Disordered snap spring identification strategy and method based on improved YOLOv11 model

Abstract

The invention discloses a disordered snap spring identification strategy and method based on an improved YOLOv11 model. The method comprises the following steps: sequentially establishing a data set of target snap springs; a YOLOv11 deep learning network is improved; training the target detection network by using a snap spring data set to obtain a target identification network; matching a clamp spring and a small dismounting hole thereof by utilizing an association strategy; an industrial camera is used for obtaining a clamp spring image from a placing platform, the clamp spring image is input into a target recognition network, feature association is conducted, and then clamp spring grabbing information is obtained. A C3k2 module is replaced by a C3k2DEConv module, the capacity of capturing key details on the edge of a small target is enhanced, a parallel patch awareness attention (PPA) module is used for replacing traditional convolution operation, the model can better pay attention to important information of the small target, and an NWD loss function is used for replacing an original loss function; according to the method, the model is ensured to accurately identify the snap spring and the disassembly small hole thereof, and an association strategy of the snap spring and the disassembly small hole thereof is provided, so that the identified snap spring is matched with the corresponding disassembly small hole, and prior preparation is made for subsequently clamping the disordered snap spring by using a mechanical arm.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine vision auxiliary disordered grabbing, and particularly relates to a disordered spring identification strategy and method based on an improved YOLOv11 model. TECHNICAL BACKGROUND

[0002] Industrial robot grabbing is the most basic and difficult task in industrial automation manufacturing, and its purpose is to replace human beings to repeatedly complete tedious tasks such as industrial part feeding, assembling and sorting. In a typical robot grabbing task, the robot needs to identify the required object in a cluttered scene containing multiple objects, and pick it up and place it in a specific position. Accurate pose estimation is considered one of the key technologies to solve this problem.

[0003] At present, the demand for automatic processing of C-shaped clamps is increasingly urgent in the automobile manufacturing and mechanical assembly industries. The current target detection system based on deep learning has made significant progress in general object recognition, but still faces severe technical challenges in application scenarios such as this specific workpiece.

[0004] Firstly, the detection accuracy of small targets is insufficient, and the small hole of the clamp disassembly is only tens of pixels in the industrial vision system. Traditional convolutional neural networks are difficult to achieve stable detection due to the limitation of receptive field and loss of detailed features. The existing YOLO series algorithm has limited ability to extract micro features, resulting in low recall rate of detection, which cannot meet the industrial precision requirements. Secondly, the correlation relationship of multiple targets is complex, and in the disordered placement scene, accurately determining the attribution relationship between the small hole and the clamp body becomes a technical bottleneck. Traditional methods rely on complex post-processing algorithms, but lack of targeted modeling of the geometric characteristics of the clamp, resulting in an association accuracy that cannot meet the industrial application standard. Real-time and precision are difficult to balance, and the processing cycle has strict requirements in industrial field (usually ≤300ms), while methods based on point cloud registration or complex feature matching have huge computational overhead, making it difficult to meet the real-time requirements while ensuring accuracy.

[0005] In recent years, convolutional neural networks (CNN) have shown excellent performance in target detection, and most research on target recognition has focused on deep learning technology. The latest method uses different strategies to introduce deep learning on color images for initial pose estimation, but the calculation is relatively expensive, and cannot achieve end-to-end optimization.

[0006] The existing technical solutions, such as a point pair feature (PPF) based matching method and a multi-modal perception based grabbing system, solve the unordered grabbing problem to a certain extent, but the accuracy and efficiency are still not ideal in grabbing the specific workpiece such as the clamp spring. For example, the traditional pose estimation depends on feature matching between an image and a corresponding model, and usually adopts a manually coded template or a feature optimized by a proxy target to perform corresponding grouping and hypothesis verification, and such a method has poor real-time performance and is difficult to meet the industrial beat requirement. SUMMARY

