A flat wire winding solder joint defect detection method that does not rely on defect data
By employing transfer learning and feature library comparison, the problem of scarce and unbalanced samples in the detection of weld defects in flat wire windings was solved, achieving efficient and robust detection and localization of weld defects.
Patent Information
- Application Number
- CN202410580747.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-11
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-05-11
AI Technical Summary
In the existing flat wire winding welding process, defective weld points are scarce and the sample types are unbalanced, resulting in high costs and the inability to identify unknown defects by traditional detection methods.
A defect-independent detection method is adopted. The target detection model is trained by transfer learning, the pre-trained ResNet50 network is used to extract solder joint features, a feature library is constructed, and the KNN algorithm is used to compare solder joint features to detect defects.
It enables efficient detection of solder joint defects when defect samples are scarce, reduces memory usage, enhances detection robustness and locates defect areas, and solves the sample imbalance problem of traditional methods.
Smart Images

Figure CN118505631B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a flat wire winding solder point defect detection method, in particular to a flat wire winding solder point defect detection method that does not rely on defect data. Background Art
[0002] In recent years, electric vehicle drive motors have widely adopted flat-wire permanent magnet synchronous motors (PMSMs). This design improves the motor's slot fill factor, significantly increasing its power density and efficiency. The flat-wire winding is manufactured by removing the insulating varnish from the ends of the enameled copper wire, bending it into a U-shape, inserting it into the stator slots, and then twisting the exposed copper ends together and laser welding them. Due to the highly reflective surface of copper, defects such as insufficient welds, weld pits, and weld beads may occur during the welding process. These defects affect the mechanical and electrical properties of the weld, posing a threat to motor safety. Therefore, weld quality monitoring is essential.
[0003] Flat wire winding welding fault detection technologies can be broadly categorized into two types: physical characteristics-based and image-based. Physical characteristics-based technologies primarily observe radiation signals, optical and acoustic characteristics during the welding process, or electrical characteristics after welding to assess weld quality. Image-based deep learning methods, as non-contact, non-destructive automated detection methods, offer high efficiency and are widely used for various defect detection applications.
[0004] Currently proposed methods for flat wire winding solder joint inspection focus on defect characteristics. However, two issues warrant attention: sufficient defect sample data is required to train the model, and the second is the need to label and classify the defect types in the samples. However, on flat wire winding welding production lines, due to the instability of the welding process, it is difficult to obtain a large number of defect samples. Some studies have manipulated welding conditions to specifically create defect samples, which increases costs. Moreover, when welding failures occur on the production line, it is difficult to fully describe the defect type of the solder joint. Different studies have used different descriptions and classifications of flat wire winding solder joint defects, making it difficult for the model to identify unknown defects. Summary of the Invention
[0005] In order to solve the problems of scarcity of defective solder joints and imbalance of defect category samples in the background art, the present invention provides a method for detecting defects in solder joints of flat wire windings that does not rely on defect data.
[0006] The technical solution adopted in the present invention is:
[0007] The flat wire winding weld defect detection method of the present invention, which does not rely on defect data, comprises:
[0008] 1) Several solder joint images on a flat wire winding welding production line are collected and their positions are marked to obtain several annotated solder joint images, and a production line image training set is constructed. Each annotated solder joint image in the production line image training set is cropped to obtain several single solder joint annotated images. After defects are annotated on each single solder joint annotated image, the annotated images of single solder joints with normal morphology are extracted to construct a single solder joint dataset.
[0009] 2) Establish a target detection model, input the production line image training set into the target detection model for training, and obtain a trained solder joint defect target detection model.
[0010] 3) The single solder joint dataset is input into the solder joint defect target detection model. After feature extraction, the solder joint defect target detection model obtains several normal solder joint features and constructs a feature library.
[0011] 4) Collect the production line solder joint images to be inspected, input the production line solder joint images to be inspected into the solder joint defect target detection model, the solder joint defect target detection model outputs the solder joint positions on the production line solder joint images to be inspected after processing, crop the production line solder joint images to be inspected into several single solder joint images according to the solder joint positions, input each single solder joint image into the solder joint defect target detection model, and the solder joint defect target detection model outputs the solder joint features; for the solder joint features of each single solder joint image, use the neighboring KNN (k-NearestNeighbor) algorithm to obtain the outlier value after processing based on the solder joint features and feature library, compare the outlier value with the preset threshold value, judge the defect situation of the solder joint in the single solder joint image and locate the defect area, so as to realize the solder joint defect detection of flat wire winding.
