Microwave component welding penetration measurement method based on deep learning

Abstract

The present invention provides a method for measuring the solder penetration rate of microwave components based on deep learning, and relates to the field of digitalization and intelligence of microwave component production. In the present invention, an X-ray image of the microwave component to be measured is obtained and preprocessed; the preprocessed X-ray image is used as the input of the preset YoloV8 model deep learning model to detect the void area on the specified device area, and each void area is instance segmented to obtain a pixel-level mask image; based on the mask image, the pixel area of ​​each void area is obtained; based on the pixel area of ​​each void area, the solder void rate on the specified device area is obtained to obtain the solder penetration rate of the specified device area. The use of deep learning methods to measure solder penetration rate has bid farewell to the method of judging void pixels by a single threshold, and the void calibration accurate to pixels and the powerful generalization ability of deep learning ensure the accuracy of void position and area prediction.

Description

Technical Field

[0001] The present invention relates to the field of digital and intelligent production of microwave components, and in particular to a method for measuring the solder penetration rate of microwave components based on deep learning. Background Art

[0002] Microwave components are a common form of hybrid integrated circuits, usually composed of a housing, high-frequency and low-frequency connectors, a substrate, a chip, and various substrates. The solder penetration rate between the housing and the substrate, and between the chip and the substrate has a direct impact on the reliability of the performance of microwave components.

[0003] In the related art, the numerical detection method of microwave component welding permeability is to use a method of binarizing the image and then performing threshold segmentation. However, the method based on threshold segmentation uses a single threshold to determine whether it is a hollow pixel, which often causes large statistical errors. In addition, when the microwave components being tested have different types of detection areas, such as substrates, annular isolators, chips, substrates and other devices, the welding images of the device areas are often quite different. The method based on a single threshold segmentation often requires different parameter schemes to adapt to different device area images.

[0004] In view of the above-mentioned defects, it is necessary to provide a welding penetration rate measurement method with small error and applicable to different detection areas of microwave components. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention provides a method for measuring the solder penetration rate of microwave components based on deep learning, which solves the technical problems that the traditional method based on single threshold segmentation has large errors and cannot be applied to different detection areas of microwave components.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A method for measuring the solder penetration rate of microwave components based on deep learning, comprising:

[0010] Acquire and preprocess X-ray images of the microwave components to be tested;

[0011] The preprocessed X-ray image is used as the input of the preset YoloV8 model to detect the hole area on the specified device area, and each of the hole areas is instance segmented to obtain a pixel-level mask image; wherein the size of the hole areas in different device areas varies in magnitude;

[0012] Based on the mask image, obtaining a pixel area of ​​each of the hole regions;

[0013] Based on the pixel area of ​​each of the void regions, the solder void rate on the designated device region is obtained to obtain the solder penetration rate of the designated device region.

[0014] Preferably, the preprocessing includes image sharpening processing, histogram equalization processing and multi-image averaging processing.

[0015] Preferably, the training process of the YoloV8 model includes:

[0016] Acquire and preprocess multiple historical X-ray images of microwave components;

[0017] Perform image enhancement on the preprocessed historical X-ray images;

[0018] On each enhanced image, four-point rectangular frames are marked on the hole areas on different device regions, and contours are marked on the hole areas on different device regions;

[0019] Divide all annotated images into training set, validation set and test set;

[0020] Set the hyperparameters of the YoloV8 model;

[0021] The training set is used as the input of the YoloV8 model, the loss function is calculated, and the gradient is calculated using the back propagation algorithm until the model converges;

[0022] Using the validation set to evaluate model performance and adjust model hyperparameters;

[0023] Select multiple models that perform well on the validation set, perform a final evaluation on the test set, output model performance indicators respectively, and select the best performing model as the final YoloV8 model.

[0024] Preferably, the image enhancement is mosaic enhancement and / or mixed enhancement.

[0025] Preferably, the loss function includes a category classification loss function and a bounding box regression loss function.

[0026] Preferably, the category classification loss function adopts binary cross entropy loss, and / or the bounding box regression loss function adopts distribution focus loss and CIOU loss.

[0027] Preferably, based on the mask image, the pixel area of ​​each hole region is obtained using a polygon area formula; the polygon area formula is:

[0028]

[0029] in, AreaRepresents the area of ​​a polygon; N b Represents the number of boundary points of the polygon surrounding the hole area on the mask image, ( X i ,Y i )and( X i+1 ,Y i+1 ) represent the i、i+ The xy coordinates of a boundary point.

