Performance analysis method and system of bonding wire for LED packaging

Abstract

The invention belongs to the technical field of detection of bonding wires for LED packaging, and particularly relates to a performance analysis method of a bonding wire for LED packaging, which comprises the following steps of: 1, constructing a mechanical detection module, performing mechanical detection on the bonding wire for LED packaging by using a stress detection device, judging whether abnormity occurs or not according to a detection result, and performing the next step if no abnormity occurs; 2, constructing an X-ray optical detection module, and carrying out X-ray imaging on the bonding wire for LED packaging by adopting an X-ray imaging device to obtain an X-ray image of a bonding wire welding spot; and step 3, in the X-ray optical detection module, firstly preprocessing the image, constructing an improved U-Net convolutional neural network model to perform quality performance analysis detection, and obtaining a final analysis result of the bonding wire for LED packaging. According to the method, X-ray image detection is carried out on the bonding wire welding spot for LED packaging, image segmentation and feature extraction are carried out based on the digital image and the U-Net convolutional neural network model, and therefore the accuracy of an analysis result is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of bonding wire detection for LED packaging, and in particular relates to a performance analysis method and system for bonding wires for LED packaging. Background Art

[0002] Semiconductor lighting covers many industries such as general lighting, landscape lighting, and automotive lighting. It is another light source revolution following the traditional lighting source, and its core is LED (Light Emitting Diode); luminous flux attenuation failure is an important indicator for judging the performance of LED lighting products; in the wire bonding process, the influence of bonding parameters on bonding quality is crucial. Process parameters such as ultrasonic energy and ultrasonic welding pressure will directly affect the packaging quality of LED chips. The unique oxidation resistance, high hardness, and balling complexity of bonding wires make bonding wires more demanding on the bonding process. Therefore, a performance analysis method and system for bonding wires for LED packaging is needed.

[0003] In the prior art, for example, Chinese patent number CN118483037A discloses a bonding wire quality detection device and detection method, discloses a bonding wire quality detection device, including a bracket, and also includes: a lifting plate lifted and lowered on the bracket, wherein a tension detection component is fixedly arranged at the bottom of the lifting plate, and two sets of fixed rollers are respectively provided with through holes, and the axes of the two sets of through holes coincide; a clamping plate is slidably arranged on the fixed seat, a clamping pad is fixedly arranged on the inner wall of the clamping plate, and a plurality of capsules are fixedly arranged on the inner wall of the clamping pad; in the prior art, by opening coaxial The through hole can make the axis of the sample consistent with the stretching direction, thereby improving the accuracy of the test data. In addition, an arc-shaped clamping pad is used to clamp the sample wrapped on the fixed roller. The clamping pad can wrap the sample on the fixed roller and fully contact the surface of the sample, so that the original line contact becomes surface contact. The existing technology only performs quality inspection on the bonding wire from a mechanical perspective, and is unable to perform all-round inspection on the bonding wire on the already welded LED. For this reason, we propose a performance analysis method and system for bonding wire for LED packaging, which can adapt to LED production lines and improve detection efficiency and production efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide a performance analysis method and system for LED packaging bonding wires, which construct a mechanical detection module, use a stress detection device to perform mechanical detection on the LED packaging bonding wires, and determine whether an abnormality occurs according to the detection result, and if no abnormality occurs, proceed to the next step; if an abnormality occurs, secondary reinforcement is performed for the LED packaging bonding wire solder joints with abnormal mechanical detection, and step 2 is directly performed after the secondary reinforcement; then, an X-ray optical detection module is constructed, and an X-ray imaging device is used to perform X-ray imaging on the LED packaging bonding wires to obtain X-ray images of the bonding wire solder joints; finally, in the X-ray optical detection module, the image is first preprocessed, and an improved U-Net convolutional neural network model is constructed to perform quality performance analysis and detection, and the final analysis results of the LED packaging bonding wires are obtained; by performing X-ray image detection on the LED packaging bonding wire solder joints, image segmentation and feature extraction are performed based on digital images and the U-Net convolutional neural network model, thereby improving the accuracy of the analysis results of the LED packaging bonding wire solder joints and improving the detection efficiency.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A performance analysis method for a bonding wire for LED packaging, comprising the following steps:

[0007] Step 1: Build a mechanical detection module, use a stress detection device to perform mechanical detection on the bonding wire for LED packaging, and determine whether there is an abnormality based on the detection results. If there is no abnormality, proceed to the next step;

[0008] Step 2: construct an X-ray optical detection module, use an X-ray imaging device to perform X-ray imaging on the bonding wire for LED packaging, and obtain an X-ray image of the bonding wire solder joint;

[0009] Step 3: In the X-ray optical detection module, the image is first preprocessed, and an improved U-Net convolutional neural network model is constructed to perform quality performance analysis and detection, and the final analysis results of the bonding wire for LED packaging are obtained.

