A method and system for particle size monitoring for metal powder production

Abstract

The present application relates to the technical field of metal powder production monitoring, and discloses a kind of metal powder production granularity monitoring method and system, the method includes obtaining the synchronous imaging data of high-speed particle flow in vacuum air atomization powder production tower observation window, the imaging correction is carried out to the synchronous imaging data, and first particle image is obtained;The first particle image is input into the pre-trained particle artifact recognition neural network model to carry out particle artifact recognition, and reflection mark and trailing mark are obtained;According to the reflection mark and the trailing mark, the first particle image is corrected to artifact, and second particle image is obtained;According to the second particle image, contour extraction and particle size statistics are carried out, and in-situ granularity monitoring data are obtained.The present application can realize the in-situ accurate monitoring of high-speed particle granularity.

Description

Technical Field

[0001] This invention relates to the field of metal powder production monitoring technology, and in particular to a method and system for monitoring particle size in metal powder production. Background Technology

[0002] Currently, metal powders serve as fundamental materials in processes such as powder metallurgy, additive manufacturing, and precision forming. The particle size distribution of powder directly affects the density, flowability, and mechanical properties of the subsequently formed parts. Therefore, during the production of metal powders, it is necessary to continuously monitor changes in particle size and determine whether the powder production process is in a stable state based on the particle size distribution.

[0003] In existing technologies, particle size data of metal powder is typically obtained through offline sampling, sieving, or instrumental analysis. This method requires testing after the powder-making process or after interim sampling, resulting in a time lag between the test results and the actual production status, making it difficult to reflect particle size changes during particle formation within the vacuum atomization powder-making tower. With the application of image acquisition equipment and smart sensors in production sites, images of high-speed falling particles can be acquired through observation windows, and particle size can be analyzed based on image processing. However, during vacuum atomization powder-making, molten metal droplets break up and fall rapidly under the action of high-speed airflow. The observation window contains a large number of particles moving at high speeds, making particle images prone to localized reflections, motion blur, boundary adhesion, and morphological distortion. Existing image processing methods typically rely on ordinary grayscale segmentation or edge extraction steps, making it difficult to reliably distinguish individual particle boundaries from disturbed images. This leads to subsequent equivalent particle size conversion and particle size statistics deviating from the true particle state.

[0004] Existing technologies present the problem of difficulty in accurately monitoring the particle size of high-speed particles during the metal powder production process in situ. Summary of the Invention

[0005] This invention provides a method and system for monitoring particle size in metal powder production, in order to solve the problem in the prior art that it is difficult to accurately monitor the particle size of high-speed particles in the metal powder production process in situ.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for monitoring the particle size of metal powder production, comprising: Acquire synchronous imaging data of high-speed particle flow within the observation window of the vacuum atomizing powder tower, perform imaging correction on the synchronous imaging data, and obtain the first particle image; The first particle image is input into a pre-trained particle artifact recognition neural network model to identify particle artifacts, and reflective marks and trailing marks are obtained. Based on the reflective marker, reflective pixel regions are extracted from the first particle image. Neighborhood reconstruction is performed on the reflective pixel regions to obtain a reflective corrected image. Based on the shadow marker, shadow pixel regions are extracted from the reflective corrected image. Image restoration is performed on the shadow pixel regions to obtain a second particle image. The second particle image is segmented into a particle foreground to obtain a particle binary image. The particle contour is extracted from the particle binary image to obtain a particle contour set. The particle contour set is screened according to preset morphological quality conditions to obtain the target contour set; Based on the target profile set, particle size conversion is performed to obtain an equivalent particle size set. Particle size statistics are then performed on the equivalent particle size set to obtain in-situ particle size monitoring data.

[0007] In a second aspect, the present invention provides a particle size monitoring system for metal powder production, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention acquires synchronous imaging data of high-speed particle flow in the observation window of the vacuum atomization powder making tower, and performs imaging correction on the synchronous imaging data to form a first particle image suitable for subsequent identification. This enables the particle size monitoring process to be directly oriented towards the high-speed particle flow image at the powder making site, avoiding the monitoring lag problem caused by simply relying on offline sampling and detection, and providing a stable data foundation for in-situ particle size analysis in the metal powder production process.

[0009] (2) The present invention inputs the first particle image into a pre-trained particle artifact recognition neural network model to obtain reflective markers and trailing markers, and corrects the reflective pixel area and the trailing pixel area according to the reflective markers and trailing markers respectively, so that the contour interference caused by local reflection and motion trailing in the high-speed particle image is processed by partitioning, thereby reducing the influence of artifacts on the extraction of the true boundary of the particle and improving the reliability of subsequent particle contour recognition.

[0010] (3) The present invention obtains a set of particle contours by performing particle foreground segmentation and particle contour extraction on the second particle image, and obtains a set of target contours by screening according to preset morphological quality conditions. Then, the particle size conversion and particle size statistics are performed on the target contours to generate in-situ particle size monitoring data. This makes the particle size statistics based on the particle targets after artifact correction and contour screening, reduces the interference of abnormal contours on the equivalent particle size calculation, and improves the accuracy and usability of the metal powder particle size monitoring results. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a particle size monitoring method for metal powder production provided in the first embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for monitoring particle size in metal powder production, comprising the following steps: S1, acquire synchronous imaging data of high-speed particle flow within the observation window of the vacuum atomizing powder tower, perform imaging correction on the synchronous imaging data, and obtain the first particle image; S2, input the first particle image into a pre-trained particle artifact recognition neural network model to perform particle artifact recognition and obtain reflective marks and trailing marks; S3, extract the reflective pixel region from the first particle image according to the reflective mark, perform neighborhood reconstruction on the reflective pixel region to obtain the reflective correction image, extract the trailing pixel region from the reflective correction image according to the trailing mark, perform image restoration on the trailing pixel region to obtain the second particle image; S4, perform particle foreground segmentation on the second particle image to obtain a particle binary image, and extract particle contours from the particle binary image to obtain a particle contour set; S5, perform contour filtering on the particle contour set according to preset morphological quality conditions to obtain the target contour set; S6. Based on the target profile set, the particle size is converted to obtain an equivalent particle size set. The equivalent particle size set is then statistically analyzed to obtain in-situ particle size monitoring data.

