Powder spreading quality online monitoring method based on image recognition

Abstract

The invention discloses a powder laying quality online monitoring method based on image recognition, and the method comprises the steps: integrating an image collection system of a multi-angle light source on powder laying equipment, and carrying out the real-time analysis of a powder laying layer through a weak supervision semantic segmentation model based on image-level label training; according to the model, fine segmentation of a powder laying area, holes and foreign matters is realized through a double-domain pseudo-label mining and self-refining module; multi-dimensional quality indexes such as uniformity, thickness distribution and defect quantification are extracted based on a segmentation result, finally, index deviation is mapped into a process parameter adjustment instruction through an intelligent decision engine, equipment is driven to conduct real-time correction, and closed-loop control is formed; according to the method, the dependence on labeled data is greatly reduced through weak supervised learning, the adaptability of the system to different materials and environments is improved, and the problems that a traditional method is poor in environment robustness, depends on a large amount of labeled data and is single in detection dimension are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of additive manufacturing and industrial vision detection, and particularly relates to a powder laying quality online monitoring method based on image recognition. BACKGROUND

[0002] In the process of precision manufacturing such as additive manufacturing or powder metallurgy, the uniformity, thickness consistency and whether there are defects (such as holes, foreign matters, scratches) of the powder laying layer are the key steps to determine the quality of the final product. The existing monitoring of the powder laying quality is mainly through offline sampling inspection based on machine vision or online judgment based on traditional image processing algorithms. The existing technology usually adopts threshold segmentation, edge detection or template matching method, carries out binaryzation processing on the collected powder laying layer image, and evaluates the powder laying quality by calculating pixel statistical features (such as mean gray value, variance) or carrying out simple morphological analysis. Some more advanced systems try to introduce a classic machine learning classifier to learn the preset features to realize the recognition of specific type of defects.

[0003] Firstly, the traditional image processing method highly depends on the manually designed features and fixed threshold, and has poor robustness to the differences in powder material reflection characteristics, environmental light changes and powder background texture interference, which easily leads to misjudgment. Secondly, although the scheme based on supervised learning performs well under specific conditions, a large number of pixel-level accurate defect image data are needed for model training, and the labeling of such data is time-consuming and laborious, and the cost is high, which seriously restricts the rapid deployment and iterative update of the model. Thirdly, the existing methods mainly focus on the detection of single type of defects or the evaluation of overall uniformity, and lack the ability of integrated and fine real-time analysis of multi-dimensional quality indicators (such as local thickness distribution and fine segmentation of different morphological defects) of the powder laying layer.

[0004] Therefore, in view of the above problems, the present application provides a powder laying quality online monitoring method based on image recognition. SUMMARY

[0005] In order to overcome the problems of poor environmental adaptability, dependence on a large amount of labeled data and single detection dimension of the existing powder laying quality monitoring method, the present application provides a powder laying quality online monitoring method based on image recognition. The method greatly reduces the dependence of the model on labeled data by introducing an image-level label training strategy, and adopts a network structure that fuses multi-dimensional context information to improve the fine recognition ability of complex defects. Finally, the recognition result is converted into a process adjustment instruction to realize real-time perception and closed-loop optimization of the powder laying quality.

[0006] The technical scheme of the present application is: a powder laying quality online monitoring method based on image recognition, comprising the following steps: S1, setting at least one industrial camera above the powder laying device, cooperating with a multi-angle light source system, to collect real-time images of the powder layer; S2, inputting the collected powder layer images into an image processing model based on weakly supervised semantic segmentation to generate a region segmentation map of the powder layer, wherein the image processing model is trained using image-level labels and extracts global features and local detail features of the powder region through a dual-domain pseudo-label mining module; S3, based on the region segmentation map, extracting uniformity indicators, thickness distribution indicators and defect detection indicators of the powder layer, wherein the defect detection indicators include quantitative indicators of foreign matter, holes, inkjet area filament, inkjet area scratches, inkjet area accumulation, inkjet area spheroidization, non-inkjet area scratches, non-inkjet area accumulation, layer shift and warping defects; S4, according to the extracted indicators, judging whether the current powder layer meets the preset quality standard, if not, outputting the defect type and location information and generating a process adjustment instruction; S5, according to the process adjustment instruction, real-time adjusting the powder laying speed, powder thickness or vibration frequency of the powder laying device to realize closed-loop control.