[0010] Preferably, in step 1, if an abnormality occurs, secondary reinforcement is performed on the bonding wire solder joints for LED packaging that show abnormality in mechanical detection, and step 2 is performed directly after the secondary reinforcement.

[0011] Preferably, the step 3 specifically includes the following steps:

[0012] Step 301: Use the three-dimensional Otsu algorithm to segment the X-ray image of the bonding wire solder joint; and output the weld area image;

[0013] Step 302: constructing an improved U-Net convolutional neural network model. First, the acquired X-ray images of bonding wire solder joints are divided into a training set and a test set. First, the U-Net model is constructed using the training set, and a dense conditional random field is introduced. After the training is completed, the improved U-Net convolutional neural network model is obtained.

[0014] Step 303: Use the improved U-Net convolutional neural network model to process the test set and output the test results, which are performance analysis results of the bonding wire for LED packaging.

[0015] Preferably, in step 301, for the M×N bonding wire solder joint X-ray image, f(x,y) is used to represent the grayscale value of the pixel at (x,y), and g(x,y) is used to represent the average grayscale value in the k×k neighborhood at (x,y), and g(x,y) is defined as follows:

[0016]

[0017] Let h(x,y) represent the average grayscale median in the k*k neighborhood at (x,y). h(x,y) is defined as follows:

[0018]

[0019] According to the corresponding proportion of the welding area in the whole image, a factor K is set, and the corresponding category with large variance is corrected by factor K. The purpose of this is to adjust the weight of the threshold calculation formula, so as to avoid the threshold biased towards the category with large variance, that is, to avoid the optimal threshold being too high. The improved optimal threshold formula is:

[0020]

[0021] Where K represents the proportion of the welding area to the entire image.

[0022] Preferably, in step 302, when constructing the improved U-Net convolutional neural network model, a spatial attention mechanism is introduced to enhance the extraction of defect position features. The spatial attention mechanism (SAM) first performs a maximum pooling and average pooling operation on the feature map F to obtain a two-layer compressed feature map. The two layers of feature maps are superimposed and then a convolution operation is performed. The result of the convolution is normalized by an activation function to obtain a spatial weight matrix M, which is multiplied by the input feature map to finally obtain a new feature map. The operation process of the spatial attention mechanism is: M S (F) = δ(f 3×3[AvgPool(F); MaxPool(F)]); where: F is the input feature map; δ is the activation function; f is the convolution kernel of size 3×3; AvgPool is average pooling; Max-Pool is maximum pooling;

[0023] At the same time, since the defect targets are small, low-dimensional features can better preserve the location information of the features, so the spatial attention mechanism is applied to the low-dimensional feature extraction part of the network.

[0024] Preferably, when constructing the improved U-Net convolutional neural network model, a spatial pyramid pooling module is constructed, and the spatial pyramid pooling module uses three maximum pooling layers with pooling kernel sizes of 13×13, 9×9, and 5×5, respectively, to perform a maximum pooling operation on the feature map, and stack the original feature map with the local feature map obtained by the pooling process, so that the receptive field of the local area of ​​the feature map and the receptive field information close to the global are fused, thereby separating the most significant contextual features and obtaining richer local feature information.

[0025] Preferably, in step 302, a dense conditional random field is introduced to determine whether two pixels belong to the same category by calculating the similarity between them. In the model test, the probability distribution map output by the decoder is used as the unary potential energy of the dense conditional random field, and the position and color information in the binary potential energy are provided by the original image. The processing result of the dense conditional random field is used as the final output result;

[0026] Dense Conditional Random Fields define a N}, the domain of X is the label set L = {l1, l2, ..., l k}. Another one is defined in the variables {I1, I2, ..., I N}, where: N is the number of pixels in the input image; I j is the color vector of pixel j; X j is the label of pixel j; the energy function of dense conditional random field is:

[0027]