[0014] In step S1, synchronous imaging data of the high-speed particle flow within the observation window of the vacuum atomizing powder-making tower is acquired, and imaging correction is performed on the synchronous imaging data to obtain a first particle image, including: A synchronous trigger signal is sent to the pulse illumination source and the image acquisition unit to acquire synchronous imaging data of the high-speed particle flow within the observation window; The synchronous imaging data is subjected to brightness equalization correction and geometric distortion correction according to preset optical correction parameters to obtain the first particle image; The pixel size calibration coefficient corresponding to the first particle image is determined according to the preset pixel calibration relationship.

[0015] In one implementation, the synchronized imaging data is a raw image matrix formed by the image acquisition unit when the high-speed particle flow within the observation window is illuminated by a pulsed illumination source. The raw image matrix includes the grayscale value corresponding to each pixel position and the synchronization trigger time. The pulsed illumination source and the image acquisition unit receive the same synchronization trigger signal. After the synchronization trigger signal arrives, the pulsed illumination source emits a short pulse of illumination. Under the control of the same synchronization trigger signal, the image acquisition unit opens the exposure window and acquires an image of the high-speed particle flow within the observation window, thus obtaining the synchronized imaging data.

[0016] It should be noted that the trigger period of the synchronization trigger signal can be set to 2 milliseconds, corresponding to a sampling frequency of 500 Hz; the pulse width of the pulse illumination source can be set to 10 nanoseconds; and the exposure time of the image acquisition unit can be set to 1 microsecond. These values ​​correspond to the continuous sampling interval, illumination duration, and image acquisition time window of the high-speed particle stream, respectively. Those skilled in the art can replace these parameters with similar ones based on the particle stream velocity and the frame rate of the image acquisition unit. The synchronization trigger signal is output by an industrial trigger control unit, which is connected to both the trigger terminals of the pulse illumination source and the image acquisition unit.

[0017] The preset optical correction parameters include dark field offset parameters, brightness gain parameters, and geometric distortion parameters. The dark field offset parameters are obtained by blocking the pulsed illumination source and acquiring multiple frames of dark field images. The average grayscale value of the same pixel location in the multiple dark field images is taken to obtain the dark field offset parameter corresponding to each pixel location. The brightness gain parameters are obtained by placing a uniform reflection calibration plate at the observation window position and acquiring multiple frames of uniform illumination images. The response grayscale value of each pixel location is normalized to obtain the brightness gain parameter corresponding to each pixel location. The geometric distortion parameters are obtained by placing a grid calibration plate at the observation window position and acquiring grid calibration images. A distortion correction mapping relationship is established based on the pixel location of the grid intersection point in the image and the actual location of the grid intersection point on the calibration plate to obtain the geometric distortion parameters.

[0018] In this embodiment, brightness equalization correction and geometric distortion correction are performed on the synchronous imaging data according to the preset optical correction parameters. During brightness equalization correction, the dark field offset parameter corresponding to the pixel position is first subtracted from the grayscale value of each pixel in the synchronous imaging data. Then, the pixel grayscale is adjusted according to the brightness gain parameter of the corresponding pixel position to obtain a brightness equalization image. During geometric distortion correction, the pixel positions in the brightness equalization image are mapped to the corrected regular pixel coordinates according to the geometric distortion parameters, and the mapped pixel grayscale is resampled to obtain the first grain image.

[0019] It is worth noting that the preset pixel calibration relationship is the correspondence between the pixel size in the first particle image and the actual physical size. This preset pixel calibration relationship is established using a standard scale calibration plate, which is placed at the same focal plane as the high-speed particle flow imaging position. After the image acquisition unit acquires an image of the standard scale calibration plate, it reads the pixel spacing corresponding to adjacent scribe lines on the standard scale calibration plate and divides the actual distance between adjacent scribe lines by the pixel spacing to obtain the pixel size calibration coefficient. This pixel size calibration coefficient is associated with and stored in the first particle image, and is called upon for subsequent particle size conversion.

[0020] For example, in a calibration process, the actual distance between adjacent lines on the standard scale calibration plate is 100 micrometers, and the pixel spacing between adjacent lines in the image is 40 pixels. Dividing 100 micrometers by 40 yields a pixel size calibration coefficient of 2.5 micrometers per pixel. During synchronous acquisition, the pulse width of the pulsed illumination source is set to 10 nanoseconds, the exposure time of the image acquisition unit is set to 1 microsecond, and the image acquisition unit outputs synchronous imaging data with a resolution of 1024 x 1024 pixels. For higher resolution cameras, the trigger frequency can be reduced to match the bandwidth. After performing brightness equalization correction and geometric distortion correction on the synchronous imaging data, a first particle image is obtained, and 2.5 micrometers per pixel is used as the pixel size calibration coefficient corresponding to the first particle image.

[0021] In step S2, the first particle image is input into a pre-trained particle artifact recognition neural network model for particle artifact recognition, resulting in reflective markers and trailing markers, including: The first particle image is input into the particle artifact recognition neural network model to extract particle artifact features and obtain a particle artifact feature map. Reflection region identification is performed based on the particle artifact feature map to obtain a reflection probability map; and ghosting region identification is performed based on the particle artifact feature map to obtain a ghosting probability map. The reflection probability map and the ghosting probability map are marked with pixels according to the preset artifact judgment threshold to obtain reflection mark and ghosting mark.

[0022] In one implementation, the grain artifact recognition neural network model is a pre-trained image segmentation neural network model. The input to the grain artifact recognition neural network model is a first grain image, and the output is a reflection probability map and a ghosting probability map corresponding to the size of the first grain image. The pixel values ​​in the reflection probability map represent the probability that a corresponding pixel in the first grain image belongs to a reflection artifact, and the pixel values ​​in the ghosting probability map represent the probability that a corresponding pixel in the first grain image belongs to a ghosting artifact. The reflection marker is a binary marker map formed by pixel-marking the reflection probability map, and the ghosting marker is a binary marker map formed by pixel-marking the ghosting probability map.