[0007] Therefore, the present application is based on the YOLOv11 deep learning network, and a module is introduced to improve the network in order to improve the precision and speed of the detection of the small hole small target of the clamp spring and make the network complete the feature association task of the clamp spring and the small hole, and an association strategy and algorithm of the clamp spring and the small hole are proposed, which can accurately and efficiently identify the clamp spring and the small hole, match the small hole to the clamp spring, and determine the pose of the clamp spring.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] An unordered clamp spring identification strategy and method based on an improved YOLOv11 model, the method comprising:

[0010] 1) Constructing a data set:

[0011] An industrial camera is used to obtain images of unordered and non-stacked clamp springs placed on a platform, and a Labelimg tool is used to preprocess the images and label the clamp springs and their small holes, and the labeling results are saved in YOLO label format;

[0012] The labeled clamp spring images are expanded after enhancement to form a data set;

[0013] 2) Improving the target detection network:

[0014] YOLOv11 including a backbone network, a neck network, a detection head and a loss calculation module is selected as a basic identification network;

[0015] A DEConv module is introduced in C3k2, and a PPA module is added to replace the traditional convolution operation based on the detection head Detect module;

[0016] In the target position loss function, an NWD loss function is used to replace the original loss function;

[0017] The target detection network is obtained;

[0018] 3) The labeled clamp spring enhanced images in 1) are used to form a data set to train the target detection network in 2), and the best weights of the clamp spring and small hole identification models are obtained respectively;

[0019] 4) A clasp and its associated strategy and algorithm for disassembling the small hole are proposed. The associated strategy is a special grabbing strategy designed for the characteristics of C-shaped clasp to determine the ownership relationship between the disassembling small hole and the clasp.

[0020] 5) An industrial camera is used to obtain images of unordered and non-stacked clamps placed on the platform. The clamp image is input into the target recognition model for target recognition.

[0021] Further, the DEConv module is introduced into the C3k2 module of the target detection network, and the DEConv module adopts a two-stage structure of training period and inference period.

[0022] Further, the training period extracts features in parallel, with 5 parallel convolutions, including 1 ordinary convolution (VC) and 4 differential convolutions (DC); the 1 ordinary convolution extracts intensity level basic features, and the 4 differential convolutions are central difference (CDC), angle difference (ADC), horizontal difference (HDC), and vertical difference (VDC), respectively, capturing gradient level high-frequency details in different dimensions. The 5 features are fused by element addition to ensure feature richness.

[0023] Further, the inference period performs reparameterization conversion, and uses the linearity of convolution to add the weights of the 5 parallel convolutions in the training period at the corresponding positions, which is equivalent to converting 1 ordinary 3x3 convolution, eliminating the additional overhead of parallel structure.

[0024] Further, the PPA module is composed of multi-branch fusion and attention mechanism.

[0025] The multi-branch fusion includes patch-aware and series convolution, and the parameters p in the patch-aware are set to 2 and 4, respectively, representing local and global branches; for the input feature tensor F, first adjust F' through pointwise convolution (PWConv), then calculate F_local, F_global and F_conv through three branches respectively, and finally sum the three results;

[0026] The attention mechanism includes channel attention and spatial attention components, and the summed results pass through one-dimensional channel attention map and two-dimensional spatial attention map in turn, through element-level multiplication and subsequent activation function, batch normalization and other operations, and finally obtain the output F" of PPA.

[0027] Further, the NWD loss function solves the problem that IoU_Loss is very sensitive to the position deviation of small objects. The NWD loss function first models the bounding box as a two-dimensional Gaussian distribution, and for two two-dimensional Gaussian distributions and whose second-order Wasserstein distance is defined as:

[0028] ;

[0029] Simplify as:

[0030] ;

[0031] where: denotes the Gaussian distribution, and is a two-dimensional mean vector, and is a covariance matrix, and is the square root of the covariance matrix, is the trace operation of the matrix, is the norm, is the Frobenius norm;

[0032] For the Gaussian distribution and real bounding box modeled by the predicted box and , it can be further simplified as:

[0033] ;

[0034] where: is the coordinate and coordinate of the predicted box center point, is the width and height of the predicted box; is the coordinate and coordinate of the real bounding box center point, is the width and height of the real bounding box; is the transpose of the vector, is the distance metric;

[0035] The output distance metric cannot be directly used as a similarity metric like IoU whose value is between 0 and 1. Normalize it in exponential form to get a new metric:

[0036] ;

[0037] where C is a constant closely related to the data set, which is empirically set to the average value size of AI-TOD in the experiment;

[0038] The NWD loss function is designed as:

[0039] ;

[0040] wherein is a Gaussian distribution model of the predicted box P, is a Gaussian distribution model of the ground truth bounding box G.

