A quartz crucible factory quality intelligent vision full inspection system

Abstract

The application discloses a kind of quartz crucible factory quality intelligent vision full inspection system, belong to quartz crucible quality detection technology.The application obtains the two-dimensional surface image and two-dimensional layered image of crucible by first equipment and second equipment;Data acquisition unit collects these data, for providing data basis for defect detection unit and micro-bubble detection unit.Because defect detection unit needs to detect multiple defect categories, therefore, the defect detection model is constructed according to the improved YOLOv11 model;And the defect detected by the bubble detection model of micro-bubble detection unit is only one kind, and the bubble detection model is constructed according to YOLOv11 model.To this end, while ensuring the detection accuracy, reduce the overall cost of system.The transparent layer bubble detection unit eliminates the repeated counting bubble by similar bubble fusion, to ensure the reliability of bubble quantity;Visual unit provides reliable basis for crucible quality evaluation by fusing physical parameters and visual data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of quartz crucible quality detection, and particularly relates to an intelligent visual full detection system for the factory quality of a quartz crucible. BACKGROUND

[0002] At present, the photovoltaic industry is developing rapidly, and the quality of quartz crucibles, which are the core instruments for drawing single crystal silicon rods, directly affects the quality of single crystal silicon and further influences the development of the entire photovoltaic industry chain. Quartz crucibles are usually made of high-purity quartz sand and have the advantages of high purity, high temperature resistance, and good thermal stability. It is divided into a transparent layer and an opaque layer from the inside out. The transparent layer is uniform and dense, and the surface is smooth, which can enhance the strength of the crucible and reduce the temperature of the inner surface.

[0003] In order to meet the requirements of the factory standard of high-quality quartz crucibles, the current industry adopts the "random sampling + manual visual inspection" mode: random sampling at the end of the production line, rotating the crucible under a strong light lamp by professional personnel, and identifying defects by experience. This has the following three defects: first, the sampling sample cannot cover the batch difference caused by the discrete nature of the thermal expansion coefficient; second, the missed detection rate of the human eye for microcracks below 50 microns increases with the fatigue index, making it difficult to meet the requirements of zero defects downstream; third, the manual recording of data is unstructured and difficult to trace process fluctuations. More importantly, the high light transmission of quartz material and the curved surface reflection result in a high false positive rate of traditional two-dimensional visual solutions.

[0004] Therefore, in the existing technical system, there is no full detection level intelligent visual system, forming an industry gap.

[0005] In addition, for the detection of the transparent layer of the quartz crucible, the current industry mainly uses strong backlight or transmission imaging to indirectly infer the relevant characteristics of the transparent layer by observing the bubble distribution. However, this method has obvious limitations and can only obtain two-dimensional information of the bubbles, and cannot accurately judge the integrity and three-dimensional thickness distribution of the transparent layer.

[0006] Since the existing detection can only obtain two-dimensional information of the crucible, the accuracy of the calculation of the gas content rate of the transparent layer and the calculation result of the thickness are unreliable, making it difficult to provide reliable basis for the quality evaluation of the quartz crucible. SUMMARY

[0007] The technical problem to be solved by the present application is to provide an intelligent visual full detection system for the factory quality of a quartz crucible, which is novel and reasonable in design and can comprehensively detect the crucible.

[0008] To solve the above technical problems, the technical scheme adopted by the present application is:

[0009] An intelligent visual full detection system for the factory quality of a quartz crucible, the system comprising:

[0010] A crucible inspection device includes a first device and a second device. The first device includes a first placement stage and an image acquisition component. The image acquisition component includes a camera robotic arm positioned in the middle of the first placement stage and a light source robotic arm fixed above the first placement stage. The camera robotic arm is used to hold an industrial camera, and the light source robotic arm is used to hold a strip light source device to acquire a two-dimensional surface image of the inner wall of the crucible. The second device includes a second placement stage and a video acquisition component. The video acquisition component includes a camera robotic arm fixed in the middle of the second placement stage and an adjustable focus light source device. The camera robotic arm is connected to a microbubble detection camera through a camera motion module. The microbubble detection camera and the adjustable focus light source device work together to acquire a two-dimensional layered image of the inner wall of the crucible at a set speed and frame rate for each frame. Both the first and second placement stages are used to support the crucible to be inspected with its opening facing downwards.

[0011] The data acquisition unit is used to acquire two-dimensional surface images and two-dimensional layered images obtained by the first device and the second device;

[0012] The defect detection unit is used to identify the defect category and detection box location in a two-dimensional surface image through a defect detection model. The defect detection model is built based on an improved YOLOv11 model. The improved YOLOv11 model includes: an input layer, which preprocesses the multispectral image acquired by an industrial camera and then inputs it into the backbone network; a backbone network, including convolutional layers and C3k2 modules, for acquiring multi-scale features of the image; a neck network, which performs multi-scale fusion, upsampling, and stitching operations on the input feature map to output a high-quality multi-scale feature map; and a detection head, which first performs SAM processing on the high-quality feature map and then inputs it into the corresponding detection head, which outputs the confidence score, detection box location, and defect category.

