Wheelchair part welding defect online detection system and method

Abstract

The invention provides a wheelchair part welding defect online detection system and method. The wheelchair part welding defect online detection system comprises an image acquisition mechanism, a first linear motion mechanism, a second linear motion mechanism, a connecting frame, a part objective table, a third linear motion mechanism and a control unit, the control unit controls the first linear motion mechanism and the second linear motion mechanism to move the connecting frame to enable the industrial area-array camera sets on the two sides above the part objective table to move left and right, and controls the third linear motion mechanism to enable the part objective table to move front and back. The industrial area-array camera group is used for acquiring an image of a to-be-detected part on the part objective table and transmitting the image to the control unit; and the control unit performs defect detection on the acquired image by adopting a YOLOV8 improved algorithm through a defect detection model. The device has the advantages of non-contact measurement, high detection speed, high detection precision and capability of stably working for a long time.

Description

Technical Field

[0001] The present invention belongs to the technical field of wheelchair component welding defect detection, and in particular relates to an online wheelchair component welding defect detection system and method. Background Art

[0002] Wheelchairs are important tools for rehabilitation. Ordinary wheelchairs are generally composed of wheelchair frames, wheels, seats, brakes and other parts. Once a wheelchair fails, it is easy to cause catastrophic accidents that harm people's lives, and the consequences are very serious. Therefore, welding defects of wheelchair components are of great significance to the safe use of wheelchairs.

[0003] Wheelchairs are mainly composed of tubular parts welded and assembled. Tubular parts welding is to weld the circular materials with two overlapping ports into one piece by melting arc welding. Therefore, the welding points of the tubular parts are outside the tubular parts, and the welding points are distributed on the outer ring surface around the diameter of the tubular parts. The welding defects of tubular welded parts will cause the inner cavity to be exposed, which makes it easy to rust in the cavity. Weld defects will cause a significant reduction in structural stability and a corresponding significant decrease in bearing capacity. The length and depth of weld defects have an impact on the local stability of thin-walled tubular structures. When the defects reach a certain level, it will cause a qualitative change in structural stability.

[0004] Defect detection on the surface of wheelchairs is mainly carried out manually, which has the disadvantages of low detection efficiency, low accuracy of detection results and strong subjectivity. Research on wheelchair safety detection technology is not comprehensive enough. Summary of the invention

[0005] In view of the above technical problems, the present invention provides an online detection system for welding defects of wheelchair components, which has the advantages of non-contact measurement, fast detection speed, high detection accuracy, and long-term stable operation.

[0006] The present invention also provides a detection method of the online detection system for welding defects of wheelchair components.

[0007] Note that the description of these objectives does not prevent the existence of other objectives. One embodiment of the present invention does not need to achieve all of the above objectives. Objectives other than the above objectives can be extracted from the description of the specification, drawings, and claims.

[0008] The present invention achieves the above technical objectives through the following technical means.

[0009] An online detection system for welding defects of wheelchair components, comprising an image acquisition mechanism, a first linear motion mechanism, a second linear motion mechanism, a connecting frame, a parts loading platform, a third linear motion mechanism and a control unit;

[0010] The image acquisition mechanism includes a plurality of industrial area array camera groups, and a plurality of industrial area array camera groups are respectively arranged above and below the part stage. The industrial area array camera groups in the middle of the upper and lower parts stage are respectively connected to the external bracket, and the positions remain unchanged. Connecting frames are respectively arranged on both sides of the part stage. The industrial area array camera groups on both sides above the part stage are respectively arranged on the upper part of the connecting frame, and the industrial area array camera groups on both sides below the part stage are respectively arranged on the lower part of the connecting frame, and a linear guide rail is provided at the bottom of the connecting frame;

[0011] The first linear motion mechanism is connected to the industrial area array camera group on one side above the part stage, and the second linear motion mechanism is connected to the industrial area array camera group on the other side above the part stage, driving the connecting frames on both sides of the part stage to move left and right along the bottom linear guide rail to the horizontal line of the tubular part where the welding point is located, so as to realize the shooting of the welding point;

[0012] The third linear motion mechanism is arranged on both sides of the bottom of the part stage to drive the part stage to move forward and backward;

[0013] The control unit is respectively connected to the industrial area array camera group, the first linear motion mechanism, the second linear motion mechanism and the third linear motion mechanism; the control unit controls the first linear motion mechanism and the second linear motion mechanism to move the connecting frame so that the industrial area array camera groups on both sides above the part stage move left and right, and controls the third linear motion mechanism to move the part stage forward and backward. The industrial area array camera group is used to collect images of the parts to be tested on the part stage and transmit them to the control unit; the control unit uses the YOLOV8 improved algorithm through the defect detection model to perform defect detection on the collected images.

[0014] In the above scheme, the industrial area array camera group includes a fixing plate, a light source fixing module, a camera fixing module, an industrial area array camera and an industrial area array light source;

[0015] Industrial area array cameras are respectively provided at the front and rear ends of the fixing plate, and the industrial area array cameras are connected to the fixing plate through a camera fixing module;

[0016] The industrial area array light source is mounted on a fixing plate via a light source fixing module.

[0017] In the above scheme, the first linear motion mechanism and the second linear motion mechanism have the same structure, and each comprises a first driving motor, a first slider and a first guide rail;

[0018] The output shaft of the first driving motor is connected to the first sliding block, and drives the first sliding block to perform linear motion on the first guide rail.

[0019] In the above solution, the third linear motion mechanism includes a second drive motor, a second slider, a second guide rail and a straight rod;

[0020] The second guide rails are respectively arranged on both sides of the part carrier, and the two second guide rails are connected by straight rods. Second sliders are respectively provided on the second guide rails, and the two sides of the part carrier are respectively connected to the second sliders. The output shaft of the second drive motor is connected to one of the second sliders to drive the second slider to move linearly along the second guide rails.

[0021] In the above scheme, the parts loading platform is a hollow horizontal bar structure including a main frame and loading spokes;

[0022] A plurality of parallel-arranged cargo spokes are arranged in the main frame, and the cargo spokes are used for placing wheelchair components.

[0023] In the above scheme, the control unit includes a picture acquisition module, a data storage module, a defect detection module and a result display module;

[0024] The image acquisition module is used to control the movement of the second linear motion mechanism, drive the part stage to move, the wheelchair parts on the part stage pass under the industrial array camera group, control the industrial array light source of the industrial array camera group to turn on, control the industrial array cameras above and below the part stage to photograph the welding defects of the wheelchair parts, and transmit them to the data storage module;

[0025] The data storage module is used to save the collected pictures into a folder to be detected;

[0026] The defect detection module is used to perform defect detection on the collected image by adopting the defect detection model of the YOLOV8 improved algorithm;

[0027] The result display module is used to display the detection results and mark the defect type and defect number.