[0023] It should be noted that the particle artifact feature map is an intermediate feature map formed by the particle artifact recognition neural network model after performing multi-layer feature extraction on the first particle image. The particle artifact feature map includes particle brightness distribution features, particle edge direction features, and particle wake extension features. The particle brightness distribution features reflect the gray-level concentration state of locally bright areas on the particle surface; the particle edge direction features reflect the direction of gray-level change near the particle boundary; and the particle wake extension features reflect the morphology of areas where gray-level continuously decays along the particle's falling direction. The particle artifact recognition neural network model performs reflective area identification and trailing area identification based on the particle artifact feature map, obtaining a reflective probability map and a trailing probability map.

[0024] The particle artifact recognition neural network model can adopt an encoder-decoder structure. The encoder includes four feature extraction blocks, each of which sequentially includes a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional kernel size of the four feature extraction blocks is 3x3, the stride is 1, the padding is 1, and the number of output channels is 32, 64, 128, and 256, respectively. A max-pooling layer is placed between adjacent feature extraction blocks, with a window size of 2x2 and a stride of 2. The decoder includes four upsampling blocks, each of which includes a bilinear interpolation upsampling layer and a convolutional layer. The bilinear interpolation upsampling layer magnifies the height and width of the input feature map to twice its original size. The convolutional kernel size of the convolutional layer is 3x3, the stride is 1, and the padding is 1. A convolutional output layer is placed at the end of the decoder, with a convolutional kernel size of 1x1 and two output channels. The two output channels correspond to the reflection probability map and the shadow probability map, respectively. The pixel values ​​of each output channel fall between 0 and 1 after Sigmoid normalization mapping. The batch size, learning rate, optimizer, number of training rounds and stopping conditions are all set using the conventional training parameters of image segmentation networks.

[0025] It is worth noting that the training samples for the particle artifact recognition neural network model are obtained by pixel-level annotation of historical particle images. These historical particle images are particle images collected within the observation window of a vacuum atomization powder-making tower, and the number of training samples can be set to no less than 3000. These historical particle images cover imaging states under different atomizing gas pressures, different melt flow rates, different particle densities, and different exposure intensities. When annotating these historical particle images, locally bright areas with a grayscale value of no less than 240 located within the main particle area are labeled as reflective artifact areas; areas extending along the particle's falling direction with an extension length of no less than 5 pixels and a grayscale value gradually decreasing along the extension direction are labeled as motion blur artifact areas. If a pixel simultaneously meets the labeling conditions for both reflective artifact areas and motion blur artifact areas, then that pixel is labeled as a reflective artifact area. The grayscale value of no less than 240 is used to limit locally overly bright pixels in the 8-bit image, and the extension length of no less than 5 pixels is used to distinguish between motion blur artifacts and short-distance grayscale transitions at particle edges.

[0026] After annotation, the annotated historical particle images are divided into training, validation, and test sets, with a ratio of 8:1:1. To reduce single-annotation bias, historical particle images can be annotated pixel-level by two annotators. When the pixel overlap ratio of the two annotation results in the same artifact category is not less than 0.85, the corresponding annotation result is included in the training samples; when the pixel overlap ratio is less than 0.85, the corresponding historical particle image is re-annotated. The 0.85 is used to filter training annotation samples with insufficient consistency.

[0027] The training process of the particle artifact recognition neural network model uses pixel-level cross-entropy loss as the optimization objective. During training, historical particle images are input into the model, which outputs a training reflection probability map and a training ghosting probability map. Pixel-level differences are calculated between the training reflection probability map and the labeled reflection artifact region, and pixel-level differences are also calculated between the training ghosting probability map and the labeled ghosting artifact region. The network parameters of the model are updated based on these two difference results. The batch size during training can be set to 8, the initial learning rate to 0.001, the Adam optimizer to be used, the weight decay coefficient to be set to 0.0001, and the maximum number of training epochs to be set to 100. The training stopping condition is set to a decrease in validation set loss of less than 0.001 for five consecutive training epochs; this value is used to determine whether the training process of the particle artifact recognition neural network model has reached a stable state. After training, the fixed network parameters are stored as the pre-trained particle artifact recognition neural network model.

[0028] It should be noted that the preset artifact detection threshold can be set to 0.5, based on a typical value selected after balancing the accuracy and recall of reflection and ghosting recognition. This value is used to convert the pixel values ​​in the probability map into binary labels. When the pixel value of a pixel in the reflection probability map is greater than or equal to 0.5, the corresponding position of that pixel in the reflection label is marked as a reflection pixel; when the pixel value of a pixel in the ghosting probability map is greater than or equal to 0.5, the corresponding position of that pixel in the ghosting label is marked as a ghosting pixel; when the corresponding pixel value is less than 0.5, the corresponding position is marked as a non-artifact pixel. If the same pixel is marked as both a reflection pixel and a ghosting pixel, the reflection label is used as the priority label for that pixel, and the neighborhood reconstruction of the reflection pixel area is performed first, followed by the image restoration of the ghosting pixel area.

[0029] For example, when a first grainy image with a resolution of 1024 x 1024 pixels is input into a pre-trained grain artifact recognition neural network model, the model outputs a reflection probability map and a ghosting probability map, both also 1024 x 1024 pixels. If a pixel in the reflection probability map has a value of 0.87, and the same pixel in the ghosting probability map has a value of 0.12, then the corresponding position of that pixel in the reflection marker is marked as a reflective pixel, and the corresponding position in the ghosting marker is marked as a non-ghosting pixel. Similarly, if another pixel in the ghosting probability map has a value of 0.76, and the same pixel in the reflection probability map has a value of 0.18, then the corresponding position of that other pixel in the ghosting marker is marked as a ghosting pixel, and the corresponding position in the reflection marker is marked as a non-reflective pixel.