[0013] The transparent layer bubble detection unit is used to identify the detection box of bubble defects on the two-dimensional layered image according to the bubble detection model; the bubble detection model is constructed based on the YOLOv11 model;

[0014] Based on the coordinates of the detection frame position By setting the speed and frame rate data, the three-dimensional coordinate matrix of each bubble on each frame of the image is determined. Treating bubbles as standard spheres, the volume of a single bubble at the boundary of the transparent layer is calculated based on the three-dimensional coordinate matrix of similar bubbles after fusion, and the gas content is calculated in combination with the volume of the detection area.

[0015] The visualization unit is used to display the data collected by the data acquisition unit, the detection data and results of the defect detection unit and the transparent layer bubble detection unit.

[0016] Furthermore, the backbone network consists of 5 convolutional layers. Except for the first convolutional layer, each of the other convolutional layers is followed by a C3k2 module. The feature map output from the stacked last convolutional layer and the C3k2 module is used as the input to D-SPPF. The D-SPPF module is used to dynamically select the pooling kernel, and the output of the D-SPPF module is used as the input to the C2PSA module. The outputs of the stacked second convolutional layer and the C3k2 module, the stacked fourth convolutional layer and the C3k2 module, and the C2PSA module are used as the outputs of the backbone network.

[0017] Furthermore, the C2PSA module extracts features from the input feature map using a 1×1 convolutional layer, segments the extracted feature map, inputs the segmented feature map into multiple parallel PSA modules for feature processing, concatenates the feature maps processed by multiple PSA modules, and outputs the concatenated feature map after feature integration using a 1×1 convolutional layer.

[0018] The D-SPPF module first performs feature extraction on the input feature map using a 1×1 convolutional layer. The extracted feature map is then convolved with 3×3, 5×5, and 7×7 convolutional kernels. The feature maps processed by convolutional kernels of different sizes are then concatenated. The concatenated feature map is then pooled. The pooled feature map is then further pooled. Finally, the feature maps resulting from the two pooling operations are concatenated and output.

[0019] Furthermore, the SAM processing procedure is as follows: the input feature map is subjected to max pooling and average pooling to generate pooled features; the pooled features are input into a two-dimensional convolutional layer for convolution operation; and the convolution result is activated by the Sigmoid activation function to generate a spatial attention map.

[0020] The C3k2 module extracts features from the input feature map using a 1×1 convolutional layer; segments the extracted feature map; inputs the segmented feature map into multiple bottleneck layers for feature processing; concatenates the feature maps processed by the bottleneck layers; further extracts features from the concatenated feature map using a 3×3 convolutional layer; fuses the feature map output from the 3×3 convolutional layer with the feature map processed by the Sobel operator; and finally outputs the fused feature map after feature integration using a 1×1 convolutional layer.

[0021] Furthermore, the categories of defects include bubbles, black spots, deposits, white spots, pits, cracks, and crystallization points.

[0022] Furthermore, the first device also includes a dimension acquisition component, which includes a contact thickness sensor, a height measurement sensor for detecting the height of the crucible, and a diameter measurement sensor for detecting the diameter of the crucible; a camera robotic arm and a light source robotic arm hold the contact thickness sensor to acquire crucible wall thickness data;

[0023] The first and second placement platforms also include weight sensors for acquiring the weight data of the crucible;

[0024] The data acquisition unit is also used to acquire data on the weight, wall thickness, height, and diameter of the crucible obtained by the first device.

[0025] Furthermore, both the first and second devices also include positioning components. The positioning components include rubber-coated rollers driven by a screw module. The rubber-coated rollers are horizontally positioned on the first and second placement platforms perpendicular to the crucible feeding direction. Fixing members are symmetrically arranged on both sides of the rubber-coated rollers to fix the placement position of the crucible.

[0026] Furthermore, the visualization unit specifically includes: drawing a three-dimensional distribution map of bubbles based on a three-dimensional coordinate matrix, and visually displaying the data acquired by the data acquisition unit (crucible weight and crucible wall thickness) and the detection results of the defect detection unit;

[0027] When drawing a 3D bubble distribution map, the bubbles are displayed according to their radii, and bubbles with different radii are displayed in different colors.