[0028] The value range of i and j is 1~N. The first term of the energy equation ψu(x i) is a unary potential energy function, which is calculated independently by the classifier and is used to measure whether the pixel i belongs to the category label x i The probability of; the one-dimensional potential used in reality contains the description of shape, texture, position and color; the second term of the energy equation is the binary potential energy function ψp(x i , x j)It is used to calculate the relationship between each pixel and other pixels, and classify pixels with close colors and relative distances into one category. The calculation formula is:

[0029]

[0030] Where: μ(x i ,x j ) is the label probability function, which calculates the probability that pixel i and pixel j belong to the same class. i ≠x j , then μ(x i , x j )=1, otherwise 0; I i and I j is the pixel position p i and p j The color vector on w (1) and w (2) It is a linear combination weight used to adjust the weight of the appearance kernel and the smoothing kernel. The appearance kernel assumes that adjacent pixels with similar colors are likely to belong to the same category, and the role of the smoothing kernel is to eliminate isolated small areas.

[0031] Preferably, in step 303, after acquiring the X-ray image, the data is further divided into a validation set, and the data is divided into three non-overlapping parts: a training set, a validation set, and a test set. In the model test, the trained improved U-Net convolutional neural network model is used to output a probability distribution map, which is used as a unary potential energy input of a dense conditional random field, and the bubble defect segmentation result is output after being processed by the dense conditional random field.

[0032] A performance analysis system for bonding wires for LED packaging, the analysis system comprising a mechanical detection module and an X-ray optical detection module, the mechanical detection module is used to detect the mechanical properties of the bonding wires for LED packaging after welding, and the X-ray optical detection module completes quality analysis by collecting X-ray optical images of the welding points.

[0033] Preferably, an improved U-Net convolutional neural network model is constructed in the X-ray optical detection module, and the improved U-Net convolutional neural network model is used to complete the quality analysis of the bonding wire for LED packaging based on the acquired X-ray optical image.

[0034] The technical effects achieved by the present invention are:

[0035] In the present invention, firstly, a mechanical detection module is constructed, and a stress detection device is used to perform mechanical detection on the LED packaging bonding wire, and it is determined whether an abnormality occurs according to the detection result, and if no abnormality occurs, the next step is performed; if an abnormality occurs, secondary reinforcement is performed for the LED packaging bonding wire solder joints with abnormal mechanical detection, and step 2 is directly performed after the secondary reinforcement; then, an X-ray optical detection module is constructed, and an X-ray imaging device is used to perform X-ray imaging on the LED packaging bonding wire to obtain an X-ray image of the bonding wire solder joint; finally, in the X-ray optical detection module, the image is first preprocessed, and an improved U-Net convolutional neural network model is constructed to perform quality performance analysis and detection, and the final analysis result of the LED packaging bonding wire is obtained; by performing X-ray image detection on the LED packaging bonding wire solder joints, image segmentation and feature extraction are performed based on digital images and the U-Net convolutional neural network model, thereby improving the accuracy of the analysis result of the LED packaging bonding wire solder joints and improving the detection efficiency.

[0036] In the present invention, in order to solve the problem of insufficient feature extraction capability and poor generalization of the classic U-Net network, the present invention introduces a dense conditional random field Mobile-Net as the backbone feature extraction network of U-Net, improves the network's ability to obtain defect shape and position information, reduces the number of network parameters, and reduces the model's requirements for the amount of training samples; secondly, a spatial attention mechanism is introduced into the low-dimensional feature extraction part of the dense conditional random field Mobile-Net, and spatial pyramid pooling is introduced after feature extraction to improve the network's ability to extract high- and low-dimensional features of the image; after decoding, in order to solve the problem of feature information loss caused by the decoder upsampling layer, a dense conditional random field is introduced after classification is completed, and the classification results of the pixels are re-evaluated in combination with the pixel values ​​and category information of the pixels, so as to further improve the segmentation accuracy; finally, experiments are carried out on a data set of bonding wire solder joint defects for LED packaging to verify the effectiveness of the improved U-Net convolutional neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a performance analysis method of a bonding wire for LED packaging of the present invention;

[0038] Figure 2 It is a flow chart of step 3 in a performance analysis method of a bonding wire for LED packaging of the present invention;

[0039] Figure 3 It is a system block diagram of a performance analysis system for a bonding wire for LED packaging of the present invention;

