An online monitoring, diagnosis and control system for additive manufacturing structures and powder spreading anomalies
By combining industrial cameras and deep learning models, the problem of accurate identification and feedback control of defects with similar characteristics during laser powder bed melting was solved, achieving high-precision defect detection and timely feedback, ensuring part quality and equipment stability.
Patent Information
- Application Number
- CN202311778187.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-21
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-12-21
AI Technical Summary
Existing in-situ testing methods are unable to accurately distinguish and promptly detect defects with similar characteristics such as structural deformation, abnormal surface morphology, and uneven powder spreading during the laser powder bed melting process, leading to equipment damage and unstable part quality.
By combining industrial cameras with computer vision technology, using the YOLOv7 defect detection model and the U-Net image segmentation deep learning model, and combining them with a long short-term memory (LSTM) neural network, rapid identification and diagnosis of defects can be achieved, and feedback control can be provided through warning lights.
It achieves high-precision in-situ detection of three types of defects with similar image features, and can detect high-hazard defects five layers in advance to prevent equipment damage and deterioration of part quality. The detection accuracy reaches 94%, 96% and 100%.
Smart Images

Figure CN118115429B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of metal powder bed melting technology, specifically relating to an online monitoring, diagnosis and control system for additive manufacturing structures and powder spreading anomalies. Background Technology
[0002] Laser powder bed fusion (L-PBF) is a widely used additive manufacturing process that can directly manufacture complex or customized parts in one piece, without being limited by part shape or cumbersome traditional processing steps. However, the complex laser-material interaction leads to a variety of abnormal defects in the printing process, affecting manufacturing stability. Among these, macroscopic defects pose the greatest direct threat; severe part deformation and fracture can not only compromise part integrity but also cause irreversible damage to the equipment. Therefore, in-situ detection technology is essential during the printing process to accurately, reliably, and rapidly detect and control process defects, ensuring product quality and maintaining stable manufacturing.
[0003] Previous studies have proposed various signal sensing approaches and data processing methods to capture different features during the printing process and diagnose abnormal defects, such as industrial cameras, high-speed cameras, photodiodes, and ultrasonic sensors. Among the various in-situ monitoring strategies developed for the L-PBF process, using layer-by-layer signals from the deposited layer and powder layer to detect part quality anomalies is a commonly used strategy. Industrial cameras, with their large imaging area, ability to detect a wide variety of defects, and low cost, have become one of the most suitable in-situ inspection methods for industrial applications, driven by the rapid development of machine vision technology.
[0004] Structural deformation, abnormal surface morphology, and uneven powder spreading are three common defects that occur during the printing process. Abnormal surface morphology manifests as dense protrusions at the edges of the part's upper surface. This is because when a high-energy-density laser scans the part's edges, the molten pool absorbs powder from 3-4 sides, whereas inside the sample, it typically absorbs powder from only 2 sides. Excessive powder accumulation leads to protrusions at the edges exceeding the powder spreading thickness. Structural deformation is a macroscopic deformation defect caused by the accumulation of residual stress within the part. During manufacturing, the high-temperature gradient and rapid cooling rate generated by multiple thermal cycles cause complex temperature and stress evolution in the part, and the accumulated residual stress leads to structural deformation defects. Uneven powder spreading generally does not result in obvious abnormalities in the part's morphology. It is usually due to unstable operation of the powder spreading device or slight over-prominence on the part's surface, resulting in insufficient powder spreading to the predetermined thickness on the upper part of the part during the powder spreading process.
[0005] In-situ powder bed detection still faces the following challenges. First, current in-situ detection solutions mainly utilize computer vision technology to identify defects with highly differentiated image features, such as scraper marks, fragments, and structural deformation. However, when faced with defects exhibiting similar characteristics—structural deformation, abnormal surface morphology, and uneven powder bed distribution—simple target detection models perform poorly. For example, early-stage structural deformation defects may be misdiagnosed as abnormal surface morphology, and later-stage structural deformation defects may be misdiagnosed as uneven powder bed distribution. Facing rapidly evolving defects like structural deformation, the inability to make timely and accurate type diagnoses and implement corrective measures could potentially cause irreversible damage to the equipment. Second, in-situ detection during the printing process primarily focuses on the powder bed surface, lacking awareness of the internal state of parts and their longitudinal deformation, thus failing to determine the actual condition and provide decision-making solutions.