[0007] As a preferred, in step S1, the industrial camera is a high frame rate area array camera or a line array camera with a resolution not less than 5 million pixels, preferably with a polarizer to suppress the specular reflection of the metal or ceramic powder surface; the multi-angle light source system includes independently programmable control ring LED light source, coaxial light source and at least two lateral linear light sources, the system automatically switches or combines the illumination mode according to the reflection characteristics of the powder material and the current ambient light intensity, for example, for high-reflectivity metal powder, low-angle ring diffuse light is preferred combined with coaxial light, and for dark polymer powder, high-brightness lateral light is preferred.

[0008] As a preferred, in step S2, the weakly supervised semantic segmentation model adopts a network structure based on dual-domain pseudo-label mining and self-refinement strategy, specifically including: a global inference unit based on graph convolution, for modeling the context dependency relationship between the powder region and the non-powder region in the feature map, enhancing the recognition of large-area uniform region and boundary transition region; a significant region expansion controller, for dynamically adjusting the feature activation threshold according to the initial activation map, and gradually expanding the coverage of the powder region to the complete boundary; a self-refinement module, which adopts a boundary feature conversion mechanism, introduces a boundary confidence map in the feature propagation process to enhance the consistency of the internal features of the powder region and suppress the false recognition of the background and noise region; a cross-pixel dependency module, which further optimizes the detail accuracy of the powder region segmentation by calculating the similarity weight between pixels in the feature map and fusing multi-scale context information.

[0009] As preferred, in step S3: The uniformity index is obtained by calculating the standard deviation and coefficient of variation of the pixel gray value in the powder laying area. When the coefficient of variation is greater than a preset threshold, it is determined that the uniformity is unqualified. The thickness distribution index is calculated by combining the three-dimensional point cloud data of the powder laying layer reconstructed by the structured light projection system or the binocular stereo vision system. The range, mean and variance of the thickness distribution are analyzed. The warping defect is quantified by analyzing the normal vector change or height profile curvature of the edge region point cloud. The extraction of the defect detection index includes: S301, for foreign matter, hole, extracting its area, position, shape factor; S302, for inkjet area filament and non-inkjet area scratch, distinguishing and quantifying by calculating the aspect ratio, skeleton length and strike of the connected domain; S303, for the accumulation defect of inkjet area and non-inkjet area, quantifying by analyzing the ratio of protruding height to base area of local three-dimensional topography; S304, for the spheroidization defect of inkjet area, quantifying by calculating the circularity and height difference with the surrounding powder; S305, for layer shift defect, quantifying by comparing the center of gravity or contour of the feature area (such as edge, marker point) in the continuous layer segmentation graph.

[0010] As preferred, the method further includes distributed monitoring node setting: at least three monitoring nodes are equally spaced on the scraper travel path of the powder laying equipment. Each node corresponds to an independent image acquisition unit. The system processes and extracts features of the images of each node in parallel, and uses a multi-source information fusion algorithm to integrate the results of each node to generate an overall powder laying quality evaluation report.

[0011] As preferred, the method further includes real-time alarm, data tracing and online learning functions: When the powder laying quality anomaly is detected, the system automatically triggers an audible and visual alarm, and displays the defect position on the human-machine interface. At the same time, the system automatically packages the abnormal image, corresponding timestamp, process parameter sequence, quality index data and processing suggestion, and stores them in a distributed quality database with timestamp, supporting batch, equipment and defect type tracing query. The system will regularly add the defect labeling images confirmed by the operator to the incremental training set, and perform online incremental training on the weakly supervised semantic segmentation model, gradually improving the adaptability and generalization ability of the model to different materials, new process parameters and new defects.

[0012] As preferred, the method is applicable to at least one powder laying process in selective laser melting, binder jetting and powder pressing forming, and the quality data and process adjustment instructions generated by the method can be integrated into the manufacturing execution system of the factory to realize the whole-process traceability of the powder laying quality data and the continuous optimization of the process knowledge.

[0013] Advantages of the present application: 1. The present application can effectively overcome the complex industrial field interference such as metal powder reflection, low contrast of ceramic powder and substrate, and environmental light fluctuation, greatly improve the environmental adaptability and imaging stability of the monitoring system under different materials and different working conditions, and solve the problems of high false detection rate and poor robustness caused by fixed illumination and simple processing in traditional methods, by constructing a high-fidelity image acquisition system composed of a programmable multi-angle light source, a polarization filter and a high-resolution camera, and designing a "material-illumination" mapping table that can automatically match the optimal illumination mode according to the optical properties of the powder material, and performing adaptive correction and enhancement.