[0028] A control method according to the online detection system for welding defects of wheelchair components comprises the following steps:

[0029] Step S1, loading: placing the wheelchair component to be tested on the component loading platform and fitting it with the loading spokes;

[0030] Step S2, position adjustment: the control unit controls the first linear motion mechanism and the second linear motion mechanism to move the corresponding industrial area array camera group to the horizontal line where the tubular defect of the wheelchair component is located;

[0031] Step S3, adjusting the industrial area array camera group: adjusting the focal length and light source intensity of the industrial area array camera group;

[0032] Step S4, image acquisition: the control unit controls the second linear motion mechanism to move, driving the part stage to move, the wheelchair component on the part stage passes under the industrial area array camera group, the control unit controls the industrial area array light source of the industrial area array camera group to turn on, controls the industrial area array cameras above and below the part stage to photograph the welding defects of the wheelchair component, and transmits the images to the control unit;

[0033] Step S5, image defect detection: the control unit uses the YOLOV8 improved algorithm through the defect detection model to perform defect detection on the collected image;

[0034] Step S6, result display: the control unit displays the detection results on the display interface, marking the defect type and defect number.

[0035] In the above solution, the image defect detection in step S5 specifically includes the following steps:

[0036] S5.1. Photograph the welding defect sample set of wheelchair parts and perform preprocessing:

[0037] Welding defect types include weld holes, weld seams, and weld slag flakes. Images of the welding parts of wheelchair components were taken as sample sets. The data set was expanded using data augmentation, and the entire data set was randomly divided into a training set and a validation set.

[0038] All defect images in the training set and validation set are labeled with boxes, including the specific location and category information of the three defects;

[0039] Convert the annotated XML file format to YOLO data format. The generated txt file includes the image file name, the corresponding annotation file, the category label of each annotated area, and the bounding box coordinate information. Generate the training set train.txt and the validation set val.txt file lists;

[0040] S5.2. Improve the YOLOV8n model:

[0041] In the YOLOV8n network, the backbone network and the neck network use the C2fGhostv2 feature extraction module to replace the C2f module, and the C2fGhostv2 module uses the Ghostv2Bottleneck to replace the Bottleneck in the C2f module;

[0042] Add the SEAttention module to the YOLOV8n model, input the captured image into the YOLOV8n model, and generate the intrinsic feature map through the convolution of C2fGhostv2. Then, each channel of the intrinsic feature map generates the Ghost feature map through linear transformation. Finally, the intrinsic feature map and the Ghost feature map are concatenated into the final feature map. The feature map X is compressed to compress the original feature map size C×H×W to C×1×1. The compression processing formula is as follows:

[0043]

[0044] Among them, Z c is the compressed feature information of channel c, c is the number of channels, F sq is the compression function, U C is the result after convolution, H is the height of the image, W is the width of the image, u c (i, j) is the element at position (i, j) in the cth channel of the original feature map;

[0045] By using the information aggregated in the squeeze operation to fully capture the channel dependency, the attention weight S is obtained as follows:

[0046] S=F ex (z,W)=σ(W2δ(W1z))

[0047] Among them, F ex is the activation function, σ is the sigmoid function, δ is the ReLU function, W1 is the first fully connected layer parameter, W2 is the second fully connected layer parameter, the attention weight S is weighted to each channel feature, and finally the feature map with channel attention is output;

[0048] The importance of the channel is modeled through the SE attention mechanism, and the SE attention mechanism is integrated into the backbone network to output the feature map The formula is as follows:

[0049]

[0050] Among them, F s％ale For the scaling operation, s c is the weight of the cth channel, u c is the feature of the cth channel in the original feature map;

[0051] A small object detection head is added to the YOLOV8n model, with a detection feature map resolution of 160X160, which is used to detect targets with a pixel size of 4X4 ​​or larger.

[0052] In the YOLOV8n network, Wise_IoU is used to replace CIoU as the loss function. The Wise_IoU formula is as follows:

[0053] L WIoU =r×R WIoU ×L IoU

[0054]

[0055] L IoU =1-IoU

[0056]

[0057] Among them, L IoU ∈(0,1) represents a medium quality anchor box, R WIoU ∈(0,e) represents a high-quality anchor box, r is a non-monotonic focusing coefficient, β represents the outlier degree, For variables The constant to convert, is the running average of the momentum m, tn is the number of training iterations; α, β are learning parameters. The learning parameter α is used to control the weight of the classification loss, and the learning parameter β is used to control the outlier degree in the loss function. x, y are the coordinates of the center point of the prediction box. gt ,y gt is the coordinate of the center point of the real frame, W g and H g The width and height of the bounding rectangle of the predicted box and the true box;

[0058] The improved YOLOV8n model is trained using the training set, and the model is dynamically adjusted according to the verification results of the verification set until the expected surface defect recognition accuracy is achieved, thereby obtaining the trained YOLOV8n model;

[0059] S5.3. Detect welding defects based on the trained YOLOV8n model: Input the images collected by the industrial area array camera group into the improved YOLOV8n model, and use the improved YOLOV8 algorithm to perform defect detection on the collected images to obtain the defect type and number of defects.

[0060] In the above scheme, the data enhancement method in the preprocessing of step S5.1 includes rotating the image and adding blur, and the data set is expanded by rotating the sample set photos by 180° and adding blur noise.

[0061] The above scheme also includes a second detection step S7.

[0062] If a serious defect is detected in the first detection, it is necessary to re-capture the image for re-inspection and confirmation:

[0063] S7.1. Remove the defects detected in the first round, and perform a second round of detection on the images that have no defects detected in the first round;

[0064] S7.2, a second round of detection is performed on the positions of the defects detected in the first round, and the system will return to the image acquisition step, re-acquire the image and perform defect detection analysis on the re-taken image.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The control unit of the present invention controls the first linear motion mechanism and the second linear motion mechanism to move the connecting frame so that the industrial area array camera groups on both sides above the part stage move left and right to the horizontal line of the tubular part where the welding point is located, and controls the third linear motion mechanism to move the part stage forward and backward. The industrial area array camera group is used to collect images of the parts to be tested on the part stage and transmit them to the control unit; the control unit performs defect detection on the collected images by using the defect detection model of the improved YOLOV8 algorithm, and can realize defect detection of welding parts of various models of wheelchairs in the enterprise production line; the present invention can realize high-precision detection of various welding defects. The present invention shortens the welding defect detection time of a single wheelchair component; the present invention has the advantages of non-contact measurement, fast detection speed, high detection accuracy, and long-term stable operation.