[0006] Therefore, there is an urgent need for an in-situ defect detection, diagnosis, and feedback control technology for laser powder bed melting processes. This technology aims to establish an in-situ monitoring system that can quickly and accurately distinguish between defects with similar characteristics, promptly identify high-risk defects that threaten equipment, and make feedback decisions to ensure part quality and manufacturing process stability. Summary of the Invention
[0007] The purpose of this invention is to provide an online monitoring, diagnosis and control system for additive manufacturing structures and powder spreading anomalies, solving the problem of high-precision in-situ diagnosis and feedback control of defects with similar image features.
[0008] The technical solution for realizing the present invention is as follows:
[0009] An online monitoring, diagnosis, and control system for additive manufacturing structures and powder bed anomalies includes a hardware system and a software system. The hardware system includes an industrial camera, a computer, and an alarm light. The industrial camera is mounted on a rangefinder inside the L-PBF forming chamber to capture powder bed images during the printing process. The computer is mounted outside the L-PBF forming chamber and connects to the industrial camera via a GigE Ethernet interface to receive the powder bed images captured by the industrial camera. The software system is used to realize online monitoring, diagnosis, and control of additive manufacturing structures and powder bed anomalies. The alarm light is connected to the computer via a USB serial port. The software system in the computer generates a feedback control strategy to control the alarm light to illuminate green, yellow, and red lights.
[0010] The software system functions include powder bed image acquisition, perspective transformation and correction of acquired images, illumination homogenization processing of corrected images, monitoring and diagnosis of structural and powder spreading defects in the processed images, control strategy feedback, and warning light control.
[0011] A method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies includes the following steps:
[0012] S1: Image acquisition and preprocessing of powder bed, including perspective transformation correction and illumination homogenization.
[0013] S2: Using the trained YOLOv7 defect detection model, quickly detect exposed defects in the preprocessed powder bed image in S1;
[0014] S3: Based on the detection results of exposed defects in the parts, use the ROI (Region of Interest) extraction algorithm to crop out the exposed defect area image of the parts from the preprocessed powder bed image;
[0015] S4: Using the trained U-Net image segmentation deep learning model, segment the defect shape of the exposed defect area image of the part and calculate the defect area;
[0016] S5: Create a defect diagnosis training dataset and use it to train a Long Short-Term Memory (LSTM) neural network model to diagnose the types of defects.
[0017] Further, step S1 includes the following steps:
[0018] S11: An industrial camera is mounted off-axis inside the L-PBF forming chamber to capture powder bed images during the printing process. The perspective transformation method is used to correct the images captured by the industrial camera. After correction, the image borders completely coincide with the edges of the substrate.
[0019] S12: The Gamma transformation illumination adaptive correction method is used to improve the brightness of dark areas in the image with uneven lighting after correction, and to display the content details of the dark areas.
[0020] S13: Uses anti-sharpening mask filtering to enhance the edge contrast and details of the parts with exposed defects in the image, and enhances the features of the exposed defects.
[0021] Further, step S2 includes the following steps:
[0022] S21: Based on the Labelimg data annotation software, labels are defined for the exposed defects of parts in the preprocessed powder bed images, and a dataset of exposed defects of parts is created.
[0023] S22: Train the YOLOv7 defect detection model using the part exposure defect dataset. Use the trained YOLOv7 defect detection model to quickly detect part exposure defects in the preprocessed powder bed image and generate a defect information file. The defect information file includes the defect type, X coordinate of the defect bounding box center point, Y coordinate of the defect bounding box center point, defect bounding box width, and defect bounding box length.
[0024] Further, step S3 includes the following steps:
[0025] S31: Based on the detection results of exposed defects in the parts, the ROI (Region of Interest) extraction algorithm is used to crop the exposed defect area image of the preprocessed powder bed image. The cropping process is as follows: the preprocessed powder bed image is cropped with a square formed by multiplying the larger value of the defect boundary box length and the defect boundary box width by the scaling factor δ, with the defect boundary box center point as the center, to obtain a cropped square image of the exposed defect area of the parts, where δ = 1.2.