[0014] 2. The present application introduces a weakly supervised semantic segmentation model based on image-level label training, which automatically generates and optimizes pixel-level pseudo labels in the training stage through a dual-domain pseudo label mining module, and iteratively improves the segmentation accuracy combined with a self-refinement module, so that the model can achieve defect recognition and segmentation performance comparable to that of a fully supervised model only with image-level annotation (such as "with holes") without expensive and time-consuming pixel-level fine annotation, reducing data annotation cost by more than 90%, greatly reducing the landing threshold of the technology, and solving the problem of deep learning model's serious dependence on large amounts of labeled data.

[0015] 3. The present application is based on high-precision segmentation map, parallel extraction and quantification of uniformity indicators including gray variation coefficient and texture consistency, thickness and flatness indicators combined with three-dimensional topography reconstruction, and geometric and material feature indicators covering multiple defects such as holes, foreign matter and agglomerates, forming a comprehensive and three-dimensional digital twin description of the physical state of the powder layer, overcoming the limitations of traditional methods that detect only one dimension and cannot quantitatively classify and analyze the root cause of defects. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The system framework schematic diagram of the present application is shown; Figure 2 The workflow schematic diagram of the present application is shown. DETAILED DESCRIPTION

[0017] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0018] Referring to Figure 1 The present application provides an embodiment: an image recognition-based powder laying quality online monitoring method, comprising the following steps of: (1) fixing and installing an image acquisition unit behind or above a powder laying device (such as a scraper or a roller), the unit mainly comprising an imaging module, an illumination module and a synchronization and control unit, specifically: The imaging module adopts one or more high-resolution industrial area array cameras (such as resolution ≥ 5 million pixels), and the frame rate needs to match the highest movement speed of the powder laying device, so as to ensure no motion blur. A telecentric lens with low distortion and high depth of field is selected to ensure that the magnification is consistent throughout the field of view, and a linear polarizing mirror with adjustable angle is additionally installed in front of the lens to suppress the specular reflection of high-reflectivity materials (such as aluminum and titanium alloy powder).

[0019] The illumination module is a programmable multi-angle light source system, which comprises a ring-shaped LED diffuse light source for providing uniform basic illumination, a coaxial light source for highlighting surface texture and small undulations, and two groups of symmetrically arranged lateral strip light sources for generating grazing light to enhance the particle feeling and edge profile. The brightness and switching combination of all light sources can be automatically controlled by the upper computer software according to the “material-illumination” mapping table (for example, for stainless steel powder, the “30% ring light + 70% coaxial light” mode is adopted; for nylon powder, the “turn off the coaxial light and enable the two lateral grazing lights” mode is adopted).

[0020] The synchronization and control unit adopts a high-precision encoder or a photoelectric sensor to obtain a real-time position signal of the powder laying device, triggers the camera to perform image acquisition at the moment when each layer of powder laying is completed and the device returns to the initial position, so as to ensure that the image timing is strictly synchronized with the process cycle.

[0021] The acquired original image enters a preprocessing pipeline, is first dark-field and flat-field corrected to eliminate the camera background noise and uneven illumination, is then automatically cropped in a region of interest (ROI) to focus on the actual powder laying area, and is finally subjected to local contrast enhancement based on CLAHE (contrast limited adaptive histogram equalization) to highlight the powder distribution details. The preprocessed image stream is sent to subsequent steps in real time.

[0022] (2) Transform the pre-processed RGB or grayscale image into a pixel-level segmentation map that accurately identifies the "qualified powder spreading area", "thin / missing area (hole)", "powder agglomeration area", "foreign matter / pollution area", and "background (substrate or previous cladding layer)". First, construct an end-to-end weakly supervised semantic segmentation network. The backbone network can use ResNet or VGG pre-trained on ImageNet.

[0023] The dual-domain pseudo-label mining module of the network is embedded in the middle layer of the backbone network. It contains a global inference unit based on graph convolution, which projects the feature map to the interaction space, captures long-range context dependencies within the powder spreading area by constructing a feature node graph and performing graph convolution, and enables the model to understand the semantics of "even if the local texture changes slightly, it belongs to the same uniform powder spreading area". At the same time, the salient region expansion controller in the module is responsible for dynamically managing the threshold of feature activation. It generates a control vector by analyzing the global average pooling results of the feature map, which determines the extent to which the initially activated "most discriminative" region (such as the typical qualified powder region) can expand to the surrounding similar regions, thereby gradually growing the segmentation region from "point" or "small block" to cover the complete physical powder spreading range, effectively solving the problem of sparse initial response region under weak supervision.