[0041] The NWD loss function can provide gradients in and or G.

[0042] Further, the spring and the associated strategy of the dismounting hole comprise the following steps:

[0043] S1, reading the label file to parse all objects, including the spring (spring is 0) and the hole (hole is 1);

[0044] S2, traversing each spring to find all holes within the bounding box of the spring;

[0045] S3, if the number of holes in the spring is equal to 2, directly associate;

[0046] S4, if the number of holes in the spring is greater than 2, select the best two holes;

[0047] S5, if the number of holes in the spring is less than 2, mark as association failed, skip this spring;

[0048] S6, calculate the grasping information (center point, angle, width);

[0049] S7, output the visual association result;

[0050] Further, the step of traversing each spring in the S2 step comprises:

[0051] S21, obtaining the label file of the current spring;

[0052] S22, inverse normalization: converting the normalized coordinates and width into pixel coordinates;

[0053] S23, calculating the bounding box (top-left corner, bottom-right corner) of each object;

[0054] S24, for the spring, recording its bounding box and center point;

[0055] S25, for the hole, recording its center point;

[0056] Further, the method of selecting the best two holes in the S4 step, if the number of holes in the spring is greater than 2, comprises:

[0057] S41, calculating the distance from the hole to the center of the spring, and screening out the holes whose values are within half the width of the spring;

[0058] S42, if the number of screened small holes is 2, directly select the best small hole pair;

[0059] S43, if the number of screened small holes is greater than 2, further calculate the center distance between each group of small hole pairs, and screen out small hole pairs with a certain distance as the best small hole pair.

[0060] The beneficial effects of the present application are:

[0061] 1. The DEConv module is used to optimize the target detection model of YOLOv11, which solves the problem that the traditional ordinary convolution in the YOLOv11 target detection network lacks relevant experience knowledge of target details, especially the insufficient capture ability of key details such as the small target edge of the small hole of the clamp spring disassembly and the target and background boundary, while avoiding the introduction of additional parameter quantity and calculation overhead, to adapt to the dual requirements of YOLOv11 for real-time inference performance and detection accuracy, and to enhance the precision and speed of the recognition network.

[0062] 2. The PPA module (parallel patch perception attention module) is innovatively used to reconstruct the detection head of YOLOv11, which adopts a multi-branch feature extraction strategy to extract features of different scales and levels through different branches; after multi-branch feature extraction, efficient channel attention and spatial attention components are used for adaptive feature enhancement. The key information of small targets is highlighted, which improves the accuracy of small hole detection of clamp spring disassembly; combined with the NWD loss function, the instability problem of IoU loss in small target detection is effectively alleviated.

[0063] 3. The association strategy and algorithm of clamp spring-small hole based on spatial relationship reasoning are innovatively proposed, which realizes an association accuracy of 95% through multi-level geometric constraints and an adaptive threshold mechanism. The algorithm fully excavates the relevant experience knowledge of the clamp spring structure, combines distance distribution analysis and direction consistency verification, effectively solves the attribution judgment problem of small holes in a multi-target scene, and determines the optimal posture for subsequent robot grasping to realize the automatic grasping of the clamp spring in the industrial field. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0065] Figure 1 The step flow chart of the disordered clamp spring recognition method in the embodiments of the present application;

[0066] Figure 2 A schematic diagram of the DEConv module structure in the embodiment of the present application is shown in the figure.

[0067] Figure 3 A schematic diagram of the PPA module structure in the embodiment of the present application is shown in the figure.

[0068] Figure 4 A schematic diagram of the improved YOLOv11 structure in the embodiment of the present application is shown in the figure.