[0026] Further, step S4 includes the following steps:
[0027] S41: Collect images of the defective areas of the parts, use Labelme mask image creation software to segment the shape of the defects in the images of the defective areas of the parts, generate binary mask image labels of the defects, and make an image segmentation dataset;
[0028] S42: Train the U-Net image segmentation deep learning model using an image segmentation dataset;
[0029] S43: Use the trained U-Net image segmentation deep learning model to segment the shape of defects in the image of the exposed defect area of the part, generate a binary mask image of the defects, and calculate the defect area of the current printing layer.
[0030] Further, step S5 includes the following steps:
[0031] S51: Use a simulation model to simulate the three-dimensional shape of the part in the current printing layer state, determine which type of defect the current defect belongs to, such as structural deformation, abnormal surface morphology, or uneven powder spreading, collect the area data of defects in n consecutive layers (n≥10) with the current printing layer as the last layer as input data, and use the defect type corresponding to the three-dimensional shape of the part in the current printing layer state as output to create a defect diagnosis training dataset, and use the defect diagnosis training dataset to train a long short-term memory (LSTM) neural network model.
[0032] S52: Using a trained Long Short-Term Memory (LSTM) neural network model, with the area of the first n layers, including the current layer, as input, diagnose the type of the defect.
[0033] A feedback control method for in-situ diagnosis of powder bed anomalies based on defect evolution mechanism includes the following steps:
[0034] S61: Combining simulation model results and printing experiment results, calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process;
[0035] S62: Assess the degree of harm to part quality and equipment operation caused by the height exceeding the powder bed surface, and formulate feedback strategies for different types of defects in the parts based on the degree of harm, wherein the different types of defects in the parts are diagnosed using the above method.
[0036] Furthermore, S61 includes the following:
[0037] S611: Use a simulation model to simulate the part shape during the printing process of parts that undergo structural deformation. Calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process based on the simulation results. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer.
[0038] S612: Measure the dimensions of parts with abnormal surface morphology and uneven powder distribution in the printing experiment results, and calculate the height of the part above the powder bed surface in the direction perpendicular to the powder bed surface. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer.
[0039] Furthermore, S62 includes the following:
[0040] S621: For parts whose structure has undergone irreversible deformation and whose structural deformation defects extend h millimeters (h≥2) beyond the powder bed surface in the direction perpendicular to the powder bed surface during the part's evolution, after the powder bed anomaly is diagnosed in situ based on the defect evolution mechanism, a decision is made to stop printing and the warning light flashes red.
[0041] S622: For surface morphology defects that protrude from the part and extend beyond the powder bed surface by m millimeters (0≤m≤2) in the direction perpendicular to the powder bed surface during the part's evolution, after the powder bed anomaly is diagnosed in situ based on the defect evolution mechanism, a decision is made to continue printing, and the warning light flashes green.
[0042] S623: For uneven powder distribution defects where the part has not deformed, after the powder bed abnormality is diagnosed in situ based on the defect evolution mechanism, a decision is made to re-spread the powder and reprint, and the warning light flashes yellow.
[0043] Compared with the prior art, the significant advantages of this invention are:
[0044] Based on simulation models and experimental results, this invention reveals the evolution patterns and hazards of different defects. The designed two-stage similarity feature detection and diagnosis model combines the "defect evolution pattern" and the "target detection model" to achieve high-precision in-situ detection and diagnosis of three types of image feature similar defects while ensuring image processing speed. Compared with the traditional YOLOv7 target detection model, the two-stage model of this invention can detect high-hazard defects 5 layers in advance, make timely decision feedback, and prevent defects from further damaging the quality of parts and equipment. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the hardware system for online monitoring, diagnosis, and control of L-PBF forming apparatus, additive manufacturing structure, and powder spreading anomalies.
[0046] Figure 2 Diagram of an online monitoring, diagnosis, and control software system for additive manufacturing structures and powder spreading anomalies.
[0047] Figure 3 The flowchart shows the adaptive homogenization and desharpening mask filtering processes for light.
[0048] Figure 4 Flowchart for online monitoring and diagnosis of additive manufacturing structure and powder spreading anomalies.
[0049] Figure 5 The diagram illustrates three types of defects: structural deformation, abnormal surface morphology, and uneven powder spreading.
[0050] Figure 6 A diagram illustrating a method for determining defect labels for a training dataset using a simulation model.