[0024] The self-refinement and boundary optimization module of the network receives the coarse segmentation features output by the backbone network and the dual-domain module. Its core is the boundary feature transformation mechanism, which first predicts a "boundary confidence map" from the features through a lightweight subnetwork, where high values indicate a high likelihood of the pixel being located on the boundary between different regions. When the features are propagated forward, according to this confidence map, the pixel features belonging to the same region are strongly fused (strong connection), and the pixel features on both sides of the boundary between different regions are weakly suppressed (weak connection). This operation makes the network explicitly aware of the "boundary" at the feature level, greatly improving the accuracy of the segmentation boundary. Further, the cross-pixel dependency module, based on this, establishes a global pixel relationship model by calculating the attention weights between all pairs of pixels, so that the final class judgment of each pixel not only depends on its own features, but also fully considers the context information of all similar pixels globally, thus having strong robustness to uneven lighting, local shadows, and other disturbances, and being able to fill in small segmentation holes caused by noise.

[0025] The network model only needs to collect a large number of powder layer images and label each image with an image-level label, such as [qualified, with holes, with foreign matter], during training. It does not need to perform extremely time-consuming pixel-level contour labeling. Its training adopts a multi-stage training strategy. In the first stage, the image-level label and the dual-domain pseudo-label mining module are used to train the network to generate an initial pseudo-segmentation mask (i.e. the pixel-level label "guessed" by the model itself). In the second stage, the self-generated pseudo mask with high quality is used as a supervision signal, and the self-refining module is used to fine-tune the entire segmentation network including the backbone network in an end-to-end manner. This process is accompanied by an online screening mechanism that only selects pseudo labels with high prediction confidence of the model to participate in training, gradually iterates, and improves the model performance. This strategy enables the model to achieve a segmentation accuracy close to that of a fully supervised model, but the data preparation cost is reduced by more than 90%.

[0026] (3) After obtaining the fine segmentation map described above, the uniformity index, the thickness and flatness index, and the defect identification and quantification index are calculated in parallel, specifically: The uniformity index includes the gray level variation coefficient (CV) and the texture consistency index. For the gray level variation coefficient (CV), in the segmented "qualified powder area", the ratio of the standard deviation of all pixel gray values to the average value is calculated. The smaller the CV value, the more uniform the powder distribution. A threshold (such as 0.12) is set, and if it is exceeded, an alarm is given. For the texture consistency index, the texture feature variance in the segmented area is calculated by local binary pattern (LBP) or gray level co-occurrence matrix (GLCM), quantifying the texture changes caused by powder aggregation or sparse distribution.

[0027] The thickness and flatness index includes three-dimensional reconstruction auxiliary measurement, which reconstructs the three-dimensional morphology of the powder layer by combining the structured light projector installed at another angle, or using the principle of binocular stereo vision. According to the guidance of the segmentation map, only the point cloud of the "qualified powder area" is analyzed, and the thickness range, average thickness, and absolute deviation from the set thickness are calculated. The flatness can be evaluated by calculating the root mean square error of the plane fitting of the point cloud in this area.

[0028] The defect identification and quantification indicators include hole defects, foreign matter / pollution defects, and insufficient / excessive powder areas, specifically: for hole defects, the areas marked as "holes" in the segmentation map. Extract the features of each connected hole: area (number of pixels or actual square millimeters), equivalent diameter, position coordinates (X, Y), circularity. A circularity close to 1 indicates a circular hole, and much less than 1 may be a scratch or a crack. For foreign matter / pollution defects, the areas marked as "foreign matter". In addition to area and location, color / material features (such as color difference ΔE from standard powder in RGB space) and edge sharpness analysis are added to distinguish between foreign particles, hair, or oil stains. For insufficient / excessive powder areas, combined with three-dimensional data and the segmentation map, identify areas with a thickness less than 80% or more than 120% of the set value, and calculate the total area ratio and distribution dispersion.