[0040] Figure 4 The invention discloses an operation block diagram of an improved U-Net convolutional neural network model in a performance analysis system of a bonding wire for LED packaging. DETAILED DESCRIPTION

[0041] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0042] like Figure 1-4 As shown, a performance analysis method for a bonding wire for LED packaging comprises the following steps:

[0043] Step 1: Construct a mechanical detection module, use a stress detection device to perform mechanical detection on the LED package bonding wire, and determine whether there is an abnormality based on the detection results. If there is no abnormality, proceed to the next step; if there is an abnormality, perform secondary reinforcement on the LED package bonding wire solder joints with abnormal mechanical detection, and directly proceed to step 2 after the secondary reinforcement;

[0044] Step 2: construct an X-ray optical detection module, use an X-ray imaging device to perform X-ray imaging on the bonding wire for LED packaging, and obtain an X-ray image of the bonding wire solder joint;

[0045] Step 3: In the X-ray optical inspection module, the image is first preprocessed, and an improved U-Net convolutional neural network model is constructed to perform quality performance analysis and inspection, and the final analysis results of the bonding wire for LED packaging are obtained.

[0046] It should be noted that the electrical performance test of the bonding wire is performed before welding the bonding wire, so it is no longer necessary to perform electrical performance test on the bonding wire here.

[0047] like Figure 1-4 As shown, in step 1, if an abnormality occurs, secondary reinforcement is performed on the LED package bonding wire solder joints with abnormal mechanical detection, and step 2 is directly performed after the secondary reinforcement. In the present invention, by first performing a mechanical test on the LED package bonding wire, if the test result is abnormal, secondary welding is required on the solder joints, so that the bonding wires of all samples can meet the mechanical performance requirements, thereby improving the qualified rate of the product.

[0048] In the present invention, X-ray image detection is performed on the solder joints of the bonding wires for LED packaging, and image segmentation and feature extraction are performed based on digital images and a U-Net convolutional neural network model, thereby improving the accuracy of the analysis results of the solder joints of the bonding wires for LED packaging.

[0049] Step 3 specifically includes the following steps:

[0050] Step 301: Use the three-dimensional Otsu algorithm to segment the X-ray image of the bonding wire solder joint; and output the weld area image;

[0051] Step 302: constructing an improved U-Net convolutional neural network model. First, the acquired X-ray images of bonding wire solder joints are divided into a training set and a test set. First, the U-Net model is constructed using the training set, and a dense conditional random field is introduced. After the training is completed, the improved U-Net convolutional neural network model is obtained.

[0052] Step 303: Use the improved U-Net convolutional neural network model to process the test set and output the test results, which are performance analysis results of the bonding wire for LED packaging.

[0053] like Figure 1-4 As shown, in the present invention, in order to solve the problem of insufficient feature extraction ability and poor generalization of the classic U-Net network, the present invention introduces a dense conditional random field Mobile-Net as the backbone feature extraction network of U-Net, improves the network's ability to obtain defect shape and position information, reduces the number of network parameters, and reduces the model's requirements for the amount of training samples; secondly, a spatial attention mechanism is introduced into the low-dimensional feature extraction part of the dense conditional random field Mobile-Net, and spatial pyramid pooling is introduced after feature extraction to improve the network's ability to extract high- and low-dimensional features of the image. After decoding, in order to solve the problem of feature information loss caused by the decoder upsampling layer, a dense conditional random field is introduced after classification is completed, and the classification results of the pixels are re-evaluated in combination with the pixel values ​​and category information of the pixels, so as to further improve the segmentation accuracy; finally, experiments are conducted on a data set of bonding wire solder joint defects for LED packaging to verify the effectiveness of the improved U-Net convolutional neural network model.

[0054] In step 301, for the M×N bonding wire joint X-ray image, f(x,y) is used to represent the gray value of the pixel at (x,y), and g(x,y) is used to represent the average gray value in the k×k neighborhood at (x,y). The definition of g(x,y) is as follows:

[0055]

[0056] Let h(x,y) represent the average grayscale median in the k*k neighborhood at (x,y). h(x,y) is defined as follows:

[0057]

[0058] According to the corresponding proportion of the welding area in the whole image, a factor K is set, and the corresponding category with large variance is corrected by factor K. The purpose of this is to adjust the weight of the threshold calculation formula, so as to avoid the threshold biased towards the category with large variance, that is, to avoid the optimal threshold being too high. The improved optimal threshold formula is:

[0059]

[0060] Where K represents the proportion of the welding area to the entire image.