[0069] Figure 5 A flowchart of the out-of-order spring identification strategy in the embodiment of the present application is shown in the figure.

[0070] Figure 6 A flowchart of the screening of the best small hole pair in the embodiment of the present application is shown in the figure.

[0071] Figure 7 A sample image of the industrial camera collected spring to be detected in the embodiment of the present application is shown in the figure.

[0072] Figure 8 A sample image of the spring marked after detection by the improved algorithm in the embodiment of the present application is shown in the figure.

[0073] Figure 9 A sample image of the spring associated after the association algorithm in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0074] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0075] As shown in the figure, the present embodiment discloses an out-of-order spring identification strategy and method based on an improved YOLOv11 model, including the following steps: Figure 1 1) Constructing a data set:

[0076] An industrial camera is used to obtain images of the out-of-order and non-stacked springs placed on the platform, and a Labelimg tool is used to preprocess the images and label the springs and their disassembly small holes. The labeling results are saved in YOLO label format;

[0077] The labeled spring images are expanded after enhancement to form a data set;

[0078] In the present embodiment, the expanded spring data set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1 for training the improved algorithm model.

[0079] In the present embodiment, the expanded spring data set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1 for training the improved algorithm model.

[0080] 2) Improve the target detection network:

[0081] YOLOv11, which includes a backbone network, a neck network, a detection head, and a loss calculation module, was selected as the basic recognition network.

[0082] In C3k2, the DEConv module is introduced, and the PPA module is added on the basis of the Detect module to replace the traditional convolution operation;

[0083] In the target location loss function, the NWD loss function is used instead of the original loss function;

[0084] Obtain the target detection network;

[0085] 3) The target detection network in 2) is trained using the annotated snap ring enhanced images from 1) to form the dataset, and the optimal weights for identifying snap rings and disassembling small holes are obtained respectively.

[0086] 4) A correlation strategy and algorithm for the circlip and its disassembly hole are proposed. The correlation strategy is a special grasping strategy designed for the characteristics of C-type circlips to determine the relationship between the disassembly hole and the circlip.

[0087] 5) Use an industrial camera to acquire images of disordered, unstacked snap rings placed on the platform, and input the snap ring images into the target recognition model for target recognition.

[0088] In this embodiment, the DEConv module is introduced into the C3k2 module of the object detection network. DEConv adopts a two-stage structure of training and inference:

[0089] During training, features are extracted in parallel using five parallel convolutions: one regular convolution (VC) and four differential convolutions (DC). The regular convolution extracts intensity-level basic features, while the four differential convolutions are central difference (CDC), angular difference (ADC), horizontal difference (HDC), and vertical difference (VDC), capturing high-frequency details at the gradient level in different dimensions. The five features are fused element-wise to ensure feature richness.

[0090] During inference, a reparameterization transformation is performed. Taking advantage of the linear additivity of convolution, the weights of the five parallel convolutions during training are added at the corresponding positions, which is equivalent to a single ordinary 3×3 convolution, eliminating the additional overhead of the parallel structure.

[0091] In this embodiment, the PPA module employs a multi-branch feature extraction strategy, extracting features at different scales and levels through different branches. It processes the input feature tensor using local, global, and sequential convolutional branches. The distinction between local and global branches is achieved by controlling the `patchsize` parameter, and the attention matrix between non-overlapping patches is calculated to realize local and global feature extraction and interaction. PPA consists of two parts: multi-branch fusion and an attention mechanism.

[0092] 1) Multi-branch fusion includes patch-aware and concatenated convolution. In patch-aware, the parameter p is set to 2 and 4, representing local and global branches respectively. For the input feature tensor F, F' is first adjusted by point convolution (PWConv), and then F_local, F_global and F_conv are calculated by the three branches respectively. Finally, the three results are summed.

[0093] 2) The attention mechanism includes channel attention and spatial attention components. The summation result passes through a one-dimensional channel attention map and a two-dimensional spatial attention map in sequence. Through element-wise multiplication and subsequent activation functions, batch normalization and other operations, the output F” of PPA is finally obtained.