[0028] Furthermore, the similar bubble fusion process is as follows: if the spatial distance tolerance and radius tolerance of two bubbles satisfy the following condition formula, then the two bubbles are determined to be similar bubbles, and bubble fusion processing is performed to merge the two bubbles into one bubble, taking the average of the coordinates and radius; the condition formula is:

[0029]

[0030]

[0031] in, The spatial distance between the two bubbles, ( )and( All of these are three-dimensional coordinate matrices of the bubble. For spatial distance tolerance, For radius tolerance;

[0032] The steps for determining the three-dimensional coordinate matrix of each bubble are as follows:

[0033] 1) Calculate the center coordinates of the bubble: Convert the normalized coordinates of the bubble's 3D coordinate matrix into actual pixel coordinates;

[0034]

[0035]

[0036] in, The center coordinates of the bubble, The center coordinates of the detection box The image width in pixels. The pixel value representing the image height;

[0037] 2) Calculate bubble depth coordinates: Calculate the depth interval for each frame based on the set speed and frame rate, and then calculate the corresponding depth coordinate value for that frame based on the calculated depth interval. The calculation formula is as follows:

[0038]

[0039] in, For the first The depth coordinates of the frame. For the depth interval of each frame, To set the speed, Set the frame rate for the industrial camera;

[0040] 3) Calculate the bubble radius: Determine the bubble radius by taking the larger value between the detection frame width and height. The expression is:

[0041] in, For the detection frame width, This is the height of the detection frame.

[0042] Furthermore, the transparent layer bubble detection unit also includes determining the transparent layer boundary and thickness. The process is as follows: the inner boundary depth is the transparent layer surface depth; the number of bubbles is counted frame by frame; the bubble count is compared with a set bubble threshold in ascending order of time; if the number of bubbles in a certain frame exceeds the set bubble threshold, the depth of the previous frame is determined as the outer boundary depth; the thickness of the transparent layer is determined based on the depths of the inner and outer boundaries.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] This invention acquires two-dimensional surface images and two-dimensional layered images of the crucible using a first device and a second device. A data acquisition unit collects this data to provide a basis for the defect detection unit and the microbubble detection unit. Since the defect detection unit needs to detect multiple defect categories, its defect detection model is constructed based on an improved YOLOv11 model. In contrast, the microbubble detection unit's bubble detection model detects only one type of defect, and its bubble detection model is also constructed based on the YOLOv11 model. This approach ensures detection accuracy while reducing the overall system overhead. The transparent layer bubble detection unit first infers the horizontal and vertical coordinates and radius of the bubble using the bubble detection frame position coordinate data. It then infers the vertical coordinate of the bubble by setting speed and frame rate data. Next, it eliminates duplicate bubble counts by merging similar bubbles, thus ensuring the reliability of the bubble count. Dynamic boundary determination based on a bubble count threshold reduces thickness measurement errors. The gas content is calculated based on the bubble count and the defined transparent layer boundaries, ensuring the accuracy of the gas content. The visualization unit integrates four-dimensional data correlation display of weight, wall thickness, defects, and bubbles. By fusing physical parameters with visual data, engineers can intuitively analyze the root causes of process defects, thereby solving the problem of unreliable calculations of transparent layer gas content and thickness in existing technologies, which makes it difficult to provide a reliable basis for quartz crucible quality assessment.

[0045] This invention aims to provide a system that integrates weight and thickness measurement, internal wall defect detection, and transparent layer microbubble detection, thereby improving the accuracy, efficiency, and automation of quartz crucible inspection, reducing manual intervention, lowering the false judgment rate, and meeting the photovoltaic industry's demand for high-quality quartz crucible inspection. Specifically, it addresses the problems of limited detection functions, insufficient accuracy, low automation, and low efficiency in existing technologies, achieving comprehensive, accurate, and efficient inspection of quartz crucibles.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles according to the present invention.

[0048] Figure 2 The image shows a black spot defect detected by the defect detection unit in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention.

[0049] Figure 3 The image of white spot defects detected by the defect detection unit in an embodiment of the intelligent visual full inspection system for the quality of quartz crucibles leaving the factory according to the present invention;

[0050] Figure 4Image of crystallization point defects detected by the defect detection unit in an embodiment of the intelligent visual full inspection system for the factory quality of quartz crucibles of the present invention;

[0051] Figure 5 An image of bubble defects detected by the microbubble detection unit in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0052] Figure 6 This is a schematic diagram of the display interface of the visualization unit in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0053] Figure 7 This is a schematic diagram of the interface of the visualization unit with a three-dimensional bubble distribution map in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention.

[0054] Figure 8 The diagram shows the structure of ST1 and ST2 devices in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention.

[0055] Figure 9 This is a schematic diagram of the ST1 device, an embodiment of the intelligent visual full inspection system for the factory quality of quartz crucibles of the present invention.

[0056] Figure 10 This is a schematic diagram of the ST2 device, an embodiment of the intelligent visual full inspection system for the factory quality of quartz crucibles of the present invention.

[0057] Figure 11 This is a schematic diagram of the weighing structure of the ST1 device in an embodiment of the intelligent visual full inspection system for the factory quality of quartz crucibles of the present invention;

[0058] Figure 12 This is a schematic diagram of the industrial camera installation position of the ST1 device in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention.

[0059] Figure 13 This is a schematic diagram of the microbubble camera installation location structure of the ST2 device in an embodiment of the intelligent visual full inspection system for the factory quality of quartz crucibles of the present invention.