[0067] Note that the description of these effects does not prevent the existence of other effects. One embodiment of the present invention does not necessarily have all of the above effects. Effects other than the above can be clearly seen and extracted from the description of the specification, drawings, claims, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic structural diagram of an online detection system for welding defects of wheelchair components according to an embodiment of the present invention;

[0069] Figure 2 It is a partial structural schematic diagram of an online detection system for welding defects of wheelchair components according to an embodiment of the present invention;

[0070] Figure 3 It is a schematic structural diagram of an industrial area array camera group according to an embodiment of the present invention;

[0071] Figure 4 is a schematic diagram of training results according to an embodiment of the present invention, wherein Figure 4 (a) is the bounding box loss value during the training process of this model, Figure 4 (b) is the classification loss value during the training process of this model, Figure 4 (c) is the distributed focal loss value during the training process of this model, Figure 4 (d) is the accuracy of the model in bounding box prediction, Figure 4 (e) is the recall rate of this model on bounding box prediction, Figure 4 (f) is the bounding box loss value during the model verification process. Figure 4 (g) is the classification loss value during the model validation process. Figure 4 (h) is the distributed focus loss value in the verification process of this model, Figure 4 (i) is the mAP50 in the validation process of this model, Figure 4 (j) is the mAP50-95 in the validation process of this model;

[0072] Figure 5 A schematic diagram of a software interface according to an embodiment of the present invention;

[0073] Figure 6 A schematic diagram of welding defects of wheelchair components according to an embodiment of the present invention;

[0074] Figure 7 Schematic diagram of welding defect detection results of wheelchair components according to one embodiment of the present invention Figure 1 ;

[0075] Figure 8 Schematic diagram of welding defect detection results of wheelchair components according to one embodiment of the present invention Figure 2 ;

[0076] Fig. 9 Schematic diagram of welding defect detection results of wheelchair components according to one embodiment of the present invention Figure 3 . DETAILED DESCRIPTION

[0077] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0078] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "front", "back", "left", "right", "up", "down", "axial", "radial", "vertical", "horizontal", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0079] Figure 1 and 2 The figure shows a preferred embodiment of the wheelchair component welding defect online detection system, which includes an image acquisition mechanism, a first linear motion mechanism 1, a second linear motion mechanism 3, a connecting frame 8, a parts carrier 5, a third linear motion mechanism 2 and a control unit.

[0080] The image acquisition mechanism includes multiple industrial area array camera groups 4. Multiple industrial area array camera groups 4 are respectively arranged above and below the part stage 5. The industrial area array camera groups 4 in the middle above and below the part stage 5 are respectively connected to external brackets and their positions remain unchanged. Connecting frames 8 are respectively provided on both sides of the part stage 5. The industrial area array camera groups 4 on both sides above the part stage 5 are respectively arranged on the upper part of the connecting frame 8, and the industrial area array camera groups 4 on both sides below the part stage 5 are respectively arranged on the lower part of the connecting frame 8. A linear guide rail is provided at the bottom of the connecting frame 8.

[0081] The first linear motion mechanism 1 is connected to the industrial area array camera group 4 on one side above the part stage 5, and the second linear motion mechanism 3 is connected to the industrial area array camera group 4 on the other side above the part stage 5, driving the connecting frames 8 on both sides of the part stage 5 to move left and right along the bottom linear guide rail, so that the industrial area array camera group 4 moves to the horizontal line of the tubular part where the welding point is located.

[0082] The third linear motion mechanism 2 is arranged on both sides of the bottom of the part stage 5 to drive the part stage 5 to move forward and backward.

[0083] The control unit is respectively connected to the industrial area array camera group 4, the first linear motion mechanism 1, the second linear motion mechanism 3 and the third linear motion mechanism 2; the control unit controls the first linear motion mechanism 1 and the second linear motion mechanism 3 to move the connecting frame 8 so that the industrial area array camera group 4 on both sides above the part stage 5 moves left and right, and controls the third linear motion mechanism 2 to move the part stage 5 forward and backward, which is conducive to the defect detection of welding points of wheelchair components of different sizes and models. The industrial area array camera group 4 is used to collect images of the parts to be tested on the part stage 5 and transmit them to the control unit; the control unit performs defect detection on the collected images by adopting the defect detection model of the YOLOV8 improved algorithm.

[0084] like Figure 3 As shown, the industrial area array camera group 4 includes a fixing plate 401, a light source fixing module 402, a camera fixing module 403, an industrial area array camera 404 and an industrial area array light source 405; the front and rear ends of the fixing plate 401 are respectively provided with industrial area array cameras 404, and the industrial area array camera 404 is connected to the fixing plate 401 through the camera fixing module 403; the industrial area array light source 405 is installed on the fixing plate 401 through the light source fixing module 402.

[0085] The first linear motion mechanism 1 and the second linear motion mechanism 3 have the same structure, and each comprises a first drive motor, a first slider and a first guide rail; the output shaft of the first drive motor is connected to the first slider to drive the first slider to perform linear motion on the first guide rail.

[0086] The third linear motion mechanism 2 includes a second drive motor, a second slider, a second guide rail and a straight rod 201; the second guide rails are respectively arranged on both sides of the part stage 5, and the two second guide rails are connected by a straight rod 201. Second sliders are respectively arranged on the second guide rails, and the two sides of the part stage 5 are respectively connected to the second sliders. The output shaft of the second drive motor is connected to one of the second sliders to drive the second slider to move linearly along the second guide rails.

[0087] The parts loading platform 5 is a hollow horizontal bar structure including a main frame and loading spokes 6; a plurality of parallel arranged loading spokes 6 are provided in the main frame, and the loading spokes 6 are used to place wheelchair components 7. The loading spokes 6 can ensure that the welding parts of the wheelchair components will not be blocked, and can ensure that the upper and lower industrial area array cameras of the industrial area array camera group 4 can capture clear images, thereby meeting the requirements for welding defect detection of wheelchair components of different models and sizes.

[0088] The control unit includes an image acquisition module, a data storage module, a defect detection module and a result display module; the image acquisition module is used to control the second linear motion mechanism 2 to move at a uniform speed, drive the part stage 5 to move, and the wheelchair component 7 on the part stage 5 passes under the industrial area array camera group 4, control the industrial area array light source 405 of the industrial area array camera group 4 to turn on, highlight the surface defects of the welding parts of the wheelchair components, and control the industrial area array cameras 404 above and below the part stage 5 to shoot the welding defects of the wheelchair components 7 and pass them to the data storage module; the data storage module is used to save the collected pictures in a folder to be detected; the defect detection module is used to perform defect detection on the collected images through a defect detection model using the YOLOV8 improved algorithm; the result display module is used to display the detection results and mark the defect type and the number of defects.