[0094] In this embodiment, the NWD loss function addresses the issue that IoU_Loss is highly sensitive to small object positional deviations. The NWD loss function first models the bounding box as a two-dimensional Gaussian distribution. For two two-dimensional Gaussian distributions... and Its second-order Wasserstein distance is defined as:

[0095] ;

[0096] Simplified to:

[0097] ;

[0098] in: Indicates a Gaussian distribution. and It is a two-dimensional mean vector. and Let covariance matrix be the variance matrix. and The square root of the covariance matrix. For the trace operation of a matrix, for Norm, It is the Frobenius norm;

[0099] For from the prediction box and the true bounding box Modeling Gaussian distribution and This can be further simplified to:

[0100] ;

[0101] in: It is the center point of the prediction box coordinates and coordinate, These are the width and height of the prediction box; It is the center point of the true bounding box coordinates and coordinate, These are the width and height of the actual bounding box; This is the transpose of the vector. It's a distance metric, but the output distance metric cannot be directly used as a similarity metric (i.e., its value is between 0 and 1 like IoU). Normalizing it to its exponential form yields a new metric:

[0102] ;

[0103] Where C is a constant closely related to the dataset, which is empirically set to the average size of AI-TOD in the experiment;

[0104] The NWD loss function is designed as follows:

[0105] ;

[0106] in It is a Gaussian distribution model for the prediction box P. It is a Gaussian distribution model of the true bounding box G.

[0107] NWD loss function in and Gradients can be provided in both cases of G and G.

[0108] The association strategy of a retaining ring and its disassembly hole in this embodiment includes the following steps:

[0109] S1. Read the label file after the improved target network recognition and parse out all objects, including springs (spring is 0) and holes (hole is 1).

[0110] S2. Traverse each snap ring, sort by area, and prioritize processing large snap rings; find all the holes within its bounding box, record the holes in a list, and make different decisions based on the number of holes within the bounding box.

[0111] S3. If the number of small holes inside the snap ring is equal to 2, directly associate the holes and mark them as used.

[0112] S4. If the number of small holes inside the snap ring is greater than 2, select the two best small holes and mark the small holes as used.

[0113] S5. If the number of small holes inside the snap ring is less than 2, it is marked as an association failure, indicating that the snap ring is located at the edge of the image or is occluded, and the snap ring is skipped.

[0114] S6. Calculate the gripping information of the successfully associated snap ring, including the center position of the small hole, the center distance, the center of the gripping point, and the gripping angle;

[0115] S7. Output visualized association results;

[0116] In this embodiment, step S2 involves iterating through each retaining ring:

[0117] S21. Obtain the label file of the current snap ring, which includes the recognized snap ring image (.jpg) and label (.txt);

[0118] S22, Denormalization: Converts normalized coordinates and width to pixel coordinates.

[0119] x_pixel = x_center * image_width,

[0120] y_pixel = y_center * image_height,

[0121] w_pixel = width * image_width,

[0122] h_pixel = height * image_height;

[0123] S23. Calculate the bounding box (top left corner, bottom right corner) for each object:

[0124] x1 = x_pixel - w_pixel / 2,

[0125] y1 = y_pixel - h_pixel / 2,

[0126] x2 = x_pixel + w_pixel / 2,

[0127] y2 = y_pixel + h_pixel / 2;

[0128] S24. For the snap ring, record its bounding box and center point (x_pixel, y_pixel).

[0129] S25. For the small hole, record its center point;

[0130] In this embodiment, in step S4, if the number of small holes inside the snap ring is greater than 2, the method for selecting the two optimal small holes includes:

[0131] S41. Calculate the distance from the small hole to the center of the retaining ring, and select the small holes within a certain range of half the width of the retaining ring. This range is adjusted according to the size of the retaining ring.

[0132] S42. If the number of holes selected is 2, then directly select it as the best hole pair;

[0133] S43. If the number of selected holes is greater than 2, the center distance between each pair of holes is further calculated, and the hole pairs with a certain distance are selected as the best hole pairs. The setting of the certain value is affected by the size of the snap ring.