[0061] When selecting the optimal value, the three-dimensional Otsu algorithm traverses the entire image during calculation, that is, traverses from gray level 0 to gray level L-1, which brings certain redundant calculations. If the range of gray level traversal can be reduced, the running time can be reduced. Figure 1 It generally presents a double-peak shape, and the segmentation threshold is generally near the valley; the proportion of the weld area in the entire image is generally less than 1 / 2, and the weld area is generally the darker type, and its segmentation threshold is generally in the (α, β) interval, α is the minimum gray value of the weld image, and β is the average gray value of the weld image; that is, this can reduce the traversal range of the weld image gray level and thus reduce the amount of calculation, thereby improving the calculation efficiency.

[0062] In step 302, when constructing the improved U-Net convolutional neural network model, a spatial attention mechanism is introduced to enhance the extraction of defect position features. The spatial attention mechanism (SAM) first performs a maximum pooling and average pooling operation on the feature map F to obtain a two-layer compressed feature map. The two layers of feature maps are superimposed and then a convolution operation is performed. The convolution result is normalized by the activation function to obtain a spatial weight matrix M, which is multiplied with the input feature map to finally obtain a new feature map. The operation process of the spatial attention mechanism is: M S (F) = δ(f 3×3 [AvgPool(F); MaxPool(F)]); where: F is the input feature map; δ is the activation function; f is the convolution kernel of size 3×3; AvgPool is average pooling; Max-Pool is maximum pooling;

[0063] At the same time, since the defect targets are small, low-dimensional features can better preserve the location information of the features, so the spatial attention mechanism is applied to the low-dimensional feature extraction part of the network.

[0064] When constructing the improved U-Net convolutional neural network model, a spatial pyramid pooling module is constructed. The spatial pyramid pooling module uses three maximum pooling layers with pooling kernel sizes of 13×13, 9×9, and 5×5, respectively, to perform maximum pooling operations on the feature maps, and stack the original feature maps with the local feature maps obtained by pooling processing, so that the receptive field of the local area of ​​the feature map and the receptive field information close to the global area are fused, thereby separating the most significant contextual features and obtaining richer local feature information.

[0065] In step 302, a dense conditional random field is introduced to determine whether two pixels belong to the same category by calculating the similarity between them. In the model test, the probability distribution map output by the decoder is used as the unary potential energy of the dense conditional random field, and the position and color information in the binary potential energy are provided by the original image. The processing result of the dense conditional random field is used as the final output result;

[0066] Dense Conditional Random Fields define a N}, the domain of X is the label set L = {l1, l2, ..., l k}. Another one is defined in the variables {I1, I2, ..., I N}, where: N is the number of pixels in the input image; I j is the color vector of pixel j; X j is the label of pixel j; the energy function of dense conditional random field is:

[0067]

[0068] The value range of i and j is 1~N. The first term of the energy equation ψu(x i) is a unary potential energy function, which is calculated independently by the classifier and is used to measure whether the pixel i belongs to the category label x i The one-dimensional potential used in reality contains the description of shape, texture, position and color. The second term of the energy equation is the binary potential energy function ψp(x i , x j) It is used to calculate the relationship between each pixel and other pixels, and classify pixels with close colors and relative distances into one category. The calculation formula is:

[0069]

[0070] Where: μ(x i ,x j ) is the label probability function, which calculates the probability that pixel i and pixel j belong to the same class. i ≠x j , then μ(x i , x j )=1, otherwise 0; I i and I j is the pixel position p i and p j The color vector on w (1) and w (2) It is a linear combination weight, which is used to adjust the weight of the appearance kernel and the smoothing kernel. The appearance kernel assumes that adjacent pixels with similar colors are likely to belong to the same category, and the role of the smoothing kernel is to eliminate isolated small areas;

[0071] The function of the above formula is to determine whether similar pixels belong to the same category. When the pixels belong to the same category, the value of the energy function is small; in the X-ray chip image segmentation, the edge where the weld bubble defect contacts the welding surface is difficult to accurately divide, which affects the subsequent measurement of the defect size. Using this energy function, the defect and the welding surface segmentation is more accurate. The pixel values ​​inside the bubble defect are similar, and there are differences in the pixel values ​​of the background and the bubble defect. When the pixels in similar areas are judged as different categories or the areas with differences are judged as the same category, a larger energy value will be generated. Through multiple inferences, the value of the energy function is minimized to obtain the final result. In this way, the information of the entire image is used to refine the defect edge and improve the accuracy of segmentation.