[0089] The present invention illuminates defective parts of wheelchair components from multiple angles through an industrial area array light source 405, collects images of defective parts of wheelchair components from multiple angles through an industrial area array camera 404, and transmits and saves the real-time collected images to a control unit.

[0090] The present invention uses pictures of wheelchair component defects taken by the industrial array camera 404 as a sample set, preprocesses the sample set, expands the sample set to increase the data set, divides the training set and the test set, uses the YOLOV8 improved algorithm to train the pictures, and outputs the trained model weights to obtain a defect detection model, uses the trained defect detection model with the best weight to detect the sample set, removes the detected defects, performs a secondary detection on pictures without welding defects, re-photographs the defective parts and performs defect detection, and saves the experimental results of the secondary detection and the best weights in the data storage module; preferably, as Figure 5 and Figure 6 As shown, the present invention displays the storage data of the data storage module in real time, builds a software interface in combination with the image acquisition module, and outputs the experimental results to the host computer for display. The interface is functionally developed based on the host computer software to complete the three functions of wheelchair component defect marking, wheelchair component defect statistics and result visualization.

[0091] The welding points of the wheelchair component 7 are typical ring welds, and the welding points are distributed on the outer ring surface surrounding the diameter of the tubular parts. In order to capture a complete and clear picture of the welding points of the wheelchair component 7, the present invention adopts an industrial area array camera group 4 to photograph the tubular parts. Preferably, the four industrial area array light sources 405 in a single industrial area array camera group 4 illuminate the welding points of the tubular parts from multiple angles respectively to achieve all-round lighting and greatly reduce shadows. The four industrial area array cameras 404 in a single industrial area array camera group 4 photograph the welding points of the tubular parts from multiple angles respectively. The three industrial area array camera groups 4 are used to overlap the field of view of the industrial cameras to achieve full coverage shooting of the welding parts of wheelchair components of different models and sizes and perform defect detection.

[0092] In a specific embodiment of the present invention, each industrial area array camera group 4 includes four industrial area array cameras 404 and four industrial area array light sources 405 installed at the following positions: Figure 1 , 2 As shown, the industrial area array camera 404 is fixed on the fixed plate 401 through the industrial area array camera fixing module 403, and the industrial area array light source 405 is fixed on the fixed plate 401 through the industrial area array light source fixing module 402; preferably, the angle between the industrial area array camera fixing module 403 and the fixed plate 401 is 60°, the industrial area array light source fixing module 402 is vertically fixed on the fixed plate 401, the angle between the industrial area array light source 405 and the industrial area array light source fixing module 402 is 45°, the industrial area array camera group 4 connects the upper and lower fixed plates 401 by the connecting frame 8, the industrial area array camera 404 is 100 mm away from the part stage 5, the industrial area array light source 405 has a good lighting effect, the industrial area array camera 404 has clear imaging, and can achieve complete shooting of the welding part.

[0093] The industrial area array camera group 4 on both sides of the upper and lower parts stage 5 is connected by the industrial area array camera group connection module 8 to the upper and lower fixing plates 401, and realizes simultaneous synchronous long movement.

[0094] Preferably, the length of the industrial area array camera group connection module 8 is 390 mm.

[0095] A control method according to the online detection system for welding defects of wheelchair components comprises the following steps:

[0096] Step S1, loading: placing the wheelchair component 7 to be tested on the component loading platform 5 and fitting it with the loading spoke 6;

[0097] Step S2, position adjustment: the control unit controls the first linear motion mechanism 1 and the second linear motion mechanism 3 to move the corresponding industrial area array camera group 4 to the horizontal line where the tubular defect of the wheelchair component 7 is located;

[0098] Step S3, industrial area array camera group adjustment: adjusting the focal length of the industrial area array camera 404 and the light source intensity of the industrial area array light source 405 in the industrial area array camera group 4, so that the pictures taken by the industrial area array camera group 4 achieve a clear effect;

[0099] Step S4, image acquisition: the control unit controls the second linear motion mechanism 2 to move at a uniform speed, driving the part stage 5 to move, and the wheelchair component 7 on the part stage 5 passes under the industrial area array camera group 4. The control unit controls the industrial area array light source 405 of the industrial area array camera group 4 to turn on, highlighting the surface defects of the welding parts of the wheelchair components, and controls the industrial area array cameras 404 above and below the part stage 5 to photograph the welding defects of the wheelchair component 7 and transmit them to the control unit;

[0100] Step S5, image defect detection: the control unit uses the YOLOV8 improved algorithm through the defect detection model to perform defect detection on the collected image;

[0101] Step S6, result display: the control unit displays the detection results on the display interface, marking the defect type and defect number.

[0102] The image defect detection step S5 specifically includes the following steps:

[0103] S5.1. Photograph the welding defect sample set of wheelchair component 7 and perform preprocessing:

[0104] Welding defect types include weld holes (WH), weld seams (WL), and weld slag (WS). Pictures of 7 welding parts of wheelchair components were taken as sample sets. The sample set photos were rotated 180° and blurred noise was added to expand the data set to ensure that the data set includes various types of defects to improve the generalization ability of the model. The entire sample set was randomly divided into 80% training set and 20% validation set according to the ratio of 4:1. Dividing the data set into validation sets helps to effectively evaluate the model performance during the training process and prevent overfitting.

[0105] Use Labelimg open source image annotation tool to accurately annotate all defect images in the training set and validation set. The annotation content includes the specific location and category information of the three defects.