[0051] Figure 7 Feedback control strategy for additive manufacturing structures and powder spreading anomalies. Detailed Implementation
[0052] The present invention will be further described in detail below with reference to specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0053] To achieve the above objectives and realize online diagnostic feedback for defects with similar characteristics, the present invention will be further described in detail below with reference to the accompanying drawings:
[0054] S1: Using an additive manufacturing structure and powder bed anomaly online monitoring, diagnosis and control system, acquire and preprocess powder bed images, and perform perspective transformation correction and illumination homogenization processing on the powder bed images.
[0055] S11: The additive manufacturing structure and powder bed anomaly online monitoring, diagnosis, and control system of the present invention includes a hardware system and a software system. The powder bed image is acquired using the additive manufacturing structure and powder bed anomaly online monitoring, diagnosis, and control system.
[0056] S111: Figure 1 This relates to an L-PBF forming apparatus and an additive manufacturing structure and powder spreading anomaly online monitoring, diagnosis, and control hardware system. (Reference) Figure 1 The L-PBF forming apparatus used includes a laser, galvanometer, doctor blade, substrate, and powder bed. The hardware system for online monitoring, diagnosis, and control of additive manufacturing structures and powder spread anomalies includes an industrial camera, a computer, and warning lights. The industrial camera, mounted on a rangefinder inside the L-PBF forming chamber, captures images of the powder bed during the printing process. The computer, located outside the L-PBF forming chamber, connects to the industrial camera via a GigE Ethernet interface to receive the powder bed images and runs software to achieve online monitoring, diagnosis, and control of additive manufacturing structures and powder spread anomalies. The warning lights connect to the computer via a USB serial port. The software system on the computer generates feedback control strategies to control the warning lights to illuminate green, yellow, and red lights.
[0057] S112: Figure 2 This is a software system for online monitoring, diagnosis, and control of additive manufacturing structures and powder spreading anomalies, running on a computer within the hardware system. Functions include powder bed image acquisition, perspective transformation and correction of acquired images, illumination homogenization processing of corrected images, monitoring and diagnosis of structural and powder spreading defects in the processed images, control strategy feedback, and warning light control.
[0058] S12: Furthermore, a perspective transformation method is used to perform perspective transformation correction on the collected powder bed image of the printing process. After correction, the image border completely coincides with the edge of the substrate. This technology has been maturely applied and will not be described in detail here.
[0059] S13: Employing a Gamma transformation-based adaptive illumination correction method, this method enhances the brightness of dark areas in the corrected image with uneven lighting, thereby revealing details in those dark areas. The specific process is as follows: Figure 3 As shown, the image with uneven lighting is converted from the RGB color space to the HSV color space to obtain V channel data. Based on the Retinex model, a filter is used to extract the illumination components, which are then introduced as adaptive parameters into the Gamma correction curve for lighting correction. The image with corrected lighting is then converted from the HSV color space to the RGB color space to obtain the Gamma-transformed adaptively corrected lighting image.
[0060] S14: Uses anti-sharpening mask filtering to enhance the contrast of Gamma transform illumination adaptive correction image.
[0061] S2: Figure 4 The present invention provides a process for an additive manufacturing structure and powder bed online monitoring and diagnosis system. First, a trained YOLOv7 defect detection model is used to quickly detect exposed defects in the pre-processed powder bed image. If defects are found, the in-situ diagnosis process continues. If no defects are found, the next layer of powder is melted, and the powder bed image after the next layer of powder is laid is acquired and detected. The specific operations are as follows:
[0062] S21: Based on the Labelimg data annotation software, labels are defined for exposed defects in parts in the preprocessed powder bed images, creating a dataset of exposed defects in parts. For example... Figure 5 The defects exposed in the parts include three types: structural deformation, abnormal surface morphology, and uneven powder spreading. Among them, structural deformation and abnormal surface morphology are manifestations of abnormal part structure, while uneven powder spreading is a manifestation of abnormal powder spreading.
[0063] S22: Train the YOLOv7 defect detection model using the part exposure defect dataset. Use the trained YOLOv7 defect detection model to quickly detect part exposure defects in the preprocessed powder bed image and generate a defect information file. The defect information file includes the defect type, X coordinate of the defect bounding box center point, Y coordinate of the defect bounding box center point, defect bounding box width, and defect bounding box length.