[0030] In step S3, reflective pixel regions are extracted from the first particle image based on the reflective markers, and neighborhood reconstruction is performed on the reflective pixel regions to obtain a reflective corrected image. Motion shadow pixel regions are extracted from the reflective corrected image based on the motion shadow markers, and image restoration is performed on the motion shadow pixel regions to obtain a second particle image.

[0031] Specifically, the process involves extracting reflective pixel regions from the first particle image based on the reflective markers, performing neighborhood reconstruction on the reflective pixel regions to obtain a reflective-corrected image, including: The coordinates of the reflective pixels in the first particle image are located according to the reflective markers; Based on the coordinates of the reflective pixels, a connected region analysis is performed to obtain the reflective pixel region; Normal neighbor pixels are extracted from the periphery of the reflective pixel region, and grayscale is filled into the reflective pixel region based on the normal neighbor pixels to obtain a reflection correction image.

[0032] In one implementation, the reflective marker is a binary marker image with the same pixel size as the first particle image. The reflective pixels in the reflective marker represent the pixel positions in the first particle image that are identified as reflective artifacts by the particle artifact recognition neural network model. When locating the coordinates of reflective pixels in the first particle image based on the reflective marker, the reflective marker is traversed in pixel row and column order, and the positions marked as reflective pixels are recorded as reflective pixel coordinates.

[0033] It should be noted that when performing connected component analysis based on the reflected pixel coordinates, the eight-neighbor connectivity rule is used to merge the reflected pixel coordinates. The eight-neighbor connectivity rule means that a reflected pixel coordinate is grouped into the same connected component along with the reflected pixel coordinates in its adjacent positions above, below, left, right, upper left, upper right, lower left, and lower right. After merging all reflected pixel coordinates, each connected component is defined as a reflected pixel region.

[0034] When extracting normal neighboring pixels from the periphery of the reflective pixel region, a preset neighborhood width is extended outward based on the boundary of the reflective pixel region to obtain the neighborhood extraction range. The preset neighborhood width can be set to 5 pixels, which is used to limit the local image range around the reflective pixel region that participates in grayscale filling. Subsequently, the pixels in the neighborhood extraction range are screened for normality. Pixels that are not marked as reflective pixels by the reflective marker, are not marked as ghosting pixels by the ghosting marker, and whose grayscale value is within 30 grayscale levels above and below the median grayscale value of non-artifact pixels in the neighborhood extraction range are determined as normal neighboring pixels. If the grayscale value corresponding to the median value deviates significantly from the grayscale value of normal particle surfaces, for example, the average grayscale value of pixels directly adjacent to the periphery of the reflective region differs by more than 50 grayscale levels, then the average grayscale value of pixels on the outer boundary of the reflective region is used as the filling reference. The median grayscale value of the non-artifact pixel is obtained by sorting the grayscale values ​​of pixels in the neighborhood extraction range that are not marked by the reflective marker and the trailing marker; the 30 grayscale levels are used to exclude other bright pixels at the boundaries of other particles and abnormally dark noise pixels in the neighborhood.

[0035] If the number of normal neighboring pixels within the extracted neighborhood is less than a preset lower limit for the number of neighboring pixels, the neighborhood width is gradually increased by 5 pixels until the number of normal neighboring pixels reaches the preset lower limit, or the neighborhood width reaches a preset maximum neighborhood width. The preset lower limit for the number of neighboring pixels can be set to 16 pixels, which corresponds to the minimum number of samples required to form a stable grayscale estimate during local grayscale filling. The preset maximum neighborhood width can be set to 15 pixels, which limits the neighborhood extraction range from exceeding the current particle's local structure by too much. If the number of normal neighboring pixels is still less than 16 pixels when the neighborhood width reaches 15 pixels, grayscale filling is performed using the currently obtained normal neighboring pixels, and the reflective pixel area is marked as a low-neighborhood support area. The neighborhood width of 5 pixels, the lower limit of 16 pixels, and the maximum neighborhood width of 15 pixels are determined based on the typical width of the particle boundary transition area and the minimum number of samples required for stable grayscale estimation.

[0036] When performing grayscale filling on the reflective pixel region based on the normal neighbor pixels, for each pixel to be filled within the reflective pixel region, the nearest normal neighbor pixel is selected; when there are multiple nearest normal neighbor pixels, the average grayscale value of the multiple normal neighbor pixels is calculated; the grayscale value of the nearest normal neighbor pixel or the average grayscale value is written to the position of the pixel to be filled. After completing grayscale filling on all pixels to be filled, a reflection-corrected image is obtained. The reflection-corrected image has the same pixel coordinate system and image size as the first particle image.

[0037] It is worth noting that the reflection-corrected image inherits the grayscale values ​​of non-reflective pixels in the first grain image and replaces the grayscale values ​​in the reflective pixel region. The trailing mark has the same pixel size as the first grain image, therefore the trailing mark can be directly mapped to the corresponding pixel position in the reflection-corrected image.

[0038] Specifically, the process of extracting the ghosting pixel region from the reflection-corrected image based on the ghosting marker, and performing image restoration on the ghosting pixel region to obtain a second grainy image includes: The motion blur pixel coordinates are located in the reflection-corrected image based on the motion blur markers; The extension direction and extension length of the trailing pixel region are determined based on the trailing pixel coordinates and the neighborhood grayscale distribution. The trailing pixel region is directionally grayscale corrected according to the extension direction and the extension length to obtain a second particle image.

[0039] In one implementation, the ghosting marker is a binary marker image with the same pixel size as the reflection-corrected image. The ghosting pixels in the ghosting marker represent the pixel positions in the reflection-corrected image that are identified as ghosting artifacts by the grain artifact recognition neural network model. When locating the ghosting pixel coordinates in the reflection-corrected image based on the ghosting marker, the ghosting marker is traversed in pixel row and column order, and the positions marked as ghosting pixels are recorded as ghosting pixel coordinates; before ghosting restoration, pixels in the ghosting marker that overlap with the reflection marker need to be set to zero.