[0134] To verify the effectiveness of the method of the present invention, based on the above embodiments, a comparison is made between the improved YOLOv11 model of the present invention and the original YOLOv11 model:

[0135] Images of randomly arranged C-type snap rings of different sizes were acquired using Daheng Mercury series industrial cameras, forming an initial dataset of 600 images with multiple combinations. To further enrich the dataset and expand the image background information, various image enhancement techniques were introduced based on the initial dataset, including random flipping, cropping, image scaling, tone alteration, and noise addition, expanding the dataset to 1200 images. The expanded snap ring dataset was divided into a training set, validation set, and test set in an 8:1:1 ratio. The dataset was trained for 80 epochs on both the YOLOv11 model and the improved YOLOv11 model. The experimental results are shown in Table 1.

[0136] Table 1 Comparison of detection algorithms before and after improvement

[0137]

[0138] mAP: short for Mean Average Precision, refers to the average precision across all classes. It is a commonly used evaluation metric in the field of object detection to measure the performance of object detection algorithms. mAP50 is the average precision across all classes when the IoU threshold of the detection box is 50%, while mAP50-95 is a comprehensive metric under multiple IoU thresholds, which can better reflect the model's ability to accurately locate objects.

[0139] As shown in Table 1, the improved YOLOv11 model of this invention achieves higher recognition accuracy. The mAP50 increases from 89.5% to 95.2%, an absolute improvement of 5.7 percentage points, reaching the industrial-grade high-precision standard; the mAP50-95 increases from 84.6% to 87.9%, an absolute improvement of 3.3 percentage points, indicating that the model's performance is also improved under a stricter IoU threshold; the inference speed increases by nearly 9 percentage points, ensuring that the real-time requirements of most industrial application scenarios are met.

[0140] This invention proposes a ring snap-hole association strategy algorithm based on multi-level spatial constraints. Through a triple verification mechanism of distance distribution, direction consistency and regional distribution, it achieves an association accuracy of 95%. For unordered ring snaps without stacking, it completely solves the problem of hole ownership determination in multi-target scenarios.

[0141] The above description constitutes an embodiment of the present invention. The foregoing descriptions are preferred embodiments of the present invention. Unless there is a clear contradiction or a prerequisite for a particular preferred embodiment, the preferred embodiments can be arbitrarily combined and used. The embodiments and specific parameters described are merely for clarifying the verification process of the invention and are not intended to limit the scope of patent protection of the present invention. The scope of patent protection of the present invention is still determined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention should be included within the scope of protection of the present invention.

Claims

1. A strategy and method for identifying disordered snap rings based on an improved YOLOv11 model, characterized in that, The method includes: 1) Constructing the dataset: An industrial camera is used to acquire images of disordered, unstacked snap rings placed on a platform. The images are preprocessed using the Labelimg tool, and the snap rings and their disassembly holes are labeled. The labeling results are saved in YOLO label format. The labeled snap ring images are augmented to form a dataset; 2) Improve the target detection network: YOLOv11, which includes a backbone network, a neck network, a detection head, and a loss calculation module, was selected as the basic recognition network. In C3k2, the DEConv module is introduced, and the PPA module is added on the basis of the Detect module to replace the traditional convolution operation; In the target location loss function, the NWD loss function is used instead of the original loss function; Obtain the target detection network; 3) The target detection network in 2) is trained using the annotated snap ring enhanced images from 1) to form the dataset, and the optimal weights for identifying snap rings and disassembling small holes are obtained respectively. 4) A correlation strategy and algorithm for the circlip and its disassembly hole are proposed. The correlation strategy is a special grasping strategy designed for the characteristics of C-type circlips to determine the relationship between the disassembly hole and the circlip. 5) Use an industrial camera to acquire images of disordered, unstacked snap rings placed on the platform, and input the snap ring images into the target recognition model for target recognition.

2. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 1, characterized in that, The target detection network introduces a DEConv module into its C3k2 module. The DEConv module adopts a two-stage structure consisting of a training phase and an inference phase. During the training phase, features are extracted in parallel using five parallel convolutions, including one ordinary convolution and four differential convolutions. The ordinary convolution extracts intensity-level basic features, while the four differential convolutions are central difference, angular difference, horizontal difference, and vertical difference, respectively capturing high-frequency details at the gradient level in different dimensions. The five features are fused by element-wise addition to ensure feature richness. The inference phase undergoes a reparameterization transformation. Utilizing the linear additivity of convolutions, the weights of the five parallel convolutions from the training phase are added at corresponding positions, equivalent to a single ordinary 3×3 convolution, eliminating the additional overhead of the parallel structure.

3. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 1, characterized in that, The PPA module consists of two parts: multi-branch fusion and attention mechanism. The multi-branch fusion includes patch-aware and concatenated convolution, where the parameter p in patch-aware is set to 2 and 4, representing local and global branches, respectively. For the input feature tensor F, F' is first adjusted by pointwise convolution (PWConv), then F_local, F_global and F_conv are calculated by three branches respectively, and finally the three results are summed. The attention mechanism includes channel attention and spatial attention components. The summation result of the multi-branch fusion output passes sequentially through the one-dimensional channel attention map and the two-dimensional spatial attention map. Through element-wise multiplication and subsequent activation functions and batch normalization operations, the output F" of PPA is finally obtained.

4. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 1, characterized in that, The NWD loss function first models the bounding box as a two-dimensional Gaussian distribution. For two two-dimensional Gaussian distributions... and Its second-order Wasserstein distance is defined as: ； Simplified to: ； in: Indicates a Gaussian distribution. and It is a two-dimensional mean vector. and Let covariance matrix be the variance matrix. and The square root of the covariance matrix. For the trace operation of a matrix, for Norm, It is the Frobenius norm; For from the prediction box and the true bounding box Modeling Gaussian distribution and This can be further simplified to: ； in: It is the center point of the prediction box coordinates and coordinate, These are the width and height of the prediction box; It is the center point of the true bounding box coordinates and coordinate, These are the width and height of the actual bounding box; This is the transpose of the vector. It is a distance metric; The output distance metric cannot be directly used as a similarity measure with values ​​between 0 and 1 like IoU. Normalizing it to its exponential form yields a new metric: ； Where C is a constant closely related to the dataset, which is empirically set to the average size of AI-TOD in the experiment; The NWD loss function is designed as follows: ； in It is a Gaussian distribution model for the prediction box P. It is a Gaussian distribution model of the true bounding box G; NWD loss function in and Gradients can be provided in both cases of G and G.

5. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 1, characterized in that, The associated strategy of the snap ring and its disassembly hole includes the following steps: S1. Read the tag file and parse out all objects, including springs (spring is 0) and holes (hole is 1). S2. Traverse each snap ring and find all the holes within its bounding box; S3. If the number of small holes inside the snap ring is equal to 2, it is directly related; S4. If the number of small holes inside the snap ring is greater than 2, select the two best small holes; S5. If the number of small holes inside the snap ring is less than 2, mark it as an association failure and skip the snap ring. S6. Calculate the grabbing information, including the center point, angle, and width; S7. Output the visualized association results.

6. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 5, characterized in that, The step of iterating through each snap ring in step S2: S21. Obtain the label file for the current snap ring; S22, Denormalization: Converts normalized coordinates and width to pixel coordinates; S23. Calculate the bounding box for each object; S24. For the snap ring, record its bounding box and center point; S25. For the small hole, record its center point.

7. The disordered snap ring recognition strategy and method based on the improved YOLOv11 model according to claim 5, characterized in that, In step S4, if the number of small holes inside the snap ring is greater than 2, the method for selecting the two optimal small holes includes: S41. Calculate the distance from the small hole to the center of the snap ring, and select the small holes whose values ​​are within a small range of half the width of the snap ring. S42. If the number of holes selected is 2, then directly select it as the best hole pair; S43. If the number of selected holes is greater than 2, then further calculate the center distance between each pair of holes and select the pairs of holes with a certain distance as the best pairs of holes.

Images