[0060] Figure 14 A schematic diagram of the improved YOLOv11 model structure of an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0061] Figure 15 This is a schematic diagram of the SAM structure of an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0062] Figure 16 This is a schematic diagram of the C2PSA module structure of an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0063] Figure 17 This is a schematic diagram of the D-SPPF module structure of an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention;

[0064] Figure 18 This is a schematic diagram of the C3k2 module structure in an embodiment of the intelligent visual full inspection system for the outgoing quality of quartz crucibles of the present invention.

[0065] Explanation of reference numerals in the attached figures:

[0066] 11. First placement stage; 12. Camera robotic arm; 13. Light source robotic arm; 14. Industrial camera; 15. Strip light source device; 16. Height measurement sensor; 17. Contact thickness sensor;

[0067] 18. Diameter measuring sensor;

[0068] 21. Second placement stage; 22. Camera robotic arm; 23. Adjustable focus light source device;

[0069] 24. Microbubble detection camera; 25. Camera motion module;

[0070] 31. The crucible to be tested;

[0071] 41. Rubber-coated rollers; 42. Fixing components; 43. Weight sensor. Detailed Implementation

[0072] Example of an intelligent visual full-inspection system for the quality of quartz crucibles leaving the factory:

[0073] like Figures 1-7 As shown, the intelligent visual full inspection system for the quality of quartz crucibles includes a crucible inspection device, a data acquisition unit, a defect detection unit, a transparent layer bubble detection unit, and a visualization unit.

[0074] like Figures 8-13 As shown, the crucible inspection device includes a first device and a second device, namely ST1 device and ST2 device. The first device includes a first placement stage 11 and an image acquisition component; the image acquisition component includes a camera robotic arm 12 disposed in the middle of the first placement stage 11 and a light source robotic arm 13 fixed above the first placement stage 11; the camera robotic arm 12 is used to hold an industrial camera 14, and the light source robotic arm 13 is used to hold a strip light source device 15, so as to acquire a two-dimensional surface image of the inner wall of the crucible.

[0075] The first device also includes a dimension acquisition component, which includes a contact thickness sensor, a height measurement sensor 16 for detecting the height of the crucible, and a diameter measurement sensor 18 for detecting the diameter of the crucible; the camera robotic arm 12 and the light source robotic arm 13 hold the contact thickness sensor 17 to acquire crucible wall thickness data.

[0076] The second device includes a second placement stage 21 and a video acquisition component; the video acquisition component includes a camera robotic arm 22 fixed in the middle of the second placement stage 21 and an adjustable focus light source device 23. The camera robotic arm 22 is connected to a microbubble detection camera 24 through a camera motion module 25; the microbubble detection camera 24 and the adjustable focus light source device 23 work together to acquire two-dimensional layered images of the inner wall of the crucible at a set speed and frame rate; both the first placement stage 11 and the second placement stage 21 are used to support the crucible 31 to be tested with its opening facing downward.

[0077] Preferably, the camera robotic arm 12, the light source robotic arm 13, and the camera robotic arm 22 are all 6-axis robotic arms.

[0078] The two devices share a similar structure, both including a positioning assembly. This assembly comprises a rubber-coated roller 41 driven by a lead screw module. The roller 41 is horizontally positioned on the first placement stage 11 and the second placement stage 21, perpendicular to the crucible feeding direction. Symmetrical fixing members 42 are arranged on both sides of the roller 41 to fix the crucible's placement position. The positioning assembly controls the roller 41 based on dimensional data collected by the height measuring sensor 16 and the diameter measuring sensor 18, enabling the roller 41 to transport the crucible to the detection position.

[0079] The first placement stage 11 and the second placement stage 21 also include a weight sensor 43 for acquiring the weight data of the crucible. Preferably, the weight sensor 43 is a weight sensor with a range of 500 kg; when the quartz crucible is placed on the drum, the weight sensor can measure the weight of the crucible in real time and transmit the weight data to the control component.

[0080] To achieve automation, the crucible inspection device can also be equipped with control components, such as PLC+ and HMI control. The PLC is responsible for controlling the actions of various components of the equipment, such as starting and stopping the roller conveyor, moving the positioning components, moving the robotic arm, and taking pictures with the camera; the HMI is used for operators to interact with the equipment, displaying the equipment's operating status and inspection data, and also allowing for parameter settings, formula selection, and other operations.

[0081] The data acquisition process of the crucible inspection device is as follows: The operator places the crucible 31 to be inspected on the rubber-coated roller 31 of the ST1 equipment. The weight sensor 43 collects the weight data. The control component controls the rubber-coated roller 31 to move to the inspection position based on the weight data. The fixing component 42 fixes the crucible in place. The camera robotic arm 12 and the light source robotic arm 13 work together to enable the contact thickness sensor 17 to detect the crucible wall thickness, and the industrial camera to acquire a two-dimensional surface image of the crucible. The height measurement sensor 16 and the diameter measurement sensor 18 collect the corresponding height and diameter data. Similarly, the crucible 31 to be inspected is placed on the ST2 equipment. After positioning, the microbubble inspection camera 24 and the adjustable focus light source device 23 work together to acquire a two-dimensional layered image of the crucible.