[0040] It should be noted that the neighborhood grayscale distribution refers to the grayscale change state of pixels surrounding the trailing pixel coordinates in the reflection-corrected image. When determining the extension direction and length of the trailing pixel region based on the trailing pixel coordinates and the neighborhood grayscale distribution, the eight-neighbor connectivity rule is first used to merge adjacent trailing pixel coordinates into a trailing candidate region. Then, the grayscale values ​​of the trailing candidate region and its surrounding adjacent pixels are read. For each trailing candidate region, a sequence of continuously decreasing grayscale pixels is statistically analyzed along the vertical, horizontal, left diagonal, and right diagonal directions. The direction corresponding to the sequence with the largest number of continuously decreasing grayscale pixels is determined as the extension direction, and the number of pixels is determined as the extension length. The continuously decreasing grayscale pixel sequence refers to a sequence of adjacent pixels arranged in the same direction with progressively decreasing grayscale values. The grayscale difference between adjacent pixels can be set to be greater than or equal to three grayscale levels. This value is used to distinguish between trailing grayscale changes and image acquisition noise fluctuations, and is set according to the typical fluctuation range of image acquisition noise.

[0041] It is worth noting that when performing directional grayscale correction on the trailing pixel region based on the extension direction and the extension length, the trailing candidate region is determined as the trailing pixel region. The particle body side and the trailing end side in the trailing pixel region are identified along the extension direction. The particle body side is the side with a higher grayscale value and closer to the particle outline, while the trailing end side is the side with a lower grayscale value and farther from the particle outline. The average grayscale value of the background pixels surrounding the trailing pixel region is read, and the pixel grayscale value of the trailing end side is replaced with this average grayscale value. Then, along the extension direction, from the particle body side boundary to the trailing end side boundary, the pixel grayscale is linearly interpolated. The grayscale value of the body side boundary remains unchanged, while the grayscale value of the end side boundary is replaced with the average grayscale value of the surrounding background pixels. The middle pixels transition linearly according to the distance ratio. After completing the directional grayscale correction for all trailing pixel regions, a second particle image is obtained. The second particle image inherits the grayscale values ​​of the non-trailing pixels in the reflection-corrected image and replaces the grayscale values ​​in the trailing pixel region.

[0042] For example, after expanding a reflective pixel region outward by 5 pixels, there are 128 pixels in the neighborhood extraction range. Among them, 18 pixels are marked as reflective pixels by reflective markers, 9 pixels are marked as motion blur pixels by motion blur markers, and the remaining 101 pixels are used to calculate the median grayscale. If the median grayscale of the non-artifact pixels is 142, then pixels with grayscale values ​​between 112 and 172 that are not marked by reflective or motion blur markers are determined as normal neighbor pixels. If the number of normal neighbor pixels is 64, then the 64 normal neighbor pixels are used to fill the reflective pixel region with grayscale to obtain a reflection-corrected image.

[0043] For example, a certain motion blur candidate region includes 32 motion blur pixel coordinates. The number of pixels with continuously decreasing grayscale values ​​is 11 in the vertical direction, 4 in the horizontal direction, 6 in the left diagonal direction, and 5 in the right diagonal direction. Therefore, the vertical direction is determined as the extension direction, and the 11 pixels are determined as the extension length. If the average grayscale value of the background pixels surrounding the motion blur pixel region is 28, and the grayscale value of the pixels at the end of the motion blur is 46, then the grayscale value of the pixels at the end of the motion blur is adjusted to 28, and a linear transition adjustment is performed on the pixel grayscale values ​​between the particle body side and the end of the motion blur to obtain the second particle image.

[0044] In step S4, the second particle image is segmented into a particle foreground to obtain a particle binary image. The particle contour is extracted from the particle binary image to obtain a particle contour set.

[0045] Specifically, the second particle image is segmented into a particle foreground to obtain a particle binary image, including: Local sharpness evaluation is performed on the second particle image to obtain the image sharpness distribution; Based on preset sharpness conditions, blurry areas are removed from the image sharpness distribution to obtain the image region to be segmented; Threshold segmentation is performed on the image region to be segmented to obtain a particle foreground region and a background region. Then, a binary mapping is performed on the particle foreground region and the background region to obtain a particle binary map.

[0046] In one implementation, the image sharpness distribution is a set of sharpness values ​​corresponding to each local window in the second particle image. When evaluating the local sharpness of the second particle image, the image is divided into multiple local windows according to a preset window size, and the grayscale change intensity is calculated for each local window. The preset window size can be set to 16 by 16 pixels, which is used to limit the local image range for a single sharpness evaluation and is determined based on the typical pixel diameter of the particle in the image. The grayscale change intensity is obtained by calculating the sum of the absolute values ​​of the grayscale differences of all horizontally adjacent pixels and vertically adjacent pixels within the local window, and then dividing by the total number of adjacent pixels. A larger grayscale change intensity indicates a sharper particle boundary within the local window; a smaller grayscale change intensity indicates a more blurred particle boundary within the local window.

[0047] It should be noted that the preset sharpness conditions include a lower limit value for sharpness. This lower limit value can be set to 8 gray levels, based on the statistical lower limit of the local gray-level change intensity of the focused area. This value is used to distinguish between low-contrast blurred areas and segmentable image areas in an 8-bit grayscale image. When removing blurred areas from the image sharpness distribution according to the preset sharpness conditions, local windows with gray-level change intensity less than 8 gray levels are identified as blurred areas, and local windows with gray-level change intensity greater than or equal to 8 gray levels are identified as image areas to be segmented. If multiple adjacent local windows meet the sharpness conditions, the adjacent local windows are merged into a continuous image area to be segmented. The area of ​​the removed blurred areas typically does not exceed 5% of the total image area. If it exceeds 5%, the lower limit value for sharpness is lowered to 6 gray levels for re-segmentation, or a prompt is made to re-acquire the image.

[0048] It is worth noting that the local sharpness evaluation is part of the image region selection process before grain foreground segmentation. The image region to be segmented is the sharp image region retained in the second grain image, and the blurred region does not participate in the binary mapping between the grain foreground and background regions. When using the above method, the actual object of grain foreground segmentation is the image region to be segmented, rather than the complete second grain image.