[0030] A microwave component welding brazing penetration measurement device based on deep learning, comprising:

[0031] An image preprocessing module, used for acquiring and preprocessing an X-ray image of the microwave component to be tested;

[0032] A hole segmentation module is used to use the preprocessed X-ray image as the input of a preset YoloV8 model to detect the hole area on the specified device area, and perform instance segmentation on each of the hole areas to obtain a pixel-level mask image; wherein the size of the hole areas in different device areas has order of magnitude differences;

[0033] A pixel area calculation module, used for obtaining the pixel area of ​​each of the hole regions based on the mask image;

[0034] The solder penetration rate acquisition module is used to acquire the solder void rate on the designated device area based on the pixel area of ​​each of the void areas, so as to acquire the solder penetration rate of the designated device area.

[0035] A microwave component welding permeability measurement system based on deep learning, including an X-ray tube, a stage, a flat panel detector and a computer workstation;

[0036] The X-ray tube is used to emit X-rays excited by high-energy electrons to the microwave component to be tested placed on the stage;

[0037] The flat panel detector is used to receive X-ray radiation penetrating the microwave component to be tested, generate an electrical signal proportional to the radiation intensity, and display it through an X-ray image;

[0038] The computer workstation is used to execute the microwave component welding permeability measurement method as described above.

[0039] (III) Beneficial effects

[0040] The present invention provides a method for measuring the solder penetration rate of microwave components based on deep learning. Compared with the prior art, it has the following beneficial effects:

[0041] In the present invention, an X-ray image of the microwave component to be tested is obtained and preprocessed; the preprocessed X-ray image is used as the input of the preset YoloV8 model deep learning model to detect the void area on the specified device area, and each of the void areas is segmented to obtain a pixel-level mask image; based on the mask image, the pixel area of ​​each void area is obtained; based on the pixel area of ​​each void area, the welding void rate on the specified device area is obtained to obtain the welding penetration rate of the specified device area. The deep learning method is used to measure the welding penetration rate, which bids farewell to the method of judging void pixels by a single threshold. The void calibration accurate to pixels and the powerful generalization ability of deep learning ensure the accuracy of the prediction of void position and area. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 A block diagram of a method for measuring the solder penetration rate of microwave components based on deep learning provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the structure of a YoloV8 model provided in an embodiment of the present invention;

[0045] FIG3 (a) is a schematic diagram of mosaic enhancement provided by an embodiment of the present invention;

[0046] FIG3 (b) is a schematic diagram of a hybrid enhancement provided by an embodiment of the present invention;

[0047] FIG4 (a) is a schematic diagram of a CIOU loss (CIOU Loss) provided in an embodiment of the present invention;

[0048] FIG4( b ) is a schematic diagram of an intersection-over-union ratio of a predicted detection box area and a real detection box area provided by an embodiment of the present invention;

[0049] FIG4( c ) is a schematic diagram of a distributed focus loss (DFL Loss) provided by an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of the structure of a microwave component welding brazing penetration measurement system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] The embodiment of the present application provides a method for measuring the solder penetration rate of microwave components based on deep learning, thereby solving the technical problems that the traditional method based on single threshold segmentation has large errors and cannot be applied to different detection areas of microwave components.

[0053] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0054] The traditional microwave component welding permeability measurement method based on a single threshold only considers the grayscale value of isolated pixels, resulting in a large amount of loss and waste of image feature information. A single threshold cannot cope with the complex actual production conditions in the device area. The inconsistent solder thickness caused by the inability to control the amount of solder during the welding process will cause the background grayscale value of the device area of ​​the void pattern to change randomly, which is also the biggest source of error in this measurement method.

[0055] In contrast, the embodiment of the present invention uses a deep learning method to predict the welding void rate. Since this method predicts the void pixel point by point, the convolution operation can be used to extract the correlation information between pixels of different scales. At the same time, adding welding X-ray images of different equipment areas to the training data set of the model improves the generalization ability of the model.

[0056] At the same time, the applicant found in practice that microwave multi-channel components have complex backgrounds (metal shells and substrates of different materials and densities) and many types of components compared to other types of electronic components, such as IGBT modules. When actually measuring the void area, i.e., bubbles, there are characteristics of inconsistent bubble shapes and sizes. The traditional void measurement method based on pre-value segmentation and graphic fitting methods cannot be well applied to the void area measurement scenario with such complex backgrounds and complex patterns.