[0012] During defect detection, the solder joint defect target detection model is first used to detect the target position, and then the backbone network of the model is reused to extract the detailed features of the position for comparison to see if there are any abnormal defects.
[0013] In the step 1), the solder joint image is annotated with the solder joint position, and the position of each solder joint in the solder joint image is annotated; the solder joint annotated image is cropped, and the obtained single solder joint annotated image contains only one solder joint and has the same resolution; the single solder joint annotated image is annotated with defects, and if the solder joint in the single solder joint annotated image has a defect, it is annotated and classified as a single solder joint with abnormal morphology; if the solder joint in the single solder joint annotated image has no defect, it is annotated and classified as a single solder joint with normal morphology, that is, the current single solder joint annotated image is a single solder joint annotated image with normal morphology.
[0014] In the step 2), the target detection model is specifically established by replacing the backbone network in the YOLOv5 network with a ResNet50 network pre-trained based on the computer vision standard ImageNet dataset; when training the target detection model, the production line image training set is input into the target detection model for training, and the ResNet50 network in the target detection model is frozen by the transfer learning method. Only the neck and detection head parts in the target detection model are trained, and the model loss is calculated and back propagation is performed until the target detection model converges and the training is completed and the weights of the target detection model are saved, and finally a trained solder joint defect target detection model is obtained.
[0015] The ResNet50 network pre-trained based on the ImageNet dataset has the ability to extract surface features of general objects. By using the pre-trained ResNet50 network to replace the backbone network of the YOLOv5 model, feature information of different levels of flat wire winding solder joints can be extracted without retraining the backbone network. The features expressed at these levels, from low to high, range from details to semantics.
[0016] In step 3), the single solder joint dataset is input into the solder joint defect target detection model. For each single solder joint annotated image with normal morphology in the single solder joint dataset, the pre-trained ResNet50 network in the solder joint defect target detection model is used to extract features of the single solder joint annotated image. The single solder joint annotated image is sequentially downsampled in the pre-trained ResNet50 network to obtain feature maps of four different levels. The feature maps of the second and third levels are fused and then subjected to specific aggregation and flattening processing, and then saved in the preset original library in sequence. Then, it is sampled and constructed into a feature library The feature library contains the feature information of normal solder joints in each single solder joint annotated image.
[0017] The specific aggregation process is as follows:
[0018]
[0019]
[0020] Where f represents the aggregation function; represents a square area with the position (a, b) as the center and p as the side length; φ ij (m,n) represents the feature map generated by the i-th single solder joint annotation image with normal morphology at the j-th level in the single solder joint dataset in the square area Features at position (m,n); λ(m,n) represents the square area The weighting coefficient of the feature at the inner position (m,n); Indicates the square area The distance from the inner position (m,n) to the center (a,b).
[0021] use Represents the training dataset of normal solder joint images, n is the number of images. Image x i The backbone network is extracted into feature maps of different levels, using φ ij Represents image x i At level j (j∈{1,2,3,4}), the feature map It is a three-dimensional tensor with depth c, height h, and width w, using φ ij (a,b) represents the feature slice of the feature map at position (a,b), where a∈(1,2,...,h) and b∈(1,2,...w).
[0022] In order to increase the receptive field without reducing the resolution, the adjacent features of each pixel on the feature map are aggregated and the Represents a rectangular block area with (a, b) as the center and a side length of p. The feature representation of this area is as follows:
[0023]
[0024] Where (m,n) is the area The specific location in the region; f is the aggregation function used to aggregate the region The feature information of all locations in the .
[0025] In the step 4), each single solder joint image is input into the solder joint defect target detection model, and the pre-trained ResNet50 network in the solder joint defect target detection model is used to extract features of the single solder joint image. The feature maps of the second and third levels of the single solder joint image in the pre-trained ResNet50 network are fused to obtain the solder joint features.