[0029] (4) For simple and clear defects, predefined rules are used (for example: "if a single hole area > hole area threshold, determine that the scraper is sticking powder, instruction: increase the scraper ultrasonic vibration frequency by 10%"; "if the uniformity CV > gray variation coefficient threshold, determine that the powder humidity or particle size distribution is uneven, instruction: start the powder screening system for one-time recirculation, and reduce the current powdering speed by 15%").

[0030] For complex or multi-factor coupled quality problems, a model trained based on historical data is used, which takes the current quality indicator vector as input and outputs a set of optimized process parameter adjustment amounts (powdering speed, powder layer thickness setting, substrate temperature, etc.), aiming to optimize the predicted quality indicators of the next layer or batch.

[0031] (5) The final adjustment instructions are issued in real time to the control system of the powdering equipment through standard industrial communication protocols. The device controller dynamically adjusts the corresponding actuators (such as servo motors, vibrators, heaters). The system immediately collects new images for evaluation in the next powdering period after adjustment, forming a "monitoring-analysis-adjustment-remonitoring" closed loop, achieving dynamic stability and continuous optimization of quality.

[0032] Please refer to Figure 2 The specific process of the method is as follows: (1) Deploy a high-resolution industrial camera with integrated multi-angle programmable light sources and polarization filters above or behind the powdering equipment. Capture images of the powder layer surface after powdering through an encoder trigger strictly synchronized with the process cycle, and output standardized images with uniform illumination and clear details through a preprocessing pipeline that includes dark field correction and adaptive contrast enhancement.

[0033] (2) The collected images are immediately sent to an image processing model based on weakly supervised semantic segmentation. The model uses global graph reasoning and salient region expansion control to learn and generate refined pixel-level segmentation maps from training data with only image-level labels through an embedded dual-domain pseudo-label mining module. The model accurately distinguishes between qualified powder areas, holes, foreign objects, and backgrounds. Then, through boundary feature transformation and cross-pixel dependency optimization in the self-refinement module, the model significantly improves the recognition accuracy and robustness of complex textures and weak defect boundaries.

[0034] (3) Based on the obtained accurate segmentation map, the system performs parallel multi-dimensional quality indicator quantization extraction. The gray level variation coefficient and texture consistency of the powder area are calculated to evaluate uniformity. The thickness distribution range and flatness are calculated in combination with three-dimensional topographic data. The area, position, shape factor, and material characteristics of the identified defects such as holes and foreign objects are quantitatively analyzed. The intuitive image is converted into a series of calculable and comparable objective quality parameters.

[0035] (4) The quantized indicators are input into an intelligent decision-making module. The module combines a pre-defined rule base and a prediction model trained based on historical data to map specific quality deviations to root cause process parameter adjustments (e.g., reducing powdering speed based on non-uniformity or adjusting scraper vibration parameters for specific types of holes), thereby generating executable process adjustment instructions.

[0036] (5) Finally, the adjustment instructions are real-time issued to the control system of the powdering equipment through an industrial communication protocol, driving servo motors, vibrators, and other actuators to dynamically adjust key parameters such as powdering speed, thickness, or vibration frequency. A new round of image acquisition and evaluation is immediately started in the next powdering cycle, forming a full-automatic closed loop of "monitoring-analysis-decision-making-control-re-monitoring" to achieve continuous online optimization and stable control of powdering quality.

[0037] The present application provides a comparative example 1: This embodiment is aimed at the use of expensive and slightly reflective titanium alloy powder in the metal selective laser melting (SLM) process. The subtle defects in powdering quality directly affect the mechanical properties of the final part and may cause process defects such as spheroidization. Meanwhile, the change in particle size distribution of the powder during recycling further increases the complexity and challenge of quality monitoring.

[0038] This embodiment selects an 850nm wavelength near-infrared coaxial light source to effectively penetrate the fine plasma interference that may occur during the process. A narrow-band filter is installed at the camera end to ensure image clarity. The 5000 titanium alloy powdering images used for model training only require image-level labels, significantly reducing the threshold and cost of data preparation.

[0039] To verify the effect of the present application, two comparative examples are selected for verification. Comparative Example 1-1 is a traditional image processing method, which adopts fixed threshold segmentation + connected domain analysis. Comparative Example 1-2 adopts a full-supervised deep learning method, adopts a U-Net model, but needs 2000 images with accurate pixel labeling for training.