[0057] The difference is that the embodiment of the present invention introduces the YOLOV8 deep learning model, which cleverly achieves a high detection rate for hole areas with various sizes of orders of magnitude by changing the relevant parameters of the rectangular box, i.e., the anchor box size setting. There is no problem that the traditional fitting method cannot effectively determine the fitting parameters when the hole area is too large or too small.

[0058] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0059] Embodiment 1:

[0060] like Figure 1 As shown, an embodiment of the present invention provides a method for measuring the solder penetration rate of microwave components based on deep learning, comprising:

[0061] S1. Acquire and preprocess an X-ray image of the microwave component to be tested;

[0062] S2. Using the preprocessed X-ray image as input to a preset YoloV8 model to detect the hole area on the specified device area, and performing instance segmentation on each of the hole areas to obtain a pixel-level mask image; wherein the sizes of the hole areas in different device areas vary in magnitude;

[0063] S3, based on the mask image, obtaining the pixel area of ​​each hole region;

[0064] S4. Based on the pixel area of ​​each void region, obtain the solder void rate on the specified device region to obtain the solder penetration rate of the specified device region.

[0065] The embodiment of the present invention adopts a deep learning method to measure welding penetration rate, bids farewell to the method of judging void pixels by a single threshold, and uses pixel-accurate void calibration and the powerful generalization ability of deep learning to ensure the accuracy of void position and area prediction.

[0066] Next, we will introduce the steps of the above scheme in detail:

[0067] In step S1, an X-ray image of the microwave component to be tested is acquired and preprocessed;

[0068] The preprocessing in this step includes image sharpening, histogram equalization, and multi-image averaging. Specifically:

[0069] In order to ensure the resolution and contrast of the image, the exposure time of the X-ray image is set to 200ms-400ms to avoid the image brightness being too dark or overexposed. For the exposed image, the sharpen operator is required to sharpen the image to increase the contrast of the image. In order to further ensure the accuracy of the measurement of the hole area in the image, the method of taking the average of multiple images can be used to obtain a stable hole image. The number of multi-image averages is set to 4-12 images to reduce the impact of local image jitter caused by the possible instability of X-ray intensity.

[0070] In step S2, the preprocessed X-ray image is used as the input of the preset YoloV8 model to detect the hole area on the specified device area, and each of the hole areas is instance segmented to obtain a pixel-level mask image; wherein the sizes of the hole areas in different device areas differ in magnitude.

[0071] like Figure 2 As shown in the figure, the YoloV8 model (i.e., the YoloV8 network structure) includes a backbone network, a neck and a head structure. The network adopts a PAN-FPN structure, which can extract feature information of different sizes in the image, and use the CONCAT function to fuse features of different sizes, so that the model can effectively detect large and small targets at the same time. In addition, the dual-head coupling architecture of the bounding box regression loss and the category classification loss is adopted to significantly improve the detection speed. In addition, the YoloV8 backbone network with different depths and widths has different parameter scales, which can be selected according to the characteristics of the object to be detected.

[0072] In this step, the training process of the selected YoloV8 model includes:

[0073] S10, acquiring and preprocessing a plurality of historical X-ray images of the microwave component.

[0074] The preprocessing of historical X-ray images also includes image sharpening, histogram equalization, and multi-image averaging.

[0075] S20, performing image enhancement on the preprocessed historical X-ray image.

[0076] As shown in Figures 3(a) to 3(b), the image enhancement includes mosaic enhancement and mixed enhancement to increase the number of synthesized images and avoid overfitting in training.

[0077] S30, on each enhanced image, four-point marking rectangular frames are performed on the hole regions on different device regions, and contour marking is performed on the hole regions on different device regions. The specific steps are as follows:

[0078] Using labeling software such as Labelme and Labelimg, on each enhanced image, four-point rectangular frames are annotated on the hole areas on different device regions to obtain the four vertex coordinates of the real detection frame; and polygons are used to mark the contours of the hole areas on different device regions to obtain the vertex coordinate information of the contour of the hole area.

[0079] It should be noted that the size parameters of the rectangular frame here vary in magnitude, depending on the actual sizes of different device regions and the various void regions thereon.

[0080] Then save the annotation information into the corresponding json file, and convert the json file into the txt file format required for model training through txt file conversion.

[0081] S40, dividing all the annotated images into a training set, a validation set and a test set.