[0026] In the step 4), for the solder joint features of each single solder joint image, according to the solder joint features and feature library, the neighboring features of the solder joint features of the single solder joint image to be detected are searched in the feature library using the neighboring KNN algorithm and an outlier is obtained. The outlier is compared with a preset threshold. When the outlier is less than or equal to the preset threshold, the solder joint in the single solder joint image is normal. When the outlier is greater than the preset threshold, the solder joint in the single solder joint image is abnormal, that is, the solder joint is defective. At this time, the defective area is located to realize solder joint defect detection of the flat wire winding.
[0027] The method of the present invention first trains a target detection model through transfer learning, storing the features extracted from normal solder joint samples via a backbone network to form a feature library. The detection phase consists of two steps. After the target detection model detects the location of the solder joint, the backbone network in the reused model re-extracts the features of the solder joint and compares them with the feature library to determine whether the solder joint is defective. Compared to traditional deep learning detection methods that require a large number of defective solder joint samples, the present invention can classify whether a solder joint is defective using only normal solder joint samples.
[0028] The beneficial effects of the present invention are:
[0029] 1. The method proposed in this paper detects defects in flat wire winding solder joints on production lines in two steps. First, the target detection model detects the solder joints to be inspected. Then, the backbone network is reused to extract the intermediate layer features of these solder joints. These features are then compared with the features in the feature library to determine whether defects exist. This method solves the problems of rare defective solder joints and imbalanced defect category samples in traditional deep learning methods.
[0030] 2. The present invention reduces memory usage by reusing the backbone network in the target detection model.
[0031] 3. The present invention aggregates the intermediate layer features to smooth the features between the pixels of the feature map, thereby enhancing the robustness of defect detection.
[0032] 4. Compared with the target detection algorithm, the present invention can locate the defective area of the solder joint to be inspected after detecting the solder joint to be inspected.
[0033] In summary, the method of the present invention integrates the feature library construction into the target detection model, does not rely on the flat wire winding solder joint defect data, realizes the solder joint defect detection, and solves the problem of poor detection effect caused by scarce defect samples and sample imbalance in production. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 This is a flow chart of the solder joint defect detection method of the present invention;
[0035] Figure 2 This is the target detection model diagram of the present invention;
[0036] Figure 3 Constructing a model graph for the feature library of the present invention;
[0037] Figure 4 This is a defect detection model diagram of the present invention;
[0038] Figure 5 This is a precision-recall curve of the target detection model for solder joints of the present invention;
[0039] Figure 6The following is a graph showing the area under the receiver operating characteristic curve (AUROC) of the defect detection model of the present invention at the image level and pixel level. Figure 6 (a) is the AUROC curve of the defect detection model of the present invention at the image level. Figure 6 (b) is a graph of the area under the receiver operating characteristic curve (AUROC) of the defect detection model of the present invention at the pixel level;
[0040] Figure 7 This is a result diagram of an example of defect detection of a weld to be inspected according to the present invention. DETAILED DESCRIPTION
[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0042] like Figure 1 As shown, the flat wire winding weld defect detection method of the present invention that does not rely on defect data is specifically as follows:
[0043] 1) Several solder joint images on a flat wire winding welding production line are collected and their positions are marked to obtain several annotated solder joint images, and a production line image training set is constructed. Each annotated solder joint image in the production line image training set is cropped to obtain several single solder joint annotated images. After defects are annotated on each single solder joint annotated image, the annotated images of single solder joints with normal morphology are extracted to construct a single solder joint dataset.
[0044] In step 1), the solder joint images are annotated with the solder joint positions, marking the location of each solder joint in the solder joint image. The annotated solder joint images are cropped, resulting in single solder joint annotated images containing only one solder joint and having consistent resolution. Defects are annotated on the single solder joint annotated images. If a solder joint in the single solder joint annotated image has a defect, it is labeled and classified as a single solder joint with abnormal morphology. If a solder joint in the single solder joint annotated image has no defect, it is labeled and classified as a single solder joint with normal morphology. In other words, the current single solder joint annotated image is a single solder joint annotated image with normal morphology. In specific implementations, each solder joint image contains approximately 20 winding solder joints, and the resolution after cropping is 224×224.
[0045] 2) Establish a target detection model, such as Figure 2 As shown in FIG, the production line image training set is input into the target detection model for training to obtain a trained solder joint defect target detection model.