[0064] S3: Further, based on the detection results of exposed defects in the parts, the ROI (Region of Interest) extraction algorithm is used to crop the image of the exposed defect area of the parts from the preprocessed powder bed image. The specific operation is as follows:
[0065] S31: Based on the detection results of exposed defects in the parts, the ROI (Region of Interest) extraction algorithm is used to crop the exposed defect area image of the preprocessed powder bed image. The cropping process is as follows: the preprocessed powder bed image is cropped with a square formed by multiplying the larger value of the defect boundary box length and the defect boundary box width by the scaling factor δ, with the defect boundary box center point as the center, to obtain a cropped square image of the exposed defect area of the parts, where δ = 1.2.
[0066] S4: Using the trained U-Net image segmentation deep learning model, segment the defect shape of the exposed defect area image of the part, and calculate the defect area. The specific steps are as follows;
[0067] S41: Collect images of the defective areas of the parts, use Labelme mask image creation software to segment the shape of the defects in the images of the defective areas of the parts, generate binary mask image labels of the defects, and make an image segmentation dataset;
[0068] S42: Train the U-Net image segmentation deep learning model using an image segmentation dataset;
[0069] S43: Using the trained U-Net image segmentation deep learning model, the shape of defects in the exposed defect area image of the part is segmented to generate a binary mask image of the defects. The defect area of the current printing layer is calculated as follows: using the cv2.connectedComponentsWithStats connected component statistics function in the OpenCV computer vision library, the area of each connected component in the mask image is counted, and connected components with an area smaller than the average area of all connected components are filtered out to obtain the number of pixels in the defect area in the binary mask image. The pixel size and spatial size represented by the image are extracted according to the Region of Interest (ROI), and the obtained number of pixels is converted into spatial area information. Before conversion, the unit of the value is "pixel count", and after conversion, the unit of the value is "mm". 2 ".
[0070] S5: Further, create a defect diagnosis training dataset and use it to train a Long Short-Term Memory (LSTM) neural network model to diagnose the types of defects. Specific steps include:
[0071] S51: Use a simulation model to simulate the three-dimensional shape of the part in the current printed layer state, and determine which type of defect the current defect belongs to: structural deformation, abnormal surface morphology, or uneven powder distribution. Figure 6 This paper demonstrates a method for determining defect types for a defect diagnosis training dataset using a simulation model. The method involves collecting defect area data from n consecutive layers (n≥10) with the current printed layer as the last layer as input data, and using the defect type corresponding to the 3D shape of the part in the current printed layer state as output data to create a defect diagnosis training dataset. This dataset is then used to train a Long Short-Term Memory (LSTM) neural network model.
[0072] S52: Using a trained Long Short-Term Memory (LSTM) neural network model, with the area of the first n layers, including the current layer, as input, diagnose the type of the defect.
[0073] A feedback control method for additive manufacturing structures and powder laying defects, the specific steps of which are as follows:
[0074] S61: Combining simulation model results and printing experiment results, calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process. The specific steps are as follows:
[0075] S611: Use a simulation model to simulate the part shape during the printing process of parts that undergo structural deformation. Calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process based on the simulation results. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer.
[0076] S612: Measure the dimensions of parts with abnormal surface morphology and uneven powder distribution in the printing experiment results, and calculate the height of the part above the powder bed surface in the direction perpendicular to the powder bed surface. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer.
[0077] S62: Assess the impact of height exceeding the powder bed surface on part quality and equipment operation; based on the severity of the impact, develop feedback strategies for different types of part defects, such as... Figure 7 The different types of defects in the parts are diagnosed using the method described above, with the specific steps as follows:
[0078] S621: For parts whose structure has undergone irreversible deformation and whose structural deformation defects extend h millimeters (h≥2) beyond the powder bed surface in the direction perpendicular to the powder bed surface during the part's evolution, after the powder bed anomaly is diagnosed in situ based on the defect evolution mechanism, a decision is made to stop printing and the warning light flashes red.
[0079] S622: For surface morphology defects that protrude from the part and extend beyond the powder bed surface by m millimeters (0≤m≤2) in the direction perpendicular to the powder bed surface during the part's evolution, after the powder bed anomaly is diagnosed in situ based on the defect evolution mechanism, a decision is made to continue printing, and the warning light flashes green.
[0080] S623: For uneven powder distribution defects where the part has not deformed, after the powder bed abnormality is diagnosed in situ based on the defect evolution mechanism, a decision is made to re-spread the powder and reprint, and the warning light flashes yellow.