[0082] The data acquisition unit is used to acquire two-dimensional surface images and two-dimensional layered images obtained by the first and second devices, and also to acquire data on the weight, wall thickness, height, and diameter of the crucible obtained by the first device.

[0083] The defect detection unit is used to identify the type of defects and the location of the detection box in a two-dimensional surface image through a defect detection model; the defect detection model is constructed based on an improved YOLOv11 model. Defect types include bubbles, black spots, attachments, white spots, pits, cracks, and crystallization points.

[0084] like Figures 14-18 As shown, the improved YOLOv11 model includes:

[0085] The input layer is used to preprocess multispectral images acquired by industrial cameras before inputting them into the backbone network.

[0086] The backbone network, which includes convolutional layers and C3k2 modules, is used to acquire multi-scale features of the image.

[0087] like Figure 14 As shown, the backbone network consists of 5 convolutional layers. Except for the first convolutional layer, each of the other convolutional layers is followed by a C3k2 module. The feature map output from the stacked last convolutional layer and the C3k2 module serves as the input to the D-SPPF module. The D-SPPF module dynamically selects pooling kernels, including 3×3, 5×5, and 7×7 kernels. The output of the D-SPPF module serves as the input to the C2PSA module. The outputs of the stacked second and fourth convolutional layers and the C3k2 module, and the C2PSA module are used as the outputs of the backbone network. The backbone network is sequentially composed of Conv, Conv, C3K2, Conv, C3K2, Conv, C3K2, D-SPPF, and C2PSA.

[0088] The first two convolutional layers are used to initially extract basic features such as edges and textures, and to complete the initial channel adjustment and feature downsampling; the first C3K2 module is used to enhance the features of defect edges and contours; then, convolutional layers and C3K2 modules are stacked alternately; the convolutional layers further downsample to extract deeper features; each C3K2 module enhances the features of defect edges, textures, shapes and other features at different levels; the PSA (Dynamic Channel Attention) of the C2PSA module enhances the ability to utilize multispectral features; the features are further optimized and fused.

[0089] The neck network performs multi-scale fusion, upsampling, and concatenation operations on the input feature maps, outputting high-quality multi-scale feature maps. Row upsampling expands the feature map size, and feature concatenation enriches semantic information and details.

[0090] The detection head first processes the high-quality feature map using SAM (Spatial Attention Module), and then inputs the processed data into the corresponding detection heads. The detection heads output confidence scores, bounding box positions, and defect categories. Specifically, by capturing the differences in the importance of spatial positions in the feature map, it dynamically focuses on the spatial distribution of edge contours and the spatial range of defect morphologies, suppressing spatial feature interference from background or non-defect areas, and enhancing the saliency and localization discriminability of defect spatial features. Each scale corresponds to a target of different sizes.

[0091] The 2D surface image input to the trained defect detection model is the preprocessed and enhanced image. YOLO detection outputs a detection report, which is a ".txt" file within the report.

[0092] like Figure 15 As shown, the SAM (Spatial Attention Mechanism) process involves: performing max pooling and average pooling on the input feature map to generate pooled features; inputting the pooled features into a two-dimensional convolutional layer for convolution; and activating the convolution result using the Sigmoid activation function to generate a spatial attention map. By emphasizing the features of small-sized defects, SAM processing helps improve the model's ability to recognize small defects such as bubbles and black spots. Furthermore, this SAM processing complements the PSA (Dynamic Channel Attention Mechanism) of the backbone network's C2PSA module. The former focuses on focusing on defect regions in the spatial dimension, while the latter focuses on utilizing multispectral features in the channel dimension, together strengthening the network's ability to perceive and utilize defect features from both spatial and channel dimensions.

[0093] like Figure 16As shown, the C2PSA module first performs a 1×1 convolution operation on the input feature map to segment the extracted feature map. This segmentation is then achieved through a Split operation. The segmented feature map is input to multiple parallel PSA modules for feature processing. One branch extracts features through n PSA processes, while another branch extracts features through convolution. The outputs of the two branches are concatenated along the channel dimension using a Concat operation and then fused through a 1×1 convolution before outputting the final result. This achieves multi-branch feature extraction, fusion, and optimization, enhancing the model's ability to express complex features. The C2PSA module can adaptively emphasize important spatial information while ignoring irrelevant background or noise, thereby enhancing the model's ability to express defect features. By emphasizing the features of defect regions, the C2PSA module helps improve the model's recognition accuracy for small-sized defects such as bubbles, black spots, and attachments.