[0046] In step 2), the target detection model is specifically established by replacing the backbone network in the YOLOv5 network with a ResNet50 network pre-trained based on the computer vision standard ImageNet dataset; when training the target detection model, the production line image training set is input into the target detection model for training, and the ResNet50 network in the target detection model is frozen through the transfer learning method. Only the neck and detection head parts in the target detection model are trained, and the model loss is calculated and back propagation is performed until the target detection model converges and the training is completed and the weights of the target detection model are saved, and finally a trained solder joint defect target detection model is obtained.
[0047] The ResNet50 network pre-trained based on the ImageNet dataset has the ability to extract surface features of general objects. By using the pre-trained ResNet50 network to replace the backbone network of the YOLOv5 model, feature information of different levels of flat wire winding solder joints can be extracted without retraining the backbone network. The features expressed at these levels, from low to high, range from details to semantics.
[0048] 3) The single solder joint dataset is input into the solder joint defect target detection model. After feature extraction, the solder joint defect target detection model obtains several normal solder joint features and constructs a feature library.
[0049] In step 3), the single solder joint dataset is input into the solder joint defect target detection model. For each single solder joint annotated image with normal morphology in the single solder joint dataset, the pre-trained ResNet50 network in the solder joint defect target detection model is used to extract features of the single solder joint annotated image. The single solder joint annotated image is sequentially downsampled in the pre-trained ResNet50 network to obtain feature maps of four different levels. The feature maps of the second and third levels are fused and then subjected to specific aggregation and flattening processing, and then saved in the preset original library in sequence. Then, it is sampled and constructed into a feature library The feature library contains the feature information of normal solder joints in each single solder joint annotated image, such as Figure 3 shown.
[0050] In specific implementation, a single solder joint image with a resolution of 224×224 is passed through the ResNet50 network to generate feature maps of four different levels with 4x, 8x, 16x, and 32x downsampling. The intermediate-level feature maps, i.e., 28×28 and 14×14, are used. Specific aggregation processing is performed on the feature maps of the intermediate levels, and the higher-level feature maps are spliced together with the lower-level feature maps along the depth direction after bilinear interpolation. The features are flattened, stored, and sampled to construct a feature library.
[0051] The specific aggregation processing is as follows:
[0052]
[0053]
[0054] Where f represents the aggregation function; represents a square area with the position (a, b) as the center and p as the side length; φ ij (m,n) represents the feature map generated by the i-th single solder joint annotation image with normal morphology at the j-th level in the single solder joint dataset in the square area Features at position (m,n); λ(m,n) represents the square area The weighting coefficient of the feature at the inner position (m,n); Indicates the square area The distance from the inner position (m,n) to the center (a,b).
[0055] use Represents the training dataset of normal solder joint images, n is the number of images. Image x i The backbone network is extracted into feature maps of different levels, using φ ij Represents image x i At level j (j∈{1,2,3,4}), the feature map It is a three-dimensional tensor with depth c, height h, and width w, using φ ij (a,b) represents the feature slice of the feature map at position (a,b), where a∈(1,2,...,h) and b∈(1,2,...w).
[0056] In order to increase the receptive field without reducing the resolution, the adjacent features of each pixel on the feature map are aggregated and the Represents a rectangular block area with (a, b) as the center and a side length of p. The feature representation of this area is as follows:
[0057]
[0058] Where (m,n) is the area The specific location in the region; f is the aggregation function used to aggregate the region The feature information of all locations in the .
[0059] 4) Collect the production line solder joint images to be inspected, input the production line solder joint images to be inspected into the solder joint defect target detection model, the solder joint defect target detection model outputs the solder joint positions on the production line solder joint images to be inspected after processing, crop the production line solder joint images to be inspected into several single solder joint images according to the solder joint positions, input each single solder joint image into the solder joint defect target detection model, and the solder joint defect target detection model outputs the solder joint features; for the solder joint features of each single solder joint image, use the neighboring KNN algorithm to obtain the outlier value after processing based on the solder joint features and feature library, compare the outlier value with the preset threshold value, judge the defect situation of the solder joint in the single solder joint image and locate the defective area, so as to realize the solder joint defect detection of flat wire winding.
[0060] During defect detection, the solder joint defect target detection model is first used to detect the target position, and then the backbone network of the model is reused to extract the detailed features of the position for comparison to see if there are any abnormal defects.