[0106] Convert the annotated XML file format to YOLO data format. The generated txt file includes the image file name, the corresponding annotation file, and the category label, bounding box coordinates and other information of each annotated area. Generate the training set train.txt and validation set val.txt file lists;

[0107] S5.2. Improve the YOLOV8n model:

[0108] In the YOLOV8n network, the lightweight and efficient C2fGhostv2 feature extraction module is used to replace the C2f module. Ghostv2Bottleneck is constructed by Ghost convolution combined with the decoupled fully connected (DFC) attention mechanism. It has the ability to dynamically calibrate and capture long-distance spatial information. Compared with Bottleneck, Ghostv2Bottleneck can use fewer parameters to obtain more feature information. The C2fGhostv2 module uses Ghostv2Bottleneck to replace Bottleneck in the C2f module, effectively reducing the redundant calculations brought by ordinary convolution in Bottleneck. Without weakening the feature extraction capability, it reduces the number of parameters, ensures lightweight, and improves the model detection speed;

[0109] The SEAttention module is added to the YOLOv8n model. The SE (Squeeze and Excitation) attention mechanism can dynamically learn the importance of each channel, suppress useless features according to the importance and increase the weight of useful features to improve small target detection. The captured picture is input into the YOLOv8n model and the convolution of C2fGhostv2 is used to generate an intrinsic feature map. Then each channel of the intrinsic feature map is linearly transformed to generate a Ghost feature map. Finally, the intrinsic feature map and the Ghost feature map are concatenated into the final feature map. The feature map X is compressed and the original feature map size C×H×W is compressed to C×1×1, thereby reducing the amount of calculation and the amount of parameters. The compression processing formula is as follows:

[0110]

[0111] Among them, Z c is the compressed feature information of channel c, c is the number of channels, F sq is the compression function, U C is the result after convolution, H is the height of the image, W is the width of the image, u c (i, j) is the element at position (i, j) in the cth channel of the original feature map;

[0112] By using the information aggregated in the squeeze operation to fully capture the channel dependency, the attention weight S is obtained as follows:

[0113] S=F ex (z,W)=σ(W2δ(W1z))

[0114] Among them, F exis the activation function, σ is the sigmoid function, δ is the ReLU function, W1 is the first fully connected layer parameter, the compressed feature information z of the channel is linearly transformed, W2 is the second fully connected layer parameter, the first fully connected layer is calculated with the nonlinear activation function δ, and then linearly transformed by W2, the attention weight S obtained above is weighted to each channel feature, and finally the feature map with channel attention is output;

[0115] The importance of channels is modeled through the SE attention mechanism to improve the performance of the model. After the SE attention mechanism is integrated into the backbone network (Backbone), since the main function of the Backbone is to extract the features of the input image, the SE attention mechanism can be integrated behind the Backbone to summarize the extracted features, suppress unimportant features, and transmit important feature information to the next stage to output feature maps. The formula is as follows:

[0116]

[0117] Among them, F s％ale For the scaling operation, s c is the weight of the cth channel, u c is the feature of the cth channel in the original feature map;

[0118] The YOLOV8n model has three detection heads by default for detecting targets of different sizes. The detection feature map size corresponding to P3 / 8 is 80X80, which is used to detect targets larger than 8X8. The detection feature map size corresponding to P4 / 16 is 40X40, which is used to detect targets larger than 16X16. The detection feature map size corresponding to P5 / 32 is 20X20, which is used to detect targets larger than 32X32. In order to improve the detection capability of small targets, the present invention adds a tiny object detection head to the YOLOV8n model. The detection feature map resolution size is 160X160, which is used to detect targets with a pixel size of 4X4 ​​or larger, so as to improve the detection capability of smaller wheelchair component welding defects and solve the problems of missed detection or poor detection effect due to the presence of small targets in the data set.

[0119] In order to predict the position of the object, YOLOV8n uses CIoU (Complete Intersection over Union) as the loss function, assumes that the target has good features, and continues to strengthen the fitting ability as the bounding box loss function. The YOLOV8n network of the present invention uses Wise_IoU to replace CIoU as the loss function, implements a dynamic non-monotonic mechanism, uses outliers instead of IoU to evaluate the quality of the anchor frame, and uses gradient gain allocation technology to reduce the competitive advantage of the high-quality anchor frame of the small target to reduce the negative gradient effect generated by the low-quality anchor frame of the small target, so that WIoU can focus on the anchor frame of the small target with ordinary quality, thereby improving the overall effect of the detector and having a better detection effect. The Wise_IoU formula is as follows:

[0120] L WIoU =r×R WIoU ×L IoU

[0121]

[0122] L IoU =1-IoU

[0123]

[0124] Among them, L IoU ∈(0,1) represents a medium-quality anchor box, which is used to reduce the quality of the anchor box, R WIoU ∈(0,e) represents a high-quality anchor frame, which is used to amplify the medium-quality anchor frame, r is the non-monotonic focusing coefficient, β represents the outlier degree, For variables The constant to convert, is the running average of the momentum m, tn is the number of training iterations; α, β are learning parameters. The learning parameter α is used to control the weight of the classification loss, and the learning parameter β is used to control the outlier degree in the loss function. x, y are the coordinates of the center point of the prediction box. gt ,y gt is the coordinate of the center point of the real frame, W g and H g The width and height of the bounding rectangle of the predicted box and the true box;

[0125] The anchor frame division standard after adding Wise_IoU is dynamic, which is conducive to making the best gain allocation strategy overall;

[0126] The improved YOLOV8n model is trained using the training set, and the model is dynamically adjusted according to the verification results of the verification set until the expected surface defect recognition accuracy is achieved, thereby obtaining the trained YOLOV8n model;

[0127] S5.3. Detect welding defects based on the trained YOLOV8n model: Input the images collected by the industrial area array camera group 4 into the improved YOLOV8n model, and use the improved YOLOV8 algorithm to perform defect detection on the collected images to obtain the defect type and number of defects.

[0128] The control method of the wheelchair component welding defect online detection system further includes step S7 secondary detection. According to the requirements, a second round of detection is performed. If a minor defect is found in the first detection and is considered acceptable, the product will pass the inspection directly; on the contrary, if a serious defect is detected, it is necessary to re-capture the image for re-inspection and confirmation:

[0129] S7.1. Remove the defects detected in the first round, and perform a second round of detection on the images that have no defects detected in the first round;

[0130] S7.2, a second round of detection is performed on the positions of the defects detected in the first round, and the system will return to the image acquisition step, re-acquire the image and perform defect detection analysis on the re-taken image.

[0131] If the product is finally judged to be qualified, the qualified results and related information will be stored in the data storage module, and the stored data will be used to generate a visual report for the next production step.