[0094] like Figure 17 As shown, the D-SPPF module first extracts features from the input feature map using a 1×1 convolutional layer. The extracted feature maps are then convolved with 3×3, 5×5, and 7×7 kernels. The feature maps processed by kernels of different sizes are then concatenated. Pooling is performed on the concatenated feature maps. Further pooling is then performed on the pooled feature maps. Finally, the feature maps resulting from the two pooling operations are concatenated and output. The concatenation is achieved using the `Concat` operation. The D-SPPF module, through convolutional kernels of different sizes and multiple pooling operations, can extract features at multiple scales, helping the model identify defects of different sizes, such as white spots, pits, and cracks. D-SPPF is also known as an improved dilated spatial pyramid pooling module. This module enhances the model's ability to detect targets at different scales by combining dilated convolution and spatial pyramid pooling techniques.

[0095] like Figure 18 As shown, the C3k2 module extracts features from the input feature map using a 1×1 convolutional layer; it then segments the extracted feature map; the segmented feature map is fed into multiple bottleneck layers for feature processing; the bottleneck-processed feature maps are concatenated; a 3×3 convolutional layer further extracts features from the concatenated feature map; the feature map output from the 3×3 convolutional layer is fused with the feature map processed by the Sobel operator for edge detection and to enhance feature details; finally, a 1×1 convolutional layer integrates the features of the fused feature map before outputting the result. Segmentation is implemented using the Split operation, and concatenation is implemented using the Concat operation. Through multiple feature segmentation, bottleneck layer processing, and concatenation operations, the C3k2 module enhances the feature fusion capability, thereby improving the model's ability to express defect features.

[0096] The transparent layer bubble detection unit is used to identify the detection boxes of bubble defects on the two-dimensional layered image according to the bubble detection model; the bubble detection model is constructed based on the YOLOv11 model. The YOLOv11 model here is not improved because only one type of defect needs to be identified.

[0097] Based on the coordinates of the detection frame position By setting the speed and frame rate data, the three-dimensional coordinate matrix of each bubble on each frame of the image is determined. Treating bubbles as standard spheres, the volume of a single bubble at the boundary of the transparent layer is calculated based on the three-dimensional coordinate matrix of similar bubbles after fusion. This volume, combined with the volume of the detection area, is then used to calculate the gas content. To avoid an excessively high proportion of duplicated bubbles, which would affect the accuracy of bubble counting...

[0098] The steps for determining the three-dimensional coordinate matrix of each bubble are as follows:

[0099] 1) Calculate the center coordinates of the bubble: Convert the normalized coordinates of the bubble's 3D coordinate matrix into actual pixel coordinates;

[0100]

[0101]

[0102] in, The center coordinates of the bubble, The center coordinates of the detection box The image width in pixels. The pixel value representing the image height;

[0103] 2) Calculate bubble depth coordinates: Calculate the depth interval for each frame based on the set speed and frame rate, and then calculate the corresponding depth coordinate value for that frame based on the calculated depth interval. The calculation formula is as follows:

[0104]

[0105] in, For the first The depth coordinates of the frame. For the depth interval of each frame, To set the speed, the unit is... , This is the set frame rate for the industrial camera, measured in fps. The set speed here is equal to the movement speed of the camera's robotic arm 3.

[0106] 3) Calculate the bubble radius: Determine the bubble radius by taking the larger value between the detection frame width and height. The expression is:

[0107] in, For the detection frame width, This is the height of the detection frame.

[0108] The similar bubble fusion process is as follows: If the spatial distance tolerance and radius tolerance of two bubbles satisfy the following condition formula, then the two bubbles are determined to be similar bubbles, and bubble fusion processing is performed to merge the two bubbles into one bubble, taking the average of the coordinates and radius; the condition formula is:

[0109]

[0110]

[0111] in, The spatial distance between two bubbles, in units of , ( )and( All of these are three-dimensional coordinate matrices of the bubble. For spatial distance tolerance, This is the radius tolerance.

[0112] The formulas for calculating the bubble coordinates and radius in the fused 3D coordinate matrix are as follows:

[0113]

[0114]

[0115] in, For the number of similar bubbles, The numbers represent the sequence numbers of similar bubbles.

[0116] The process of determining the boundary and thickness of the transparent layer is as follows: The inner boundary depth is the surface depth of the transparent layer. The number of bubbles is counted frame by frame and compared sequentially with a set bubble threshold in ascending order of time. If the number of bubbles in a frame exceeds the set bubble threshold, the depth of the previous frame is determined as the outer boundary depth. The thickness of the transparent layer is determined based on the depths of the inner and outer boundaries. Here, the previous frame refers to the layered image at the boundary between the transparent and opaque layers. The bubble threshold is set according to the actual situation; for example, in a 2mm field of view, the bubble threshold is set to 30.

[0117] The formula for calculating the thickness of the transparent layer is:

[0118] in, The outer boundary depth, The depth is the inner boundary.

[0119] The formula for calculating the gas content mentioned above is:

[0120]

[0121] in, This represents the total gas-containing volume. To detect the volume of the region.

[0122] The formula for calculating the volume of the detection area is:

[0123]

[0124] in, The length of the detection area, Width of the detection area This refers to the thickness of the transparent layer.