[0061] In step 4), each single solder joint image is input into the solder joint defect target detection model, and the pre-trained ResNet50 network in the solder joint defect target detection model is used to extract features of the single solder joint image. The feature maps of the second and third levels of the pre-trained ResNet50 network are fused to obtain the solder joint features of the single solder joint image.
[0062] In step 4), for the solder joint features of each single solder joint image, according to the solder joint features and feature library, the neighboring KNN algorithm is used to search the feature library for the neighboring features of the solder joint features of the single solder joint image to be detected and obtain the abnormal value, and the abnormal value is compared with the preset threshold. When the abnormal value is less than or equal to the preset threshold, the solder joint in the single solder joint image is normal. When the abnormal value is greater than the preset threshold, the solder joint in the single solder joint image is abnormal, that is, the solder joint is defective, such as Figure 4 As shown, the defect area is located at this time, and the solder joint defect detection of the flat wire winding is realized.
[0063] Using the KNN algorithm, we search the feature library for the k nearest neighbor features of each position feature of the weld to be inspected using the Euclidean distance as the metric, and use the average value as the outlier, as follows:
[0064]
[0065] Among them, s is an outlier; m test For each feature of the solder joint to be inspected, is the feature library with m test The set of the nearest k features, m is the set Features in .
[0066] The image data of this invention was acquired from images captured by a camera on the automatic welding production line for flat wire windings. A total of 111 original images were collected, with approximately 20 weld points in each image. Figure 2 The input image is shown in the figure. The location of each solder joint is marked on the original image, and images of individual solder joints are cropped based on the marked locations, then labeled as normal / defective. The original production line dataset is divided into a 4 / 1 partition, with 89 images used as the training set and 22 images as the test set. The detection results are evaluated using average precision (AP). The detection results for the solder joints to be inspected on the production line are shown in the figure. Figure 5 As shown in the figure, the precision-recall rate of the target detection model for solder joints can reach 0.995. According to the above method, the backbone network Resnet50 is used to store the features of normal solder joints to form the final feature library. After the target detection detects the position of the solder joint on the production line image, the backbone network is used to compare the features extracted at that position with the features in the feature library to determine whether the solder joint has defects and locate the defect location. The defect detection results are evaluated with image-level AUROC and pixel-level AUROC, and the true positive rate TPR (True positive rate) and false positive rate FPR (False positive rate) curves are plotted. The results are shown in the figure. Figure 6 (a) and Figure 6 As shown in (b), the AUC can reach 97.4% and 98.0%, respectively. Some example results are as follows: Figure 7 As shown, the test image is the solder joint on the original production line image detected by the target detection algorithm, the real defect mask is the real defect position manually marked, the defect heat map is visualized according to the size of each outlier, the prediction mask is the defect position determined based on the outlier exceeding the threshold, and the defect segmentation is the defect position prediction based on the prediction mask. It can be seen that the method of the present invention can well detect solder joint defects.
[0067] The present invention addresses the problem that defect samples of flat wire winding welds in electric vehicle motors are scarce in actual industrial production, and that target detection algorithms cannot be used to directly predict defect categories. Based on feature comparison, the present invention provides a method for detecting weld defects that does not rely on defect data. The method first uses a target detection model to detect the positions of all winding welds on the production line image. It then extracts and fuses the intermediate layer features of each weld by reusing the backbone network of the model, and compares them with the features in the feature library composed of normal winding weld features, thereby determining whether the welds in the test image are defective and locating the defect positions. Actual measurement results show that the method provided by the present invention can effectively detect weld defects without relying on defect samples, thus solving the problem of scarce defect samples in industrial production.