[0040]

[0041] From the above table, it can be seen that the present application greatly reduces the data labeling cost while maintaining the detection accuracy similar to the full-supervised model, and realizes the closed-loop control that the traditional method cannot achieve, greatly improves the response speed and quality stability, and successfully realizes the reliable identification of all process-related holes with a diameter greater than 80 microns, and can stably monitor the gray scale variation coefficient of powder uniformity in the interval of 0.08-0.11 in real time, and when the system detects that the coefficient exceeds the threshold value of 0.13 for three times in a row, an alarm will be automatically triggered. In addition, the system can also effectively distinguish and detect a small amount of foreign materials such as stainless steel powder and micro-fibers in the environment. In the comparative example experiment, the present application maintains a high detection accuracy and a high foreign material recognition rate similar to the full-supervised model, and overcomes the shortcomings of the traditional method, such as sensitivity to light conditions, high false positive rate and inability to form closed-loop control.

[0042] The present application provides a comparative example 2: The present embodiment is aimed at zirconia powder used in the ceramic binder jetting process, which has inherent white characteristics and forms a very low visual contrast with the same color ceramic substrate, which brings great difficulty to accurately identify the powder laying boundary. At the same time, slight differences in powder flowability are very easy to cause "edge effect" in the edge area and cause irregular "step lines" in the cross-sectional profile, thereby verifying the monitoring sensitivity of the present application.

[0043] The present embodiment adopts a high-angle lateral grazing illumination scheme, which greatly enhances the visual contrast of the edge profile by using the magnification effect of light on the small height difference between the powder and the substrate. At the same time, the global inference unit in the segmentation model is optimized specifically, so that it can more fully utilize the context semantic information of the image in the feature fusion stage, thereby compensating for the problem of insufficient local contrast.

[0044] To verify the effect of the present application, one comparative example is selected for verification. Comparative Example 2-1 is a method of using edge detection algorithm, which adopts Canny operator combined with Hough transform to detect the powder laying boundary.

[0045]

[0046] From the above table, it can be seen that when facing the challenge of low contrast and complex background, the application far exceeds the traditional edge detection method based on low-order features by virtue of the regional segmentation ability based on semantic understanding, and realizes the accurate quantification of the powder laying geometry. From the table, it can be seen that the application has an order of magnitude advantage in edge positioning accuracy, can provide complete and smooth boundary contour lines, and can realize nearly 100% detection rate for "step line" defects. The traditional method is interrupted by internal texture and is not sensitive to step line, which fully proves the practicability of the application in solving the problem of low contrast and complex geometry monitoring.

[0047] The application provides a comparative example 3: The embodiment is aimed at the nylon powder material used in the polymer powder bed fusion process. The physical property of easy electrostatic charging often causes "dust raising" and "clumping" phenomenon, which may cause the powder laying surface to appear filamentous floating fibers or clumped agglomerates that are difficult to distinguish from normal texture. At the same time, the uniformity of powder preheating is very important for the final forming quality, so as to verify the texture differentiation of the application.

[0048] The embodiment adopts a lighting scheme mainly using annular diffuse light to uniformly and softly display surface details, and specially defines and quantifies the "agglomeration index" in the image processing model. The index can intelligently distinguish the two different abnormal phenomena of filamentous floating fibers and clumped blocks by fusing the shape factor analysis of suspicious areas and the local texture contrast calculation, so as to realize the accurate classification of typical defects of nylon powder.

[0049] In order to verify the effect of the application, a comparative example is selected for verification. Comparative example 3-1 adopts the background difference method, uses the first layer "ideal" powder laying image as the background, and performs difference between the subsequent layers and the background.

[0050]

[0051] From the above table, it can be seen that the application can not only detect static defects, but also perceive dynamic process abnormalities (such as static electricity), and has good adaptability to material batch changes, which is beyond the reach of the fixed background model method. From the table, it can be seen that the application has a very obvious advantage in the detection of agglomerates, especially in the sensitivity to filamentous materials, and is almost not affected by the color difference of powder batches or the natural fluctuation of normal texture. The traditional background difference method has high false positive rate, poor adaptability and cannot perceive dynamic static electricity interference.

[0052] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in other forms. Any skilled person in the art can modify or change the disclosed technical content into equivalent embodiments with equivalent changes, and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical solution content of the present application, still falls within the protection scope of the present application.