[0082] S50. Set the hyperparameters of the YoloV8 model.

[0083] Set the initial learning rate lr0, the final learning rate lrf, the optimizer to SGD, and the image batch processing size is generally selected to be 4, 8, or 16 images; according to the GPU memory size, turn on the mosaic enhancement function and the number of epochs for training; and configure the name of the detection object to be void and the number of categories to be 1 in the YoloV8 configuration file.

[0084] S60, using the training set as the input of the YoloV8 model, calculating the loss function, and using the back propagation algorithm to calculate the gradient until the model converges; the specific steps are as follows:

[0085] The images in the training set are used as input of the YoloV8 model to perform target detection and instance segmentation, respectively, to obtain predicted candidate boxes of hole areas on different device areas, and to obtain vertex coordinate information of the mask image and the polygons surrounding the hole areas thereon.

[0086] Construct a loss function and use it for model training.

[0087] It should be noted that a variety of optimized loss function forms are used in the embodiments of the present invention to improve the accuracy of target detection and instance segmentation, ensuring that the model can handle various detection tasks.

[0088] Specifically, the loss function includes a category classification loss function and a bounding box regression loss function.

[0089] (1) The category classification loss function adopts binary cross entropy loss (BCE Loss), which is expressed as:

[0090]

[0091] in, N Indicates the number of predicted candidate boxes, n is the number of categories (here n =1, only one category is empty), y i The candidate box predicted by the model is i The true labels of the categories, p i Indicates that the predicted candidate box is iThe probability of being classified as positive.

[0092] (2) As shown in Figures 4(a) to 4(c), the bounding box regression loss function uses distribution focus loss (DFL Loss) and CIOU loss (CIOU Loss); they are respectively expressed as:

[0093]

[0094] in, y Represents any one of the four coordinate values ​​of the real detection box. Correspondingly, for each predicted candidate box, there are four predicted coordinate values ​​of the top, bottom, left and right. In the prediction, only the vertical coordinate For example: y i =[ y max ]( y max Rounding ， y i+1= [ y max ]+1 ， y max Indicates the maximum value of the vertical coordinate on the real detection box, y i and y i+1 are the two predicted ordinates closest to the maximum ordinate value on the real detection box, and their corresponding probabilities are S i and S i+1 , the vertical coordinate to be predicted here = y i S i +y i+1 S i+1 It can be regarded as a binomial probability distribution, that is, the probability predicted by the model S i and S i+1 Respectively and set y i and y i+1 Weighted multiplication.

[0095] It should be noted that the edge grayscale of the void area often gradually changes to the background grayscale, which is a technical difficulty that cannot be solved by traditional threshold segmentation and graphic fitting methods. The proposed distributed focus loss regards the boundary parameters of the fitted rectangular frame, i.e., the anchor frame, as a statistical distribution (in line with certain statistical laws), and the neural network is good at fitting the parameters of statistical distributions. This, on the one hand, increases the good generalization ability of the approximation of the real detection frame, and on the other hand, can better adapt to the various uncertainties of the grayscale gradient of the boundary of the void area. In other words, the distributed focus loss selected in the embodiment of the present invention is closely related to the proposed application scenario, which is conducive to the accurate measurement of the solder penetration rate of the final microwave component welding.

[0096] It can be understood that the above method of putting the predicted detection box into the accurate coordinate change to predict the coordinate probability distribution greatly improves the generalization ability of the model in the case of fuzzy and uncertain empty areas, as well as the accuracy of the predicted detection box.

[0097]

[0098] As shown in Figure 4(b), IoU It represents the intersection-over-union ratio of the predicted detection box area and the real detection box area. ρ ( b, b gt ) represents the center distance between the predicted detection box and the true label detection box; c ( b,b gt ) represents the distance between the farthest two points of the predicted detection box and the real detection box; Indicates the aspect ratio of the real label detection box, Indicates the aspect ratio of the predicted detection box; α represents the equilibrium parameter; v Represents the correction factor.

[0099] Next, the back-propagation algorithm is used to calculate the gradient until the model converges.

[0100] S70. Evaluate model performance using the validation set and adjust model hyperparameters.

[0101] The solder void rate of the specified device area on the image in the validation set is calculated, and the standard deviation is calculated based on the actual void rate and the predicted void rate calculated by the model:

[0102]

[0103] in, σ is the standard deviation of the prediction, x i For the validation set i The actual void ratio of the image, The validation set for model estimation i The predicted void ratio of the image, n is the number of images in the validation set.