Claims
1. A flat wire winding weld defect detection method that does not rely on defect data, It is characterized in that include: 1) Collect several solder joint images on the flat wire winding welding production line and mark the solder joint positions to obtain several solder joint annotated images and construct a production line image training set; crop each solder joint annotated image in the production line image training set to obtain several single solder joint annotated images, annotate each single solder joint annotated image with defects, extract each single solder joint annotated image with normal morphology, and construct a single solder joint data set; 2) Establish a target detection model, input the production line image training set into the target detection model for training, and obtain a trained solder joint defect target detection model; 3) Inputting the single solder joint data set into the solder joint defect target detection model, the solder joint defect target detection model performs feature extraction, obtains several normal solder joint features and constructs a feature library; 4) Collect the production line solder joint image to be detected, input the production line solder joint image to be detected into the solder joint defect target detection model, the solder joint defect target detection model outputs the solder joint position on the production line solder joint image to be detected after processing, and cuts the production line solder joint image to be detected into several single solder joint images according to the solder joint position, inputs each single solder joint image into the solder joint defect target detection model, and the solder joint defect target detection model outputs the solder joint feature; For the solder joint features of each single solder joint image, the neighboring KNN algorithm is used to obtain the outlier value according to the solder joint features and feature library. The outlier value is compared with the preset threshold to determine the defect of the solder joint in the single solder joint image and locate the defect area, thus realizing the solder joint defect detection of flat wire windings. In the step 2), the target detection model is specifically established by replacing the backbone network in the YOLOv5 network with a ResNet50 network pre-trained based on the computer vision standard ImageNet data set; when training the target detection model, the production line image training set is input into the target detection model for training, and the ResNet50 network in the target detection model is frozen by the transfer learning method, and only the neck and detection head parts in the target detection model are trained, and the model loss is calculated and back propagated until the target detection model converges and the training is completed and the weight of the target detection model is saved, and finally a trained solder joint defect target detection model is obtained.
2. The flat wire winding weld defect detection method independent of defect data according to claim 1, Features: In the step 1), the solder joint image is annotated with the solder joint position, and the position of each solder joint in the solder joint image is annotated; the solder joint annotated image is cropped, and the obtained single solder joint annotated image only contains one solder joint and has the same resolution; the single solder joint annotated image is annotated with defects, and if the solder joint in the single solder joint annotated image has defects, it is annotated and classified as a single solder joint with abnormal morphology; if the solder joint in the single solder joint annotated image has no defects, it is annotated and classified as a single solder joint with normal morphology, that is, the current single solder joint annotated image is a single solder joint annotated image with normal morphology.
3. The flat wire winding weld defect detection method independent of defect data according to claim 1, Features: In the step 3), the single solder joint data set is input into the solder joint defect target detection model, and for each single solder joint annotated image with normal morphology in the single solder joint data set, the pre-trained ResNet50 network in the solder joint defect target detection model is used to extract features of the single solder joint annotated image. The single solder joint annotated image is sequentially down-sampled in the pre-trained ResNet50 network to obtain feature maps of four different levels, and the feature maps of the second and third levels are fused, aggregated and flattened in turn, and then saved in a preset original library in turn, and then sampled and constructed into a feature library, which contains the feature information of the normal solder joints in each single solder joint annotated image.
4. The flat wire winding weld defect detection method independent of defect data according to claim 3, Features: The polymerization process is specifically as follows: Where f represents the aggregation function; represents a square area with the position (a, b) as the center and p as the side length; φ ij (m,n) represents the feature map generated by the i-th single solder joint annotated image with normal morphology at the j-level in the single solder joint dataset in the square area Features at position (m,n); λ(m,n) represents the square area The weighting coefficient of the feature at the inner position (m,n); Represents the square area The distance from the inner position (m,n) to the center (a,b).
5. The flat wire winding weld defect detection method independent of defect data according to claim 1, Features: In the step 4), each single solder joint image is input into the solder joint defect target detection model, and the single solder joint image is feature extracted using the pre-trained ResNet50 network in the solder joint defect target detection model. The solder joint features are obtained by fusing the feature maps of the second and third levels of the single solder joint image in the pre-trained ResNet50 network.
6. The flat wire winding weld defect detection method independent of defect data according to claim 1, Features: In the step 4), for the solder joint features of each single solder joint image, according to the solder joint features and feature library, the neighboring features of the solder joint features of the single solder joint image to be detected are searched in the feature library using the neighboring KNN algorithm and the abnormal value is obtained, and the abnormal value is compared with the preset threshold. When the abnormal value is less than or equal to the preset threshold, the solder joint in the single solder joint image is normal. When the abnormal value is greater than the preset threshold, the solder joint in the single solder joint image is abnormal, that is, the solder joint is defective. At this time, the defective area is located to realize the solder joint defect detection of the flat wire winding.