Claims

1. A method for online monitoring of powder spreading quality based on image recognition, characterized in that, It includes the following steps: S1. At least one industrial camera is installed above the powder spreading equipment, and a multi-angle light source system is used to acquire real-time images of the powder spreading layer. S2, the collected powder layer image is input into the image processing model based on weakly supervised semantic segmentation to generate the region segmentation map of the powder layer. The image processing model is trained with image-level labels and extracts the global features and local detail features of the powder layer region through the dual-domain pseudo-label mining module. S3. Based on the region segmentation map, extract the uniformity index, thickness distribution index and defect detection index of the powder layer. The defect detection index includes quantitative indicators of foreign matter, holes, inkjet stringing, inkjet scratches, inkjet accumulation, inkjet spheroidization, non-inkjet scratches, non-inkjet accumulation, layer migration and warping defects. S4. Based on the extracted indicators, determine whether the current powder layer meets the preset quality standards. If it does not meet the standards, output the defect type and location information and generate a process adjustment instruction. S5 adjusts the powder spreading speed, powder spreading thickness, vibration frequency, scraper height, or binder spraying parameters of the powder spreading equipment in real time according to the process adjustment instructions to achieve closed-loop control.

2. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that: The industrial camera is a high-resolution linear or area array camera, equipped with a polarizing filter to suppress reflections. The multi-angle light source system includes a ring light source, a coaxial light source, and a side light source, which can automatically switch or combine lighting modes according to the optical characteristics of the powder material.

3. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that, The weakly supervised semantic segmentation model adopts a network structure based on dual-domain pseudo-label mining and self-refinement strategies, including: The global inference unit is used to extract the contextual dependencies of the powder-spreading region in the feature map; A significant region expansion controller is used to enhance the recognition accuracy of powder-spreading area boundaries; The self-refinement module improves the separation accuracy between the powder-covered area and the background by iteratively optimizing the initial segmentation results.

4. The online monitoring method for powder spreading quality based on image recognition according to claim 3, characterized in that: The self-refinement module employs a boundary feature transformation mechanism, which enhances the consistency of features within the powder-spreading area and suppresses misidentification of the background area by introducing a boundary confidence map during feature propagation.

5. The online monitoring method for powder spreading quality based on image recognition according to claim 4, characterized in that: The weakly supervised semantic segmentation model also includes a cross-pixel dependency module, which calculates the similarity weights between pixels in the feature map and fuses multi-scale contextual information to further optimize the segmentation accuracy of the powder-spreading region.

6. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that, In step S3: The uniformity index is obtained by calculating the standard deviation and coefficient of variation of pixel gray values ​​within the powder-spreading area. When the coefficient of variation is greater than a preset threshold, the uniformity is deemed unqualified. The thickness distribution index is calculated by combining the three-dimensional point cloud data of the powder layer reconstructed by the structured light projection system or the binocular stereo vision system. Specifically, the range, mean and variance of the thickness distribution are analyzed. The warping defect is quantified by analyzing the change of the normal vector or the curvature of the height profile of the point cloud in the edge region. The extraction of the defect detection indicators includes: S301: Extract the area, location, and shape factors of foreign objects and holes; S302 distinguishes and quantifies the stringing in the inkjet area and the scratches in the non-inkjet area by calculating the aspect ratio of the connected regions, the length of the skeleton, and the direction. S303 quantifies the accumulation defects in inkjet and non-inkjet areas by analyzing the ratio of the protrusion height to the substrate area of ​​their local three-dimensional morphology. S304 quantifies the spherical defects in the inkjet area by calculating their roundness and height difference with the surrounding powder. S305 quantifies layer displacement defects by comparing the offset of the centroid or contour of feature regions (such as edges and markers) in a continuous layer segmentation map.

7. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that: The method sets up multiple monitoring nodes during the powder spreading process. Each node corresponds to a different position of the powder spreading equipment or a different powder spreading layer. The system processes the images of each node in parallel and judges the overall powder spreading quality by combining the results of each node.

8. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that, The method also includes a real-time alarm function: when an abnormality in powder spreading quality is detected, the system automatically triggers an audible and visual alarm and records the abnormal image, timestamp, process parameters and handling suggestions, storing them in the quality database for subsequent analysis and process optimization.

9. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that, The method also includes an online learning function: the system periodically adds manually verified defect-annotated images to the training set to incrementally train the weakly supervised semantic segmentation model, gradually improving the model's adaptability and detection accuracy under different materials and process conditions.

10. The online monitoring method for powder spreading quality based on image recognition according to claim 1, characterized in that: The method is applicable to at least one powder spreading process among selective laser melting, binder spraying, or powder pressing.

Images