[0132] In a specific embodiment of the present invention, the largest wheelchair component is selected as the detection object, and the control method of the wheelchair component welding defect online detection system comprises the following steps:

[0133] S1. Loading: Place the wheelchair component 7 to be tested on the parts loading platform 5 and fit it with the loading spokes 6;

[0134] S2. Position adjustment: The control unit controls the first linear motion mechanism 1 and the second linear motion mechanism 3 so that the corresponding industrial area array camera group 4 moves to the horizontal line where the tubular defect of the wheelchair component 7 is located;

[0135] S3. Adjustment of the industrial array camera group: Adjust the focal length of the industrial array camera 404 and the light source intensity of the industrial array light source 405 in the industrial array camera group 4, so that the pictures taken by the industrial array camera group 4 can be clear;

[0136] S4. Image acquisition: The control unit controls the second linear motion mechanism 2 to move, driving the part stage 5 to move, and the wheelchair component 7 on the part stage 5 passes under the industrial array camera group 4. The control unit controls the industrial array light source 405 of the industrial array camera group 4 to turn on, and controls the industrial array cameras 404 above and below the part stage 5 to capture welding defects of the wheelchair component 7 and transmit the images to the control unit;

[0137] Further, preferably, the industrial area array camera 404 adopts the Huagu Power industrial area array camera WP-GS630M. The triggering of the Huagu Power industrial area array camera WP-GS630M for taking pictures is determined by the distance of the second linear motion mechanism 2 passing through the welding part and the running speed of the second linear motion mechanism 2. The trigger pulse for triggering the industrial area array camera 404 to take pictures can be obtained by conversion, and the trigger pulse for the industrial area array camera 404 to take pictures is determined in the JSON file. The industrial area array camera 404 in the industrial area array camera group 4 is triggered to take pictures at the welding part of the wheelchair component 7. Since the loading spokes 6 for fixing the wheelchair component 7 are thin spokes, the part loading platform 5 has a hollow spoke structure. Therefore, the industrial area array camera 404 in the industrial area array camera group 4 can comprehensively capture the welding defects of the wheelchair component 7.

[0138] S6. Image defect detection: The control unit performs defect detection on the collected images by using the defect detection model of the YOLOV8 improved algorithm. The image acquisition module of the control unit uses the Huagu Power industrial array camera WP-GS630M to collect welding defect images of wheelchair parts, and transmits the real-time collected data to the data storage module. The data storage module receives and stores the image data of the data acquisition module, the best weight obtained by training, and the data detected by the detection module. The defect detection module performs defect detection on the data collected and stored in the data storage module, and detects surface defects of parts by using the YOLOV8 improved model. The result display module displays the stored data of the data storage module in real time and interacts with the defect detection module.

[0139] Preferably, code is written to process user input: C# code is written in the source file, and when the user triggers the detection, a pre-trained defect detection model is called to analyze the imported image and parse the output results.

[0140] Preferably, the display logic of the result display module is: based on the results of the defect detection algorithm, the detection results are displayed on the display interface, the defect type and the number of defects are marked, and the result display is updated in real time to ensure that the user can intuitively see the detection progress and final results.

[0141] Preferably, the interface in the result display module has image zooming and panning functions and a statistical table of defect types and numbers, which can help staff quickly and accurately identify product defects and make corresponding decisions. At the same time, these functions also facilitate the evaluation and debugging of system performance.

[0142] The secondary inspection process ensures that each product is accurately evaluated and recorded, while providing a flexible secondary inspection mechanism, improving the accuracy and reliability of the entire product quality control system, and displaying the process trends of defect types within the same type of wheelchair components, providing an intuitive basis for process improvement.

[0143] Through the above steps, a complete wheelchair component welding defect detection system is constructed, which has a user-friendly interface and reliable defect detection capabilities.

[0144] Optimized hardware resources: GPU: 3060TI; CPU: i513490f;

[0145] Software resources: operating system: Windows 11; Python: 3.9.0; CUDA: 11.3; torch: 1.12.1; cuDNN: 8.9.3;

[0146] Hyperparameter settings: Batchsize: 4; Epoch: 1000; lr0=0.01.

[0147] The training results are evaluated, and evaluation indicators such as model detection accuracy, detection speed, and model size are selected: Model detection accuracy refers to the ratio of the number of correctly classified positive samples to the number of positive samples determined by the classifier, detection speed FPS refers to the number of frames transmitted per second (FramesPerSecond), and model size refers to the size of the optimal weight model file.

[0148] The results show that the detection accuracy of the improved YOLOV8n model is 85.7%, the detection accuracy of weld hole defects is 99.4%, and the detection accuracy of weld defects is 93.1%. The model size is 6.4MB, the detection speed is 108 frames / second, and the time required to detect defects in a single wheelchair component is less than 30 seconds.

[0149] Figure 4 This is a diagram of training results. The horizontal axis represents the training rounds, and the vertical axis represents the result value. Figure 4 (a)- Figure 4 (c) is a schematic diagram of the results of this model during the training process, which are the bounding box loss, classification loss, and distributed focal loss during the training process. Figure 4 (d) and Figure 4 (e) is the precision and recall of this model. Figure 4 (f)- Figure 4 (h) is a schematic diagram of the results of the verification process, which are the bounding box loss, classification loss, and distributed focal loss in the verification set. Figure 4 (i) and Figure 4 (j) is the mAP50 and mAP50-95 of this model during the training process.

[0150] Figure 4 In (a), train / box_loss represents the change of bounding box loss during training. This loss value measures the error between the bounding box predicted by the model and the true labeled box. The bounding box loss gradually decreases as training progresses and tends to stabilize at 0.4, which means that this model performs well in locating objects.

[0151] Figure 4 (b) train / cls_loss represents the classification loss calculated during the training process. The classification loss gradually decreases to 0.02 as the training progresses, indicating that the model's prediction of the category is becoming more and more accurate;

[0152] Figure 4 (c) train / dfl_loss represents the distributed focal loss during the training process. The distributed focal loss gradually decreases to 0.5 as the training progresses, indicating that the performance of this model in predicting bounding boxes is better;

[0153] Figure 4 (f) val / box_loss represents the bounding box loss on the validation set. This loss value measures the error between the bounding box predicted by the model and the true bounding box. The bounding box loss value gradually decreases to 1.2 as the training progresses, indicating that this model performs well in locating the target.

[0154] Figure 4 (g)val / cls_loss represents the classification loss on the validation set. The classification loss value gradually decreases to 0.01 as the training progresses, indicating that the model is more accurate in classifying the target;

[0155] Figure 4 (h)val / dfl_loss represents the distributed focal loss during the validation process. The distributed focal loss gradually decreases to 0.5 as the training progresses, indicating that the performance of this model in predicting bounding boxes is better;

[0156] Figure 4 (d) metrics / precision (B) indicates the accuracy of the model in bounding box prediction, that is, the proportion of samples detected as positive that are actually positive. The accuracy of this model reached 0.85, indicating that this model has achieved good detection results;

[0157] Figure 4 (e) metrics / recall (B) represents the recall rate of the model in bounding box prediction, that is, the proportion of samples that are actually positive that are correctly detected as positive by the model. The recall rate of this model reaches 0.75, indicating that this model has achieved good detection results;

[0158] Figure 4 (i) metrics / mAP50 (B) represents the mean average precision (meanAveragePrecision) when the IoU (Intersection over Union) threshold is 0.5. The mAP50 of this model reaches 0.75, indicating that this model has achieved good detection results;

[0159] Figure 4 (j) metrics / mAP50-95 (B) represents the average of the average precision calculated in the range of IoU thresholds from 0.5 to 0.95. This indicator provides a performance evaluation of the model under different detection difficulties, which is a more comprehensive and rigorous measurement standard. The mAP50-95 of this model reached 0.48, indicating that this model achieved good detection results;

[0160] From the above, the improved YOLOV8n model proposed in the present invention has high accuracy in detecting welding defects of wheelchair components and meets the expected requirements.