[0072] In step 303, after acquiring the X-ray image, the data is also divided into a validation set, and the data is divided into three non-overlapping parts: a training set, a validation set, and a test set. In the model test, the trained improved U-Net convolutional neural network model is used to output a probability distribution map, which is used as the unary potential energy input of the dense conditional random field, and the bubble defect segmentation result is output after being processed by the dense conditional random field.

[0073] like Figure 3-4 As shown, a performance analysis system for bonding wires for LED packaging, the analysis system includes a mechanical detection module and an X-ray optical detection module. The mechanical detection module is used to detect the mechanical properties of the bonding wires for LED packaging after welding, and the X-ray optical detection module completes quality analysis by collecting X-ray optical images of the welding points.

[0074] Preferably, an improved U-Net convolutional neural network model is constructed in the X-ray optical detection module, and the improved U-Net convolutional neural network model is used to complete the quality analysis of the bonding wire for LED packaging based on the acquired X-ray optical image.

[0075] In the present invention, firstly, a mechanical detection module is constructed, and a stress detection device is used to perform mechanical detection on the LED packaging bonding wire, and it is determined whether an abnormality occurs according to the detection result, and if no abnormality occurs, the next step is performed; if an abnormality occurs, secondary reinforcement is performed for the LED packaging bonding wire solder joints with abnormal mechanical detection, and step 2 is directly performed after the secondary reinforcement; then, an X-ray optical detection module is constructed, and an X-ray imaging device is used to perform X-ray imaging on the LED packaging bonding wire to obtain an X-ray image of the bonding wire solder joint; finally, in the X-ray optical detection module, the image is first preprocessed, and an improved U-Net convolutional neural network model (SSMob-U-Net model) is constructed to perform quality performance analysis and detection, and the final analysis result of the LED packaging bonding wire is obtained; by performing X-ray image detection on the LED packaging bonding wire solder joints, image segmentation and feature extraction are performed based on digital images and the U-Net convolutional neural network model, thereby improving the accuracy of the analysis result of the LED packaging bonding wire solder joints.

[0076] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A performance analysis method for bonding wires for LED packaging, characterized in that: The following steps are involved: Step 1: Build a mechanical detection module, use a stress detection device to perform mechanical detection on the bonding wire for LED packaging, and determine whether there is an abnormality based on the detection results. If there is no abnormality, proceed to the next step; Step 2: construct an X-ray optical detection module, use an X-ray imaging device to perform X-ray imaging on the bonding wire for LED packaging, and obtain an X-ray image of the bonding wire solder joint; Step 3: In the X-ray optical detection module, the image is first preprocessed, and an improved U-Net convolutional neural network model is constructed to perform quality performance analysis and detection, and the final analysis results of the bonding wire for LED packaging are obtained.

2. The performance analysis method of a bonding wire for LED packaging according to claim 1, characterized in that: In the step 1, if an abnormality occurs, secondary reinforcement is performed on the LED package bonding wire solder joints with abnormal mechanical detection, and step 2 is directly performed after the secondary reinforcement.

3. The performance analysis method of a bonding wire for LED packaging according to claim 1, characterized in that: The step 3 specifically includes the following steps: Step 301: Use the three-dimensional Otsu algorithm to segment the X-ray image of the bonding wire solder joint; and output the weld area image; Step 302: constructing an improved U-Net convolutional neural network model. First, the acquired X-ray images of bonding wire solder joints are divided into a training set and a test set. First, the U-Net model is constructed using the training set, and a dense conditional random field is introduced. After the training is completed, the improved U-Net convolutional neural network model is obtained. Step 303: Use the improved U-Net convolutional neural network model to process the test set and output the test results, which are performance analysis results of the bonding wire for LED packaging.

4. The performance analysis method of a bonding wire for LED packaging according to claim 3, characterized in that: In step 301, for the M×N bonding wire joint X-ray image, f(x,y) is used to represent the gray value of the pixel at (x,y), and g(x,y) is used to represent the average gray value in the k×k neighborhood at (x,y). The definition of g(x,y) is as follows: Let h(x,y) represent the average grayscale median in the k*k neighborhood at (x,y). h(x,y) is defined as follows: According to the corresponding proportion of the welding area in the whole image, a factor K is set, and the corresponding class is corrected using the factor K. The improved optimal threshold formula is: Where K represents the proportion of the welding area to the entire image.