[0104] when σ≤ 2 % , it is considered that the trained YoloV8 model has good performance and can be used for subsequent void rate measurement. Otherwise, it returns to step S30, adjusts the initial parameters before model training, and increases the number of images in the training set.

[0105] S80, select multiple models that perform well on the validation set, perform a final evaluation on the test set, output model performance indicators such as accuracy, recall, F1 score, etc., and select the best performing model as the final YoloV8 model.

[0106] After obtaining the final YoloV8 model, this step uses the preprocessed X-ray image as the input of the model to detect the hole area on the specified device area, and performs instance segmentation on each hole area to obtain a pixel-level mask image; among which, the sizes of the hole areas in different device areas vary in magnitude.

[0107] In step S3, the pixel area of ​​each hole region is obtained based on the mask image.

[0108] In this step, based on the mask image, the pixel area of ​​each hole region is obtained using the polygon area formula; the polygon area formula is:

[0109]

[0110] in, Area Represents the area of ​​a polygon; N b Represents the number of boundary points of the polygon surrounding the hole area on the mask image, ( X i ,Y i )and( X i+1 ,Y i+1 ) represent the i、i+ The xy coordinates of a boundary point.

[0111] In step S4, based on the pixel area of ​​each void region, the solder void rate on the designated device region is obtained to obtain the solder penetration rate of the designated device region.

[0112] In this step, the sum of the pixel areas of all void areas on the specified device area is counted, combined with the pixel area of ​​the specified device area itself, to calculate the welding void rate on the specified device area, and then the welding penetration rate of the specified device area is calculated using the formula "welding penetration rate = 1-welding void rate".

[0113] Embodiment 2:

[0114] The embodiment of the present invention provides a microwave component welding solder penetration rate measuring device based on deep learning, comprising:

[0115] An image preprocessing module, used for acquiring and preprocessing an X-ray image of the microwave component to be tested;

[0116] A hole segmentation module is used to use the preprocessed X-ray image as the input of a preset YoloV8 model to detect the hole area on the specified device area, and perform instance segmentation on each of the hole areas to obtain a pixel-level mask image; wherein the size of the hole areas in different device areas has order of magnitude differences;

[0117] A pixel area calculation module, used for obtaining the pixel area of ​​each of the hole regions based on the mask image;

[0118] The solder penetration rate acquisition module is used to acquire the solder void rate on the designated device area based on the pixel area of ​​each of the void areas, so as to acquire the solder penetration rate of the designated device area.

[0119] Embodiment 3:

[0120] like Figure 5 As shown, an embodiment of the present invention provides a microwave component welding permeability measurement system based on deep learning, including an X-ray tube 1, a stage 2, a flat panel detector 3 and a computer workstation 4.

[0121] The X-ray tube 1 is used to emit X-rays excited by high-energy electrons to the microwave component to be tested placed on the stage 2 .

[0122] The flat panel detector 3 is used to receive X-ray radiation that penetrates the microwave component to be tested, generate an electrical signal proportional to the radiation intensity, and display it through an X-ray image.

[0123] The computer workstation 4 is used to execute the microwave component welding permeability measurement method as described in Example 1.

[0124] In an optional embodiment, in order to penetrate the thicker chip substrate in the assembly, the maximum voltage of the X-ray tube 1 is required to reach 150 kV or more. In order to ensure that welding voids can be effectively identified, the resolution of the X-ray tube must be less than 1 um.

[0125] In an optional embodiment, in order to ensure the resolution and contrast of the acquired image, the exposure time of the X-ray tube 1 is set to 200ms-400ms to avoid the image being too dark or overexposed.

[0126] In an optional embodiment, the computer workstation 4 is installed with a high-performance GPU graphics card for image acceleration processing, and its performance is at least above GeForce RTX 4080 16G.

[0127] It can be understood that the microwave component welding permeability measurement device based on deep learning provided in Example 2 and the microwave component welding permeability measurement system based on deep learning provided in Example 3 correspond to the microwave component welding permeability measurement method based on deep learning provided in Example 1. The explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the microwave component welding permeability measurement method based on deep learning, and will not be repeated here.

[0128] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0129] 1. The embodiment of the present invention adopts a deep learning method to measure welding penetration rate, which bids farewell to the method of judging void pixels by a single threshold. The void calibration accurate to pixels and the powerful generalization ability of deep learning ensure the accuracy of void position and area prediction.

