Bolt looseness detection method based on belleville spring gasket visual target spot and computer vision technology

Abstract

The invention provides a bolt looseness detection method based on a belleville spring gasket visual target spot and a computer vision technology. The bolt looseness detection method comprises the steps that the belleville spring gasket visual target spot used for detecting bolt looseness is designed; the method comprises the following steps: acquiring a bolt image with a visual target spot of a belleville spring gasket, and manually marking to form an original data set; building a virtual data derivative network model, pre-training, and then training on an original data set to complete transfer learning; generating a large number of virtual bolt images by using the virtual data derivative network model, performing manual annotation to form a virtual data set, and combining the virtual data set with an original data set to form a training set; building a BoltYOLO bolt looseness detection deep neural network and training the BoltYOLO bolt looseness detection deep neural network; and inputting a to-be-detected image into the BoltYOLO bolt looseness detection deep neural network to obtain a detection result. The method has the advantages of being high in detection speed, high in precision, low in cost and high in reliability, meanwhile, the requirement for the number of original data sets is remarkably reduced, and the actual engineering requirement is better met.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent structures and structural health monitoring, and particularly to the identification and detection of steel structure bolt loosening defects based on computer vision technology. Background Art

[0002] The designed pre-tightening force of bolts is one of the important measures to ensure that the structure does not undergo relaxation and slip under dynamic loads and maintain integrity and reliability. The relaxation of bolt pre-tightening force will cause a significant reduction in the fatigue life of the structure under cyclic loads, seriously threatening the structural safety. The rapid detection of bolt loosening is one of the current industry difficulties. Most bolt positions are hidden and remote, and wired and patrol inspection methods are extremely inconvenient. It is difficult to timely and accurately grasp the safety status of the bolt group of large structures, and it is also impossible to conduct systematic automatic monitoring and management of bolts. Due to the lack of effective monitoring of bolt loosening, in recent years, structural collapse accidents caused by bolt failures have occurred frequently, resulting in huge economic losses and casualties.

[0003] In actual engineering, manual inspection and sensor-based non-destructive testing are common methods for detecting bolt loosening. Among them, manual inspection is the most common, mainly relying on manual regular inspections. Inspectors judge the tightness state of bolts through methods such as hammering method, torque wrench method, scribing calibration method, and magnetic strip pasting method. Although manual inspection is simple, its efficiency is very low, mostly relying on the operation experience of inspectors, with high false detection and missed detection rates, and it is difficult for manpower to cover bolts in remote locations. Sensor-based detection methods are more accurate and can be divided into force-based, piezoelectric, and ultrasonic detection technologies according to the principle. Although these detection technologies can achieve rapid and accurate detection, they usually require specific equipment, the equipment layout is also relatively complex, the price is expensive, the working conditions are harsh, and it is easily interfered by external factors. Therefore, it is difficult to be widely used for steel structure bolts, and its practicality in actual engineering is not high. In recent years, with the development and progress of computers, related technologies of computer vision and deep learning, with the characteristics of automatic, efficient, and flexible, are suitable for bolt loosening detection based on vision.

[0004] Machine learning is a technology that enables a computer to automatically learn and improve from data without explicit programming, and is used for tasks such as prediction, classification, clustering, and optimization decision-making. As a branch of machine learning, deep learning uses multi-layer neural networks to learn complex patterns in data without manual feature engineering. Deep learning shows higher accuracy and efficiency when processing large-scale data such as images and texts, and is more suitable for bolt loosening detection methods based on vision. There are two key points for bolt loosening detection based on vision:

[0005] First, vision-based bolt loosening detection requires a visual target that can accurately reflect the bolt loosening state. This visual target should be able to be captured by a camera and presented in an image, and the computer should be able to rely on this visual target to judge the tightness state of the bolt. In existing research, most methods indirectly map bolt loosening by detecting the relative rotation angle between the nut and the bolt or the length change of the exposed part of the screw. These features are indeed related to bolt loosening in theory. However, in actual engineering, it is difficult to implement methods based on the changes in bolt features before and after loosening. Moreover, due to the complex real working conditions, problems such as small changes and appearance wear of the bolt may occur, rendering such detection methods ineffective. Therefore, the current bottleneck in vision-based bolt loosening detection lies in the lack of a unified, general, and stable visual feature for bolt loosening. The industry urgently needs an effective, direct, and computer-recognizable visual target for bolt loosening.