5. The performance analysis method of a bonding wire for LED packaging according to claim 4, characterized in that: In step 302, when constructing the improved U-Net convolutional neural network model, a spatial attention mechanism is introduced to enhance the extraction of defect position features. The spatial attention mechanism (SAM) first performs a maximum pooling and an average pooling operation on the feature map F to obtain a two-layer compressed feature map. The two layers of feature maps are superimposed and then a convolution operation is performed. The result of the convolution is normalized by an activation function to obtain a spatial weight matrix M, which is multiplied by the input feature map to finally obtain a new feature map. The operation process of the spatial attention mechanism is: M S (F) = δ(f 3×3 [AvgPool(F); MaxPool(F)]); where: F is the input feature map; δ is the activation function; f is the convolution kernel of size 3×3; AvgPool is average pooling; Max-Pool is maximum pooling.

6. The performance analysis method of a bonding wire for LED packaging according to claim 5, characterized in that: When constructing the improved U-Net convolutional neural network model, a spatial pyramid pooling module is constructed. The spatial pyramid pooling module uses three maximum pooling layers with pooling kernel sizes of 13×13, 9×9, and 5×5, respectively, performs a maximum pooling operation on the feature map, and stacks the original feature map with the local feature map obtained by the pooling process, so that the receptive field of the local area of ​​the feature map and the receptive field information close to the global are fused, thereby separating the contextual features and obtaining the local feature information.

7. The performance analysis method of a bonding wire for LED packaging according to claim 6, characterized in that: In step 302, a dense conditional random field is introduced to determine whether two pixels belong to the same category by calculating the similarity between them. In the model test, the probability distribution map output by the decoder is used as the unary potential energy of the dense conditional random field, and the position and color information in the binary potential energy are provided by the original image. The processing result of the dense conditional random field is used as the final output result; Dense Conditional Random Fields define a N }, the domain of X is the label set L = {l1, l2, ..., l k }. Another one is defined in the variables {I1, I2, ..., I N }, where: N is the number of pixels in the input image; I j is the color vector of pixel j; X j is the label of pixel j; the energy function of dense conditional random field is: The value range of i and j is 1~N. The first term of the energy equation ψu(x i) is a unary potential energy function, which is calculated independently by the classifier and is used to measure whether the pixel i belongs to the category label x i The probability of the second term of the energy equation is the binary potential energy function ψp(x i , x j) It is used to calculate the relationship between each pixel and other pixels, and classify pixels with close colors and relative distances into one category. The calculation formula is: Where: μ(x i ,x j ) is the label probability function, which calculates the probability that pixel i and pixel j belong to the same class. i ≠x j , then μ(x i , x j )=1, otherwise 0; I i and I j is the pixel position p i and p j The color vector on w (1) and w (2) It is a linear combination weight used to adjust the weight of the appearance kernel and the smoothing kernel. The appearance kernel assumes that adjacent pixels with similar colors are likely to belong to the same category, and the role of the smoothing kernel is to eliminate isolated small areas.

8. The performance analysis method of a bonding wire for LED packaging according to claim 7, characterized in that: In step 303, after acquiring the X-ray image, the data is also divided into a validation set, and the data is divided into three non-overlapping parts: a training set, a validation set, and a test set. In the model test, the trained improved U-Net convolutional neural network model is used to output a probability distribution map, which is used as a unary potential energy input of a dense conditional random field, and the bubble defect segmentation result is output after being processed by the dense conditional random field.

9. A performance analysis system for bonding wires for LED packaging, characterized in that: The analysis system includes a mechanical detection module and an X-ray optical detection module. The mechanical detection module is used to detect the mechanical properties of the bonding wire for LED packaging after welding, and the X-ray optical detection module completes quality analysis by collecting X-ray optical images of the welding point.

10. The performance analysis system for bonding wires for LED packaging according to claim 9, characterized in that: An improved U-Net convolutional neural network model is constructed in the X-ray optical detection module, and the improved U-Net convolutional neural network model is used to complete the quality analysis of the bonding wire for LED packaging according to the acquired X-ray optical image.

Images