[0049] When performing threshold segmentation on the image region to be segmented, pixel grayscale values ​​are read from the image region to be segmented, and the segmentation threshold is determined based on the grayscale histogram of the image region to be segmented. The segmentation threshold can be determined by the valley between the background grayscale peak and the grain grayscale peak in the grayscale histogram; when no obvious bimodal peak is formed, the Otsu method is used to adaptively determine the segmentation threshold. Pixels with grayscale values ​​greater than or equal to the segmentation threshold are determined as grain foreground regions, and pixels with grayscale values ​​less than the segmentation threshold are determined as background regions. When performing binary mapping on the grain foreground regions and the background regions, the pixels in the grain foreground regions are assigned a value of 255, and the pixels in the background regions are assigned a value of 0, resulting in a grain binary image; the blurred region is assigned a value of 0 in the grain binary image.

[0050] It is worth noting that the particle binary image and the second particle image share the same pixel coordinate system. In the particle binary image, a pixel value of 255 represents the particle foreground, and a pixel value of 0 represents the background or blurred area. The particle binary image serves as the input for particle contour extraction.

[0051] Specifically, particle contours are extracted from the binary particle image to obtain a particle contour set, including: Perform connected component analysis on the binary graph of the particles to obtain the connected regions of the particles; Extract the boundaries of the connected regions of the particles to obtain the closed particle profile; The closed particle contours are noise-removed according to the preset contour area range to obtain a particle contour set.

[0052] In one implementation, the particle connected region is a set of interconnected foreground pixels in the particle binary graph. When performing connected component analysis on the particle binary graph, the graph is scanned according to the eight-neighbor connectivity rule, and particle foreground pixels with an interconnected pixel value of 255 are grouped into the same particle connected region. Each particle connected region corresponds to a candidate particle target or a candidate adhered particle target.

[0053] It should be noted that when extracting the boundaries of the particle connected regions, the particle foreground pixels adjacent to the background region are found in the particle connected regions, and the position coordinates of the adjacent boundary pixels are recorded sequentially in a clockwise direction to obtain a closed particle contour. If the boundary coordinates of a certain particle connected region are connected end to end, the boundary coordinates are determined as a closed particle contour; if the boundary coordinates do not form a connected structure, the particle connected region is deleted from the subsequent contour statistics.

[0054] The preset contour area range includes a lower limit and an upper limit. The lower limit can be set to 30 pixels, which is used to remove small contours formed by isolated noise points. The upper limit can be set to 8000 pixels, which is used to remove abnormal connected contours that significantly exceed the imaging range of a single particle. The lower limit of 30 pixels corresponds to the imaging area of ​​the smallest resolvable particle, and the upper limit of 8000 pixels corresponds to the maximum imaging area of ​​a typical single particle. When performing noise contour removal on the closed particle contours according to the preset contour area range, the number of particle foreground pixels contained in each closed particle contour is counted. Closed particle contours with less than 30 pixels or more than 8000 pixels of particle foreground pixels are deleted, and the remaining closed particle contours are combined into a particle contour set.

[0055] For example, the second particle image is divided into 16x16 pixel local windows. If the average gray-level difference between adjacent pixels within a local window is 12 gray levels, then the local window meets the preset sharpness condition and is included in the image region to be segmented. If the average gray-level difference between adjacent pixels within another local window is 5 gray levels, then the other local window is determined to be a blurred region and assigned a value of 0 in the particle binary image. After thresholding the image region to be segmented, a particle binary image is obtained. After performing connected component analysis on the particle binary image, 54 particle connected regions are obtained, of which 2 particle connected regions have a contour area of ​​less than 30 pixels, and 1 particle connected region has a contour area of ​​more than 8000 pixels. After deleting the closed particle contours corresponding to the above 3 particle connected regions, a particle contour set containing 51 closed particle contours is obtained.

[0056] In step S5, the particle contour set is screened according to preset morphological quality conditions to obtain a target contour set, including: Calculate the contour area and contour perimeter of each particle contour in the particle contour set; The roundness value is calculated based on the outline area and the outline perimeter, and the roundness value is the product of four times pi and the outline area divided by the square of the outline perimeter. The particle contours whose roundness values ​​meet the preset morphological quality conditions are determined as target contours, thus obtaining a target contour set.

[0057] In one implementation, the particle contour set includes multiple closed particle contours, each particle contour consisting of continuously arranged boundary pixel coordinates. The contour area is the number of particle foreground pixels contained within the particle contour, and the contour perimeter is the boundary length of the particle contour. When calculating the contour area, the particle foreground pixels within the particle contour are counted, and the count result is determined as the contour area. When calculating the contour perimeter, the connection lengths between adjacent boundary pixels are counted according to the arrangement order of the boundary pixel coordinates of the particle contour, and all connection lengths are added together to obtain the contour perimeter.

[0058] It should be noted that the roundness value is a morphological parameter characterizing how close the particle outline is to a circle. A roundness value closer to 1 indicates that the corresponding particle outline is closer to a circle; a smaller roundness value indicates that the corresponding particle outline has tailing, adhesion, flattening, or boundary damage. To calculate the roundness value, the outline area is multiplied by four times pi, and then the product is divided by the square of the outline perimeter. Since the outline area and the outline perimeter use the same pixel coordinate scale, the roundness value is dimensionless.

[0059] The preset morphological quality conditions include a lower limit for roundness. This lower limit can be set to 0.85, a value used to distinguish between particle outlines that approximate a spherical projection and those with significant morphological distortion; it is an empirical statistical threshold for the roundness of spherical particle projections in metal powder. Particle outlines with a roundness value greater than or equal to 0.85 are identified as target outlines, while those with a roundness value less than 0.85 are removed from the particle outline set. After screening all particle outlines, the remaining target outlines are combined into a target outline set.

