer are extracted and superimposed to obtain grayscale spliced images of each printing layer. If the cumulative grayscale value of a pixel after superposition exceeds the set grayscale value upper limit, the grayscale value of that pixel after superposition is set as the maximum grayscale value.
4. The method as described in claim 3, characterized in that, Step 3 includes: The slice images of each printing layer are registered with the corresponding MTT pseudo-color splicing and correction images to achieve spatial alignment between the printing outline of the slice images and the actual printing outline of the MTT pseudo-color splicing and correction images. A binary mask is generated based on the printing area of each slice image. The corresponding position of the mask in the corresponding MTT pseudo-color splicing and correction image is extracted. Each MTT pseudo-color splicing and correction image is segmented into the expected printing area MTT image and the unexpected printing area MTT image. Based on a preset signal strength comparison standard library and part process parameters, two-dimensional statistical features and spatial features of each MTT pseudo-color splicing correction image are extracted; among them, the two-dimensional statistical features include the mean and variance of grayscale within the printing layer, and the spatial features include the grayscale distribution area and perimeter within the printing layer.
5. The method as described in claim 4, characterized in that, Step 4 includes: The printing layer thickness △z is obtained based on the part's process parameters. The two-dimensional physical distances △x and △y of a single pixel are obtained based on the physical dimensions of the full-width visible light image. The voxel size is determined by △x, △y, and △z. Combined with the printing layer sequence, voxel meshes are constructed for the MTT pseudo-color stitching correction image, the MTT image of the expected printing area, and the MTT image of the unexpected printing area for each printing layer. The MTT 3D reconstruction results, the expected printing MTT 3D reconstruction results, and the unexpected printing MTT 3D reconstruction results are obtained. Based on the signal strength comparison standard library and part process parameters, the three-dimensional statistical features and morphological features of the MTT three-dimensional reconstruction results are extracted. Among them, the three-dimensional statistical features include the mean and variance of grayscale between printing layers, and the morphological features include the volume and surface area of grayscale distribution between printing layers.
6. The method as described in claim 1, characterized in that, In step 1, the molten pool time-delay tomography equipment is mounted off-axis outside the forming chamber to maintain an angle between the optical axis of the industrial area array camera and the scanning optical path; the molten pool time-delay tomography equipment includes an industrial area array camera, a lens, and an attenuator. The resolution and field of view of the industrial area array camera are determined by the forming cavity size, installation distance and monitoring accuracy, so that the physical size of the single pixel of the long exposure image meets the defect detection requirements; the photosensitive range covers the visible light and part of the near-infrared region, so as to realize the synchronous acquisition of multi-spectral information of the target in visible light and near-infrared; the frame rate is not lower than the highest frame rate corresponding to the shortest long exposure time, to ensure the continuity of long exposure acquisition; The choice of lens is determined by the installation distance and the field of view of the industrial area scan camera, so that it can cover the shaped area and meet the depth of field requirements, and focus on the acquisition plane; The attenuator satisfies the following: when the industrial area scan camera is shooting the printing substrate area and the laser is not emitting light, the image is completely dark, that is, there is no effective signal output; at the same time, when the laser is emitting light for printing, the signal amplitude is completely within the signal sensitivity range of the industrial area scan camera.
7. The method as described in claim 6, characterized in that, During laser-based printing, the laser trajectory and signal intensity at each emission point are captured within a single long exposure, thus forming a long-exposure image in which the background of the non-printing area is completely dark and the laser-printed area shows an effective monitoring signal.
8. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps of the method as described in any one of claims 1-7.
9. A storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.