[0161] Figure 7 , Figure 8 , Fig. 9 This is a schematic diagram of the identification effect of welding defects in wheelchair parts. Figure 7 The weld hole (WH) defect was accurately detected with a confidence level of 0.86, which indicates the degree of confidence of the model in the existence of the target object in the prediction box. The greater the confidence level, the more accurate the prediction result. Figure 8 Two weld (WL) defects were accurately detected with confidence levels of 0.69 and 0.88 respectively; Fig. 9 The weld hole (WH) defect and the weld slag (WS) defect were accurately detected with confidence levels of 0.77 and 0.78, respectively. Therefore, the model of the present invention can accurately detect three different types of defects.

[0162] In the backbone network and the neck network of the YOLOV8n algorithm, the present invention adopts the C2fGhostv2 feature extraction module to replace the C2f module, and the C2fGhostv2 module uses the Ghostv2Bottleneck to replace the Bottleneck in the C2f module, thereby reducing the number of parameters, ensuring lightweight, reducing the model complexity, and improving the model detection speed without weakening the feature extraction capability. The present invention adds an SE attention mechanism module, which strengthens the channel features of the input feature map, effectively fuses feature maps of different scales, and improves the feature expression capability without changing the size and number of channels of the input feature map. It can be flexibly applied to models of different scales, and the importance of the channel is modeled through the SE attention mechanism, thereby improving the performance of the model.

[0163] In view of the fact that weld holes, weld seams, and weld slag defects are small and irregularly distributed, the present invention adds a small target detection module to the head network of the YOLOV8n algorithm to improve the detection capability of smaller welding defects of wheelchair components and solve the problems of missed detection or poor detection effect due to the presence of small targets in the data set.

[0164] The present invention adopts Wise_IoU to replace CIoU as the loss function in the YOLOV8n network, and adopts an improved loss function, focusing on the anchor frame of ordinary quality of small targets, thereby improving the detection effect of the detector on small target defects.

[0165] The present invention can achieve a detection accuracy of more than 85% for three types of welding defects: weld holes, weld seams, and weld slag flakes. In the present invention, the welding defect detection time of a single wheelchair component is within 30 seconds. The present invention can achieve normal and continuous operation at -10°C to 40°C.

[0166] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0167] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. An online detection system for welding defects of wheelchair components, characterized in that: It comprises an image acquisition mechanism, a first linear motion mechanism (1), a second linear motion mechanism (3), a connecting frame (8), a parts carrier (5), a third linear motion mechanism (2) and a control unit; The image acquisition mechanism comprises a plurality of industrial array camera groups (4), wherein the plurality of industrial array camera groups (4) are respectively arranged above and below the part stage (5), the industrial array camera groups (4) in the middle above and below the part stage (5) are respectively connected to external brackets and their positions remain unchanged, a connecting frame (8) is respectively arranged on both sides of the part stage (5), the industrial array camera groups (4) on both sides above the part stage (5) are respectively arranged on the upper part of the connecting frame (8), and the industrial array camera groups (4) on both sides below the part stage (5) are respectively arranged on the lower part of the connecting frame (8), and a linear guide rail is provided at the bottom of the connecting frame (8); The first linear motion mechanism (1) is connected to an industrial array camera group (4) on one side above the part stage (5), and the second linear motion mechanism (3) is connected to an industrial array camera group (4) on the other side above the part stage (5), driving the connecting frames (8) on both sides of the part stage (5) to move left and right along the bottom linear guide rail; The third linear motion mechanism (2) is arranged on both sides of the bottom of the part carrier (5) to drive the part carrier (5) to move forward and backward; The control unit is respectively connected to the industrial area array camera group (4), the first linear motion mechanism (1), the second linear motion mechanism (3) and the third linear motion mechanism (2); the control unit controls the first linear motion mechanism (1) and the second linear motion mechanism (3) to move the connecting frame (8) so that the industrial area array camera group (4) on both sides above the part stage (5) moves left and right, and controls the third linear motion mechanism (2) to move the part stage (5) forward and backward. The industrial area array camera group (4) is used to collect images of the parts to be tested on the part stage (5) and transmit them to the control unit; the control unit uses the YOLOV8 improved algorithm through the defect detection model to perform defect detection on the collected images.

2. The wheelchair component welding defect online detection system according to claim 1 is characterized in that: The industrial area array camera group (4) comprises a fixing plate (401), a light source fixing module (402), a camera fixing module (403), an industrial area array camera (404) and an industrial area array light source (405); Industrial area array cameras (404) are respectively provided at the front and rear ends of the fixing plate (401), and the industrial area array cameras (404) are connected to the fixing plate (401) via a camera fixing module (403); The industrial area array light source (405) is mounted on a fixing plate (401) via a light source fixing module (402).

3. The wheelchair component welding defect online detection system according to claim 1, characterized in that: The first linear motion mechanism (1) and the second linear motion mechanism (3) have the same structure, and each comprises a first drive motor, a first slider and a first guide rail; The output shaft of the first driving motor is connected to the first sliding block, and drives the first sliding block to perform linear motion on the first guide rail.

4. The wheelchair component welding defect online detection system according to claim 1, characterized in that: The third linear motion mechanism (2) comprises a second driving motor, a second slider, a second guide rail and a straight rod (201); The second guide rails are respectively arranged on both sides of the part carrier (5), the two second guide rails are connected by a straight rod (201), and second sliders are respectively arranged on the second guide rails. The two sides of the part carrier (5) are respectively connected to the second sliders, and the output shaft of the second drive motor is connected to one of the second sliders to drive the second slider to move linearly along the second guide rails.

5. The wheelchair component welding defect online detection system according to claim 1, characterized in that: The parts loading platform (5) is a hollow horizontal bar structure including a main frame and loading spokes (6); A plurality of parallel-arranged cargo-carrying spokes (6) are provided in the main frame, and the cargo-carrying spokes (6) are used to place wheelchair components (7).