[0060] It is worth noting that if the particle contour corresponds to adhered particles, severely trailing particles, or particles with broken defocus boundaries, the contour perimeter is usually larger relative to the contour area, and the calculated roundness value will decrease. When screening contours using the preset morphological quality conditions, the boundary pixel coordinates of the target contours are not changed; only particle contours that do not meet the preset morphological quality conditions are deleted from the particle contour set. The target contour set serves as the input for subsequent particle size conversion.

[0061] For example, one particle contour has an area of ​​450 pixels and a perimeter of 85 pixels. In the calculation, 450 is multiplied by four times pi and then divided by the square of 85, resulting in a roundness value of approximately 0.78. Since 0.78 is less than the lower limit of roundness of 0.85, this particle contour is removed from the particle contour set. Another particle contour has an area of ​​520 pixels and a perimeter of 81 pixels. The calculated roundness value is approximately 1.00, and this other particle contour is identified as the target contour. After filtering all particle contours, the target contour set is obtained.

[0062] In step S6, particle size conversion is performed based on the target profile set to obtain an equivalent particle size set. Particle size statistics are then performed on the equivalent particle size set to obtain in-situ particle size monitoring data, including: The outline pixel area of ​​each target outline in the target outline set is calculated. Based on the outline pixel area and the preset pixel size calibration coefficient, an equal area circle conversion is performed to obtain the equivalent particle size set; The equivalent particle size set is subjected to frequency statistics according to the preset particle size interval rules to obtain the particle size frequency distribution; Based on the particle size frequency distribution, the median particle size, particle size distribution width, and fine powder yield are calculated to generate in-situ particle size monitoring data.

[0063] In one implementation, the target contour set includes multiple target contours that satisfy the preset morphological quality conditions. The contour pixel area is the number of particle foreground pixels contained within the target contour. When calculating the contour pixel area of ​​each target contour in the target contour set, the internal region of the target contour is determined according to the boundary coordinates of the target contour, and the particle foreground pixels in the internal region of the target contour are counted to obtain the contour pixel area corresponding to each target contour.

[0064] It should be noted that the preset pixel size calibration coefficient is the conversion relationship between the actual physical size and pixel size of the first particle image. The preset pixel size calibration coefficient is determined in step S1 according to the preset pixel calibration relationship and maintains the same pixel coordinate system as the first particle image, the second particle image, the particle binary image, and the target contour set. When performing the equal-area circle conversion based on the contour pixel area and the preset pixel size calibration coefficient, the contour pixel area is first converted to the pixel diameter corresponding to the circular projection with the same area. Then, the pixel diameter is multiplied by the preset pixel size calibration coefficient to obtain the equivalent particle size of the corresponding target contour. After completing the equal-area circle conversion for all target contours in the target contour set, the equivalent particle size set is obtained.

[0065] The preset particle size interval rule is a rule for dividing the equivalent particle size into multiple continuous particle size ranges, which is set according to the commonly used grading standards for metal powder particle size detection. The preset particle size interval rule can be set to less than 15 micrometers, 15 micrometers to 30 micrometers, 30 micrometers to 45 micrometers, 45 micrometers to 53 micrometers, 53 micrometers to 75 micrometers, and greater than 75 micrometers, according to the commonly used intervals for metal powder particle size detection. The boundaries of adjacent intervals in the particle size interval rule do not overlap, the upper interval does not include the right endpoint, and the lower interval includes the left endpoint. When performing frequency statistics on the equivalent particle size set according to the preset particle size interval rule, the equivalent particle size in the equivalent particle size set is read one by one, each equivalent particle size is assigned to the corresponding particle size interval, and the number of equivalent particle sizes in each particle size interval is counted to obtain the particle size frequency distribution.

[0066] It is worth noting that the median particle size is the equivalent particle size corresponding to the cumulative frequency reaching 50% of the total equivalent particle size. When calculating the median particle size, the equivalent particle size set is arranged in ascending order of equivalent particle size, and the equivalent particle size corresponding to the cumulative frequency reaching half of the total equivalent particle size is read and determined as the median particle size. The particle size distribution width is calculated based on the larger cumulative particle size, the smaller cumulative particle size, and the median particle size. The smaller cumulative particle size is the equivalent particle size corresponding to the cumulative frequency reaching 10% of the total equivalent particle size, and the larger cumulative particle size is the equivalent particle size corresponding to the cumulative frequency reaching 90% of the total equivalent particle size; where the 10% and 90% quantiles are the conventional statistical limits for the particle size distribution width of metal powder. During calculation, the larger cumulative particle size is subtracted from the smaller cumulative particle size, and the difference is divided by the median particle size to obtain the particle size distribution width. The fine powder yield is the proportion of the number of equivalent particle sizes smaller than the preset fine powder particle size limit to the total number of equivalent particle sizes. The preset fine powder particle size limit can be set to 30 micrometers. This value is used to distinguish between fine powder particles and non-fine powder particles, and it is set according to the general particle size definition of fine powder in the powder metallurgy industry.

[0067] When generating in-situ particle size monitoring data, the particle size frequency distribution, the particle size median, the particle size distribution width, and the fine powder yield are packaged at the same collection time to obtain in-situ particle size monitoring data. The in-situ particle size monitoring data includes the collection time, particle size interval, frequency corresponding to each particle size interval, particle size median, particle size distribution width, fine powder yield, and equivalent particle size count.

[0068] For example, the outline pixel area of ​​a target contour is 512 pixels, and the preset pixel size calibration coefficient is 2.5 micrometers per pixel. After converting 512 pixels into the pixel diameter corresponding to a circular projection with the same area, the pixel diameter is approximately 25.5 pixels. Multiplying 25.5 pixels by 2.5 micrometers per pixel, the equivalent particle size corresponding to the target contour is approximately 63.8 micrometers. If 48 equivalent particle sizes are obtained in one acquisition, of which 15 are smaller than 30 micrometers, the fine powder recovery rate is 31.25%. After arranging the 48 equivalent particle sizes in ascending order, the equivalent particle size at the 50% cumulative position is read as the particle size median, and the particle size distribution width is calculated by combining the equivalent particle sizes at the 10% and 90% cumulative positions, finally generating in-situ particle size monitoring data.