[0081] In summary, the defect diagnosis and feedback control method of this invention combines the "defect evolution law" and the "target detection model" to achieve high-precision in-situ diagnosis and feedback control of three types of defects with similar image features. The detection accuracy rates for surface morphology abnormalities, structural deformation, and uneven powder spreading are 94%, 96%, and 100%, respectively. Compared with the traditional YOLOv7 target detection model that directly detects the three types of defects, this invention can detect highly hazardous defects 5 layers in advance and make timely decision feedback to prevent further damage to part quality and equipment.
[0082] This invention designs a two-stage defect diagnosis and feedback control model. It utilizes a YOLOv7-based defect detection model to detect exposed defects in parts, obtaining defect location and bounding box information. A Region of Interest (ROI) extraction algorithm is used to crop the defect region, and a U-Net neural network is used to segment the defects within the region, obtaining area information. An LSTM defect diagnosis model trained using defect evolution data obtained from simulation calculations and experiments is used, integrating the area of n consecutive layers (n≥10) with the current printing layer as the last layer as input data to perform defect type diagnosis, ensuring a feature-similar defect diagnosis accuracy of over 90%, and controlling the single image processing speed to within 4 seconds. Feedback decisions are formulated for different defects, and an in-situ defect diagnosis feedback system is developed to achieve in-situ diagnosis and strategy feedback for feature-similar defects.
[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies, characterized in that, Includes the following steps: S1: Image acquisition and preprocessing of powder bed, including perspective transformation correction and illumination homogenization. S2: Using the trained YOLOv7 defect detection model, quickly detect exposed defects in the preprocessed powder bed image in S1; S3: Based on the detection results of exposed defects in the parts, use the ROI (Region of Interest) extraction algorithm to crop out the exposed defect area image of the parts from the preprocessed powder bed image; S4: Using the trained U-Net image segmentation deep learning model, segment the defect shape of the exposed defect area image of the part and calculate the defect area; S5: Create a defect diagnosis training dataset and use it to train a Long Short-Term Memory (LSTM) neural network model to diagnose the types of defects. S5 includes the following steps: S51: Use a simulation model to simulate the three-dimensional shape of the part in the current printing layer state, determine which type of defect the current defect belongs to, such as structural deformation, abnormal surface morphology, or uneven powder spreading, collect the area data of defects in n consecutive layers with the current printing layer as the last layer as input data, n≥10, and use the defect type corresponding to the three-dimensional shape of the part in the current printing layer state as the output to create a defect diagnosis training dataset. Use the defect diagnosis training dataset to train the Long Short-Term Memory (LSTM) neural network model. S52: Using a trained Long Short-Term Memory (LSTM) neural network model, with the area of the first n layers, including the current layer, as input, diagnose the type of the defect.
2. The method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies as described in claim 1, characterized in that, S1 includes the following steps: S11: An industrial camera is mounted off-axis inside the L-PBF forming chamber to capture powder bed images during the printing process. The perspective transformation method is used to correct the images captured by the industrial camera. After correction, the image borders completely coincide with the edges of the substrate. S12: The Gamma transformation illumination adaptive correction method is used to improve the brightness of dark areas in the image with uneven lighting after correction, and to display the content details of the dark areas. S13: Uses anti-sharpening mask filtering to enhance the edge contrast and details of the parts with exposed defects in the image, and enhances the features of the exposed defects.
3. The method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies as described in claim 2, characterized in that, S2 includes the following steps: S21: Based on the Labelimg data annotation software, labels are defined for the exposed defects of parts in the preprocessed powder bed images, and a dataset of exposed defects of parts is created. S22: Train the YOLOv7 defect detection model using the part exposure defect dataset. Use the trained YOLOv7 defect detection model to quickly detect part exposure defects in the preprocessed powder bed image and generate a defect information file. The defect information file includes the defect type, X coordinate of the defect bounding box center point, Y coordinate of the defect bounding box center point, defect bounding box width, and defect bounding box length.
4. The method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies as described in claim 3, characterized in that, S3 includes the following steps: S31: Based on the detection results of exposed defects in the parts, the ROI (Region of Interest) extraction algorithm is used to crop the exposed defect area image of the preprocessed powder bed image. The cropping process is as follows: the preprocessed powder bed image is cropped with a square formed by multiplying the larger value of the defect bounding box length and the defect bounding box width by the scaling factor δ, with the defect bounding box center point as the center, to obtain a cropped square image of the exposed defect area of the parts, where δ=1.