[0125] The formula for calculating the total gas-containing volume is:

[0126]

[0127] in, This represents the total number of bubbles in the detection area. For the first The radius of each bubble.

[0128] The visualization unit is used to display the data collected by the data acquisition unit, the detection data and results of the defect detection unit and the transparent layer bubble detection unit.

[0129] The visualization unit specifically includes: drawing a 3D bubble distribution map based on a 3D coordinate matrix; and visualizing the data acquired by the data acquisition unit (crucible weight and crucible wall thickness) and the detection results from the defect detection unit. To visualize the size and color of the bubbles in a clear, intuitive, and systematic way in the 3D map, key geometric planes and text labels are added to better present and convey the bubble-related data characteristics and their relationship to their spatial region. Specifically, when drawing the 3D bubble distribution map, the bubbles are displayed according to their radius, with bubbles of different radii displayed in different colors.

[0130] Preferably, the radius of the displayed bubble is equal to the bubble radius multiplied by 4 pixels. The color of the displayed bubbles is distributed with smaller radii being closer to light gray and larger radii being closer to blue. Simultaneously, the thickness data of the transparent layer is displayed at the corners of the 3D image.

[0131] This invention can be applied in various scenarios, such as photovoltaic manufacturing companies or quartz crucible manufacturers. It achieves high-precision wall thickness measurement through a contact-type thickness sensor. By combining a strip light source with an industrial camera, it eliminates interference from reflections on the curved surface of the crucible, improving the quality of acquired images. Through a bubble fusion algorithm, based on a dynamic merging strategy using dual thresholds of three-dimensional Euclidean distance and radius, it effectively handles the problem of repeated detection of the same bubble in multiple frames of images. By analyzing the thickness of the bubble distribution boundary and a preset threshold, it achieves non-contact measurement of the transparent layer thickness.

[0132] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A smart visual full-inspection system for the outgoing quality of quartz crucibles, characterized in that: The system includes: The crucible inspection device includes a first device and a second device. The first device includes a first placement stage (11) and an image acquisition component. The image acquisition component includes a camera robotic arm (12) disposed in the middle of the first placement stage (11) and a light source robotic arm (13) fixed above the first placement stage (11). The camera robotic arm (12) is used to hold an industrial camera (14), and the light source robotic arm (13) is used to hold a strip light source device (15) to acquire a two-dimensional surface image of the inner wall of the crucible. The second device includes a second placement stage (21) and a video... The acquisition component includes a camera robotic arm (22) fixed in the middle of the second placement stage (21) and an adjustable focus light source device (23). The camera robotic arm (22) is connected to a microbubble detection camera (24) through a camera motion module (25). The microbubble detection camera (24) and the adjustable focus light source device (23) work together to acquire two-dimensional layered images of the inner wall of the crucible at a set speed and frame rate. The first placement stage (11) and the second placement stage (21) are both used to support the crucible (31) to be tested with its opening facing downward. The data acquisition unit is used to acquire two-dimensional surface images and two-dimensional layered images obtained by the first device and the second device; The defect detection unit is used to identify the defect category and detection box location in a two-dimensional surface image through a defect detection model. The defect detection model is built based on an improved YOLOv11 model. The improved YOLOv11 model includes: an input layer, which preprocesses the multispectral image acquired by an industrial camera and then inputs it into the backbone network; a backbone network, including convolutional layers and C3k2 modules, for acquiring multi-scale features of the image; a neck network, which performs multi-scale fusion, upsampling, and stitching operations on the input feature map to output a high-quality multi-scale feature map; and a detection head, which first performs SAM processing on the high-quality feature map and then inputs it into the corresponding detection head, which outputs the confidence score, detection box location, and defect category. The transparent layer bubble detection unit is used to identify the detection box of bubble defects on the two-dimensional layered image according to the bubble detection model; the bubble detection model is constructed based on the YOLOv11 model; Based on the coordinates of the detection frame position Set the speed and frame rate data, among which, The center coordinates of the detection box For the detection frame width, To determine the height of the detection box, the three-dimensional coordinate matrix of each bubble on each frame of the image is determined. ,in, Let be the radius of the bubble; treating the bubble as a standard sphere, calculate the volume of a single bubble at the boundary of the transparent layer based on the three-dimensional coordinate matrix of similar bubbles after fusion, and then calculate the gas content based on this and the volume of the detection area. The visualization unit is used to display the data collected by the data acquisition unit, the detection data and results of the defect detection unit and the transparent layer bubble detection unit.

2. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: The backbone network consists of 5 convolutional layers. Except for the first convolutional layer, each of the other convolutional layers is followed by a C3k2 module. The feature map output from the stacked last convolutional layer and the C3k2 module is used as the input to the D-SPPF module. The D-SPPF module is used to dynamically select the pooling kernel, and the output of the D-SPPF module is used as the input to the C2PSA module. The outputs of the stacked second convolutional layer and the C3k2 module, the stacked fourth convolutional layer and the C3k2 module, and the C2PSA module are used as the outputs of the backbone network.

3. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 2, characterized in that: The C2PSA module extracts features from the input feature map using a 1×1 convolutional layer, segments the extracted feature map, and then inputs the segmented feature map into multiple parallel PSA modules for feature processing. The feature maps processed by multiple PSA modules are concatenated; the concatenated feature maps are then integrated using a 1×1 convolutional layer before being output. The D-SPPF module first performs feature extraction on the input feature map using a 1×1 convolutional layer, and then performs convolution operations on the extracted feature map using 3×3, 5×5, and 7×7 convolutional kernels respectively; the feature maps processed by convolutional kernels of different sizes are then stitched together. Perform pooling on the concatenated feature maps; perform further pooling on the pooled feature maps; and concatenate the feature maps after the two pooling operations and output the result.

4. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: The SAM processing procedure is as follows: the input feature map is subjected to max pooling and average pooling to generate pooled features; the pooled features are input into a two-dimensional convolutional layer for convolution operation; the convolution result is activated by the Sigmoid activation function to generate a spatial attention map. The C3k2 module extracts features from the input feature map using a 1×1 convolutional layer. The extracted feature map is segmented; the segmented feature map is then input into multiple bottleneck layers for feature processing. The feature maps processed by the bottleneck layer are concatenated; further feature extraction is performed on the concatenated feature maps using a 3×3 convolutional layer. The feature map output from the 3×3 convolutional layer is fused with the feature map processed by the Sobel operator; the fused feature map is then integrated and output after passing through a 1×1 convolutional layer.

5. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: Defects include bubbles, black spots, deposits, white spots, pits, cracks, and crystallization points.

6. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: The first device also includes a size acquisition component, which includes a contact thickness sensor (17), a height measurement sensor (16) for detecting the height of the crucible, and a diameter measurement sensor (18) for detecting the diameter of the crucible; a camera robotic arm (12) and a light source robotic arm (13) hold the contact thickness sensor (17) to acquire crucible wall thickness data; The first placement stage (11) and the second placement stage (21) also include a weight sensor (43) for acquiring the weight data of the crucible; The data acquisition unit is also used to acquire data on the weight, wall thickness, height, and diameter of the crucible obtained by the first device.

7. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: Both the first and second devices also include positioning components. The positioning components include rubber-coated rollers (41) driven by a screw module. The rubber-coated rollers (41) are horizontally positioned on the first placement platform (11) and the second placement platform (21) perpendicular to the crucible feeding direction. Fixing members (42) are symmetrically arranged on both sides of the rubber-coated rollers (41) to fix the placement position of the crucible.

8. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: The visualization unit specifically includes: drawing a three-dimensional distribution map of bubbles based on a three-dimensional coordinate matrix, and visually displaying the data acquired by the data acquisition unit (crucible weight and crucible wall thickness) and the detection results of the defect detection unit; When drawing a 3D bubble distribution map, the bubbles are displayed according to their radii, and bubbles with different radii are displayed in different colors.

9. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 1, characterized in that: The similar bubble fusion process is as follows: If the spatial distance tolerance and radius tolerance of two bubbles satisfy the following condition formula, then the two bubbles are determined to be similar bubbles, and bubble fusion processing is performed to merge the two bubbles into one bubble, taking the average of the coordinates and radius; the condition formula is: in, The spatial distance between the two bubbles, ( )and( All of these are three-dimensional coordinate matrices of the bubble. For spatial distance tolerance, For radius tolerance; The steps for determining the three-dimensional coordinate matrix of each bubble are as follows: 1) Calculate the center coordinates of the bubble: Convert the normalized coordinates of the bubble's 3D coordinate matrix into actual pixel coordinates; in, The center coordinates of the bubble, The center coordinates of the detection box The image width in pixels. The pixel value representing the image height; 2) Calculate bubble depth coordinates: Calculate the depth interval for each frame based on the set speed and frame rate, and then calculate the depth coordinate value for the corresponding frame based on the calculated depth interval. The calculation formula is as follows: in, For the first The depth coordinates of the frame. For the depth interval of each frame, To set the speed, Set the frame rate for the industrial camera; 3) Calculate the bubble radius: Determine the bubble radius by taking the larger value between the detection frame width and height. The expression is: in, For the detection frame width, This is the height of the detection frame.

10. The intelligent visual full inspection system for the outgoing quality of quartz crucibles according to claim 9, characterized in that: The transparent layer bubble detection unit also includes determining the transparent layer boundary and thickness. The process is as follows: the inner boundary depth is the transparent layer surface depth, the number of bubbles is counted frame by frame, and compared with the set bubble threshold in ascending order of time. If the number of bubbles in a certain frame exceeds the set bubble threshold, the depth of the previous frame is determined as the outer boundary depth. The thickness of the transparent layer is determined based on the depths of the inner and outer boundaries.

Images