[0130] 2. The embodiment of the present invention uses a deep learning method to predict the welding void rate. Since this method predicts the void pixel point by point, the convolution operation can be used to extract the correlation information between pixels of different scales. At the same time, different equipment areas are added to the training data set of the model to improve the generalization ability of the model.

[0131] 3. Compared with other deep learning models, the latest YoloV8 detection model used in the embodiment of the present invention has a backbone network and head network structure that have been technically optimized, greatly improving the speed of image processing. At the same time, the refined loss function design makes the regression prediction of hole pixels and detection frames more accurate.

[0132] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0133] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for measuring the penetration rate of microwave component welding based on deep learning, characterized in that: include: Acquire and preprocess X-ray images of the microwave components to be tested; The preprocessed X-ray image is used as the input of the preset YoloV8 model to detect the hole area on the specified device area, and each of the hole areas is instance segmented to obtain a pixel-level mask image; wherein the size of the hole areas in different device areas varies in magnitude; Based on the mask image, obtaining a pixel area of ​​each of the hole regions; Based on the pixel area of ​​each of the void regions, a welding void rate on the designated device region is obtained to obtain a welding penetration rate of the designated device region; The training process of the YoloV8 model includes: Acquire and preprocess multiple historical X-ray images of microwave components; Perform image enhancement on the preprocessed historical X-ray images; On each enhanced image, four-point rectangular frames are marked on the hole areas on different device regions, and contours are marked on the hole areas on different device regions; Divide all annotated images into training set, validation set and test set; Set the hyperparameters of the YoloV8 model; The training set is used as the input of the YoloV8 model, the loss function is calculated, and the gradient is calculated using the back propagation algorithm until the model converges; Using the validation set to evaluate model performance and adjust model hyperparameters; Select multiple models that perform well on the validation set, perform a final evaluation on the test set, output model performance indicators respectively, and select the best performing model as the final YoloV8 model.

2. The microwave component welding permeability measurement method according to claim 1, characterized in that: The preprocessing includes image sharpening processing, histogram equalization processing and multi-image averaging processing.

3. The microwave component welding brazing penetration measurement method according to claim 1, characterized in that: The image enhancement is mosaic enhancement and / or mixed enhancement.

4. The microwave component welding permeability measurement method according to claim 1, characterized in that: The loss function includes a category classification loss function and a frame regression loss function.

5. The microwave component welding brazing permeability measurement method according to claim 4, characterized in that: The category classification loss function adopts binary cross entropy loss, and / or the bounding box regression loss function adopts distribution focus loss and CIOU loss.

6. The microwave component welding permeability measurement method according to claim 1, characterized in that: Based on the mask image, the pixel area of ​​each hole region is obtained using a polygon area formula; the polygon area formula is: in, Area Represents the area of ​​a polygon; N b Represents the number of boundary points of the polygon surrounding the hole area on the mask image, ( X i ,Y i )and( X i+1 ,Y i+1 ) represent the i、i+ The xy coordinates of a boundary point.

7. A microwave component welding permeability measurement device based on deep learning, characterized in that: The method for measuring the penetration rate of microwave component welding according to claim 1 comprises: An image preprocessing module, used for acquiring and preprocessing an X-ray image of the microwave component to be tested; A hole segmentation module is used to use the preprocessed X-ray image as the input of a preset YoloV8 model to detect the hole area on the specified device area, and perform instance segmentation on each of the hole areas to obtain a pixel-level mask image; wherein the size of the hole areas in different device areas has order of magnitude differences; A pixel area calculation module, used for obtaining the pixel area of ​​each of the hole regions based on the mask image; The solder penetration rate acquisition module is used to acquire the solder void rate on the designated device area based on the pixel area of ​​each of the void areas, so as to acquire the solder penetration rate of the designated device area.

8. A microwave component welding penetration measurement system based on deep learning, characterized in that: It includes an X-ray tube (1), a stage (2), a flat panel detector (3) and a computer workstation (4); The X-ray tube (1) is used to emit X-rays excited by high-energy electrons toward a microwave component to be tested placed on a stage (2); The flat panel detector (3) is used to receive X-ray radiation that penetrates the microwave component to be tested, generate an electrical signal proportional to the radiation intensity, and display it through an X-ray image; The computer workstation (4) is used to execute the microwave component welding permeability measurement method as claimed in any one of claims 1 to 6.

Images