6. The wheelchair component welding defect online detection system according to claim 1, characterized in that: The control unit includes a picture acquisition module, a data storage module, a defect detection module and a result display module; The image acquisition module is used to control the movement of the second linear motion mechanism (2), thereby driving the part stage (5) to move, so that the wheelchair component (7) on the part stage (5) passes under the industrial array camera group (4), controls the industrial array light source (405) of the industrial array camera group (4) to turn on, and controls the industrial array cameras (404) above and below the part stage (5) to photograph welding defects of the wheelchair component (7) and transmit the images to the data storage module; The data storage module is used to save the collected pictures into a folder to be detected; The defect detection module is used to perform defect detection on the collected image by adopting the defect detection model of the YOLOV8 improved algorithm; The result display module is used to display the detection results and mark the defect type and defect number.

7. A control method for an online detection system for welding defects of wheelchair components according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step S1, loading: placing the wheelchair component (7) to be tested on the component loading platform (5) and fitting it with the loading spokes (6); Step S2, position adjustment: the control unit controls the first linear motion mechanism (1) and the second linear motion mechanism (3) so that the corresponding industrial area array camera group (4) moves to the horizontal line where the tubular defect of the wheelchair component (7) is located; Step S3, adjusting the industrial area array camera group: adjusting the focal length and light source intensity of the industrial area array camera group (4); Step S4, image acquisition: the control unit controls the second linear motion mechanism (2) to move, thereby driving the part stage (5) to move, and the wheelchair component (7) on the part stage (5) passes under the industrial array camera group (4). The control unit controls the industrial array light source (405) of the industrial array camera group (4) to turn on, and controls the industrial array cameras (404) above and below the part stage (5) to photograph the welding defects of the wheelchair component (7), and transmits the images to the control unit. Step S5, image defect detection: the control unit uses the YOLOV8 improved algorithm through the defect detection model to perform defect detection on the collected image; Step S6, result display: the control unit displays the detection results on the display interface, marking the defect type and defect number.

8. The control method of the wheelchair component welding defect online detection system according to claim 7, characterized in that: The image defect detection step S5 specifically includes the following steps: S5.

1. Photograph the welding defect sample set of wheelchair components (7) and perform preprocessing: Types of welding defects include weld holes, weld seams, and weld slag flakes. Images of the welding parts of wheelchair components (7) are taken as sample sets. The data set is expanded using data augmentation, and the entire data set is randomly divided into a training set and a validation set. All defect images in the training set and validation set are labeled with boxes, including the specific location and category information of the three defects; Convert the annotated XML file format to YOLO data format. The generated txt file includes the image file name, the corresponding annotation file, the category label of each annotated area, and the bounding box coordinate information. Generate the training set train.txt and the validation set val.txt file lists; S5.

2. Improve the YOLOV8n model: In the YOLOV8n network, the C2fGhostv2 feature extraction module is used to replace the C2f module, and the C2fGhostv2 module uses Ghostv2Bottleneck to replace the Bottleneck in the C2f module; Add the SEAttention module to the YOLOV8n model, input the captured image into the YOLOV8n model, and generate the intrinsic feature map through the convolution of C2fGhostv2. Then, each channel of the intrinsic feature map generates the Ghost feature map through linear transformation. Finally, the intrinsic feature map and the Ghost feature map are concatenated into the final feature map. The feature map X is compressed to compress the original feature map size C×H×W to C×1×1. The compression processing formula is as follows: Among them, Z c is the compressed feature information of channel c, c is the number of channels, F sq is the compression function, U C is the result after convolution, H is the height of the image, W is the width of the image, u c (i, j) is the element at position (i, j) in the cth channel of the original feature map; By using the information aggregated in the squeeze operation to fully capture the channel dependency, the attention weight S is obtained as follows: S=F ex (z, W)=σ(W2δ(W1z)) Among them, F ex is the activation function, σ is the sigmoid function, δ is the ReLU function, W1 is the first fully connected layer parameter, W2 is the second fully connected layer parameter, the attention weight S is weighted to each channel feature, and finally the feature map with channel attention is output; The importance of the channel is modeled through the SE attention mechanism, and the SE attention mechanism is integrated into the backbone network to output the feature map The formula is as follows: Among them, F scale For the scaling operation, s c is the weight of the cth channel, u c is the feature of the cth channel in the original feature map; A small object detection head is added to the YOLOV8n model, with a detection feature map resolution of 160X160, which is used to detect targets with a pixel size of 4X4 ​​or larger. In the YOLOV8n network, Wise_IoU is used to replace CIoU as the loss function. The Wise_IoU formula is as follows: L WIoU =r×R WIoU ×L IoU L IoU =1-IoU Among them, L IoU ∈(0,1) represents a medium quality anchor box, R WIoU ∈(0, e) represents a high-quality anchor box, r is a non-monotonic focusing coefficient, β represents the outlier degree, For variables The constant to convert, is the running average of the momentum m, tn is the number of training iterations; α, β are learning parameters. The learning parameter α is used to control the weight of the classification loss, and the learning parameter β is used to control the outlier degree in the loss function. x, y are the coordinates of the center point of the prediction box. gt ,y gt is the coordinate of the center point of the real frame, W g and H g The width and height of the bounding rectangle of the predicted box and the true box; The improved YOLOV8n model is trained using the training set, and the model is dynamically adjusted according to the verification results of the verification set until the expected surface defect recognition accuracy is achieved, thereby obtaining the trained YOLOV8n model; S5.

3. Detect welding defects based on the trained YOLOV8n model: Input the images collected by the industrial area array camera group (4) into the improved YOLOV8n model, and use the improved YOLOV8 algorithm to perform defect detection on the collected images to obtain the defect type and defect number.

9. The control method of the wheelchair component welding defect online detection system according to claim 8, characterized in that: The data enhancement method in the preprocessing of step S5.1 includes rotating the image and adding blur, and the data set is expanded by rotating the sample set photos by 180° and adding blur noise.

10. The control method of the wheelchair component welding defect online detection system according to claim 7, characterized in that: It also includes step S7 secondary detection, If a serious defect is detected in the first detection, it is necessary to re-capture the image for re-inspection and confirmation: S7.

1. Remove the defects detected in the first round, and perform a second round of detection on the images that have no defects detected in the first round; S7.2, a second round of detection is performed on the positions of the defects detected in the first round, and the system will return to the image acquisition step, re-acquire the image and perform defect detection analysis on the re-taken image.

Images