[0069] In summary, this invention acquires synchronous imaging data of high-speed particle flow within the observation window of a vacuum atomization powder production tower, and sequentially performs imaging correction, particle artifact recognition, neighborhood reconstruction of reflective pixel areas, image restoration of trailing pixel areas, particle foreground segmentation, particle contour extraction, contour screening, particle size conversion, and particle size statistics. This forms a complete processing flow from on-site image acquisition to in-situ particle size monitoring data output, enabling accurate in-situ monitoring of high-speed particle size during metal powder production.

[0070] A second embodiment of the present invention provides a particle size monitoring system for metal powder production, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0071] It should be noted that the particle size monitoring system for metal powder production provided in this embodiment of the invention is used to execute all the process steps of the particle size monitoring method for metal powder production described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0072] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A method for particle size monitoring in metal powder production, characterized in that, include: Acquire synchronous imaging data of high-speed particle flow within the observation window of the vacuum atomizing powder tower, perform imaging correction on the synchronous imaging data, and obtain the first particle image; The first particle image is input into a pre-trained particle artifact recognition neural network model to identify particle artifacts, and reflective marks and trailing marks are obtained. Based on the reflective marker, reflective pixel regions are extracted from the first particle image. Neighborhood reconstruction is performed on the reflective pixel regions to obtain a reflective corrected image. Based on the shadow marker, shadow pixel regions are extracted from the reflective corrected image. Image restoration is performed on the shadow pixel regions to obtain a second particle image. The second particle image is segmented into a particle foreground to obtain a particle binary image. The particle contour is extracted from the particle binary image to obtain a particle contour set. The particle contour set is screened according to preset morphological quality conditions to obtain the target contour set; Based on the target profile set, particle size conversion is performed to obtain an equivalent particle size set. Particle size statistics are then performed on the equivalent particle size set to obtain in-situ particle size monitoring data.

2. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The process of acquiring synchronous imaging data of the high-speed particle flow within the observation window of the vacuum atomizing powder preparation tower, and performing imaging correction on the synchronous imaging data to obtain a first particle image includes: A synchronous trigger signal is sent to the pulse illumination source and the image acquisition unit to acquire synchronous imaging data of the high-speed particle flow within the observation window; The synchronous imaging data is subjected to brightness equalization correction and geometric distortion correction according to preset optical correction parameters to obtain the first particle image; The pixel size calibration coefficient corresponding to the first particle image is determined according to the preset pixel calibration relationship.

3. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of inputting the first particle image into a pre-trained particle artifact recognition neural network model for particle artifact recognition to obtain reflective markers and trailing markers includes: The first particle image is input into the particle artifact recognition neural network model to extract particle artifact features and obtain a particle artifact feature map. Reflection region identification is performed based on the particle artifact feature map to obtain a reflection probability map; and ghosting region identification is performed based on the particle artifact feature map to obtain a ghosting probability map. The reflection probability map and the ghosting probability map are marked with pixels according to the preset artifact judgment threshold to obtain reflection mark and ghosting mark.

4. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of extracting reflective pixel regions from the first particle image based on the reflective markers, and performing neighborhood reconstruction on the reflective pixel regions to obtain a reflective-corrected image includes: The coordinates of the reflective pixels in the first particle image are located according to the reflective markers; Based on the coordinates of the reflective pixels, a connected region analysis is performed to obtain the reflective pixel region; Normal neighbor pixels are extracted from the periphery of the reflective pixel region, and grayscale is filled into the reflective pixel region based on the normal neighbor pixels to obtain a reflection correction image.

5. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of extracting the ghosting pixel region from the reflection-corrected image based on the ghosting mark, and performing image restoration on the ghosting pixel region to obtain the second grainy image includes: The motion blur pixel coordinates are located in the reflection-corrected image based on the motion blur markers; The extension direction and extension length of the trailing pixel region are determined based on the trailing pixel coordinates and the neighborhood grayscale distribution. The trailing pixel region is directionally grayscale corrected according to the extension direction and the extension length to obtain a second particle image.

6. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of performing particle foreground segmentation on the second particle image to obtain a particle binary image includes: Local sharpness evaluation is performed on the second particle image to obtain the image sharpness distribution; Based on preset sharpness conditions, blurry areas are removed from the image sharpness distribution to obtain the image region to be segmented; Threshold segmentation is performed on the image region to be segmented to obtain a particle foreground region and a background region. Then, a binary mapping is performed on the particle foreground region and the background region to obtain a particle binary map.

7. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of extracting particle contours from the binary particle image to obtain a particle contour set includes: Perform connected component analysis on the binary graph of the particles to obtain the connected regions of the particles; Extract the boundaries of the connected regions of the particles to obtain the closed particle profile; The closed particle contours are noise-removed according to the preset contour area range to obtain a particle contour set.

8. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The step of screening the particle contour set according to preset morphological quality conditions to obtain a target contour set includes: Calculate the contour area and contour perimeter of each particle contour in the particle contour set; The roundness value is calculated based on the outline area and the outline perimeter, and the roundness value is the product of four times pi and the outline area divided by the square of the outline perimeter. The particle contours whose roundness values ​​meet the preset morphological quality conditions are determined as target contours, thus obtaining a target contour set.

9. The particle size monitoring method for metal powder production according to claim 1, characterized in that, The process involves converting the particle size based on the target profile set to obtain an equivalent particle size set, and then performing particle size statistics on the equivalent particle size set to obtain in-situ particle size monitoring data, including: The outline pixel area of ​​each target outline in the target outline set is calculated. Based on the outline pixel area and the preset pixel size calibration coefficient, an equal area circle conversion is performed to obtain the equivalent particle size set; The equivalent particle size set is subjected to frequency statistics according to the preset particle size interval rules to obtain the particle size frequency distribution; Based on the particle size frequency distribution, the median particle size, particle size distribution width, and fine powder yield are calculated to generate in-situ particle size monitoring data.

10. A particle size monitoring system for metal powder production, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 9.

Images