2.
5. The method for online monitoring, diagnosis, and control of additive manufacturing structure and powder spreading anomalies as described in claim 4, characterized in that, S4 includes the following steps: S41: Collect images of the defective areas of the parts, use Labelme mask image creation software to segment the shape of the defects in the images of the defective areas of the parts, generate binary mask image labels of the defects, and make an image segmentation dataset; S42: Train the U-Net image segmentation deep learning model using an image segmentation dataset; S43: Use the trained U-Net image segmentation deep learning model to segment the shape of defects in the image of the exposed defect area of the part, generate a binary mask image of the defects, and calculate the defect area of the current printing layer.
6. A feedback control method for in-situ diagnosis of powder bed anomalies based on defect evolution mechanism, characterized in that, Includes the following steps: S61: Combining simulation model results and printing experiment results, calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process; S62: Assess the degree of harm to part quality and equipment operation caused by the height exceeding the powder bed surface, and formulate feedback strategies for different types of defects in the part based on the degree of harm, wherein the different types of defects in the part are diagnosed by the method described in any one of claims 1-5.
7. The feedback control method for in-situ diagnosis of powder bed anomalies based on defect evolution mechanism as described in claim 6, characterized in that, S61 includes the following: S611: Use a simulation model to simulate the part shape during the printing process of parts that undergo structural deformation. Calculate the height of the part exceeding the powder bed surface in the direction perpendicular to the powder bed surface during the evolution process based on the simulation results. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer. S612: Measure the dimensions of parts with abnormal surface morphology and uneven powder distribution in the printing experiment results, and calculate the height of the part above the powder bed surface in the direction perpendicular to the powder bed surface. The calculation method is: the height of the highest point of the part minus the cumulative thickness of the powder in the current printing layer.
8. The feedback control method for in-situ diagnosis of powder bed anomalies based on defect evolution mechanism as described in claim 7, characterized in that, S62 includes the following: S621: For structural deformation defects that have undergone irreversible deformation of the part structure and extend h mm beyond the powder bed surface in the direction perpendicular to the powder bed surface during the part evolution process, h≥2, after the powder bed abnormality is diagnosed in situ based on the defect evolution mechanism, a decision is made to stop printing and the warning light flashes red. S622: For surface morphology defects that protrude from the part and extend beyond the powder bed surface by m millimeters in the direction perpendicular to the powder bed surface during the part's evolution (0≤m≤2), after the powder bed anomaly is diagnosed in situ based on the defect evolution mechanism, a decision is made to continue printing, and the warning light flashes green. S623: For uneven powder distribution defects where the part has not deformed, after the powder bed abnormality is diagnosed in situ based on the defect evolution mechanism, a decision is made to re-spread the powder and reprint, and the warning light flashes yellow.
9. An online monitoring, diagnosis, and control system for additive manufacturing structures and powder spreading anomalies, wherein the online monitoring, diagnosis, and control system for additive manufacturing structures and powder spreading anomalies is used to implement the online monitoring, diagnosis, and control method for additive manufacturing structures and powder spreading anomalies as described in any one of claims 1-5, characterized in that, The additive manufacturing structure and powder bed anomaly online monitoring, diagnosis and control system includes a hardware system and a software system. The hardware system includes an industrial camera, a computer and a warning light. The industrial camera is mounted on a rangefinder inside the L-PBF forming chamber to capture powder bed images during the printing process. The computer is mounted outside the L-PBF forming chamber and uses a GigE Ethernet interface to connect to the industrial camera to receive the powder bed images captured by the industrial camera. The software system is used to realize online monitoring, diagnosis and control of additive manufacturing structure and powder bed anomalies. The warning light is connected to the computer via a USB serial port. The software system in the computer generates feedback control strategies to control the warning light to light up green, yellow and red lights. The software system functions include powder bed image acquisition, perspective transformation and correction of acquired images, illumination homogenization processing of corrected images, monitoring and diagnosis of structural and powder spreading defects in the processed images, control strategy feedback, and warning light control.
Citation Information
Patent Citations
Visual online detection method and system of SLM (selective laser melting) formed spreading powder quality
CN107402220A
Powder laying quality monitoring and control method in metal selective laser melting forming
CN114216911A