[0006] Next, vision-based bolt loosening detection requires a deep neural network model that can judge the bolt state based on the visual target. Currently, in the field of deep learning, many object detection models based on deep neural networks, such as R-CNN, YOLO, and SSD, have been proposed. These models all have powerful capabilities to identify and capture different features and can well distinguish objects with obvious appearance differences. Therefore, these models are widely used in object detection and security monitoring in daily life. However, in the task of bolt loosening detection, the positive and negative samples to be identified and distinguished are both bolts. Except for the visual target, the appearance of the positive and negative samples is basically the same. In this case, common object detection models may not be able to accurately capture and locate the visual target of bolt loosening, making it difficult to establish the connection between the visual target and the bolt state. Even worse, they may output seemingly correct detection results based on irrelevant information or other deep abstract features, leading to potential application risks. Therefore, it is necessary to develop a dedicated object detection model for bolt loosening detection. This model should be able to accurately find and focus on the visual target in the image and make decisions based on the visual target. Summary of the Invention

[0007] Based on the problems in the above-mentioned background technology, the present invention provides a bolt loosening detection method based on computer vision technology, which mainly includes: designing a new type of bolt loosening visual target based on a disc spring gasket, which can directly reflect the tightness state of the bolt and has the advantages of easy observation, easy perception, convenience and stability; building a dedicated deep neural network model BoltYOLO for bolt loosening detection based on the visual target of the disc spring gasket, which can accurately capture and locate the proposed visual target in the image and output the detection result of the bolt tightness state according to the visual target, and has the characteristics of high accuracy and fast speed; in addition, it also includes a virtual data derivation network model, a model interpretation and visualization module, etc., and finally forms an intelligent detection method for steel structure bolt loosening defects based on the visual target of the disc spring gasket and computer vision technology. Compared with the wired sensor device and the manual inspection method, the method of the present invention has the characteristics of fast detection speed, high accuracy, low cost and high reliability, and at the same time significantly reduces the quantity requirement for the original data set, which is more in line with the actual engineering requirements.

[0008] The technical solution of the present invention is as follows:

[0009] A bolt loosening detection method based on the visual target of the disc spring gasket and computer vision technology, including:

[0010] Step 1, design a disc spring gasket visual target for detecting bolt loosening;

[0011] Step 2, collect bolt images with the visual target of the disc spring gasket and manually annotate them to form an original data set;

[0012] Step 3, build a virtual data derivation network model based on the generative adversarial network, first perform pre-training, and then train on the original data set to complete transfer learning;

[0013] Step 4, use the virtual data derivation network model to generate a large number of virtual bolt images, manually annotate all the images to form a virtual data set, and then merge the virtual data set with the original data set to form a training set;

[0014] Step 5, build a BoltYOLO bolt loosening detection deep neural network and train the BoltYOLO bolt loosening detection deep neural network on the training set;

[0015] Step 6, input the image to be detected into the BoltYOLO bolt loosening detection deep neural network, and the BoltYOLO bolt loosening detection deep neural network infers and outputs a result image with a detection bounding box, a detection category and a confidence level.

[0016] The present invention has the following advantages and beneficial effects:

[0017] First, the visual target of the conical spring washer used in the present invention for detecting bolt loosening is a ring-shaped color band mark, which is extremely easy to observe and can be clearly recognized even in complex or poorly lit environments; the characteristics of the bolt with this mark during loosening are intuitive, facilitating the computer to quickly distinguish the bolt state; it is convenient to use, the conical spring washer can be directly installed without complex operations, special tools, or additional processes, and the bolt fastening state can be checked through this mark after installation; at the same time, the bolt with a conical spring washer covers the mark in the fastened state, and the integrity and readability of the mark can be maintained for a long time, and it is not easily affected by environmental factors.

[0018] Secondly, the BoltYOLO bolt loosening detection deep neural network designed for the visual target of the conical spring washer in the present invention adopts an enhanced backbone network with stronger feature extraction ability, and multiple convolutional attention mechanisms are added to the neck network, significantly enhancing the model's attention to important information related to the visual target of the conical spring washer in the input feature map. With the efficient processing ability of the BoltYOLO bolt loosening detection deep neural network, the present invention can complete the evaluation of the loosening state of a large number of bolts in a short time; through the optimization of the model structure and the application of model interpretation technology, the high accuracy and reliability of the detection results are ensured, effectively avoiding misjudgment or missed detection caused by human factors.

[0019] In addition, the present invention constructs an efficient virtual data derivation network model through a generative adversarial network, which can generate a large amount of virtual data based on a small amount of original data sets, greatly reducing the actual demand for the original data sets and being more convenient for engineering applications. Description of the Drawings

[0020] Figure 1 is a block diagram of the technical solution of the bolt loosening detection method based on the visual target of the conical spring washer and computer vision provided by the present invention;

[0021] Figure 2 is a schematic diagram of the visual target of the conical spring washer provided by the present invention;

[0022] Figure 3 is a plan view of the working mechanism of the visual target of the conical spring washer for detecting bolt loosening provided by the present invention;

[0023] Figure 4 is an elevation view of the working mechanism of the visual target of the conical spring washer for detecting bolt loosening provided by the present invention;

[0024] Figure 5 is a structure diagram of the BoltYOLO bolt loosening detection deep neural network provided by the present invention;

[0025] Figure 6It is the structural diagram of each module of the BoltYOLO bolt loosening detection deep neural network provided by the present invention;

[0026] Figure 7 It is the comparison of the output class activation heatmap results between the BoltYOLO bolt loosening detection deep neural network and YOLO11 in an embodiment provided by the present invention. Specific implementation manner

[0027] The technical solutions in the embodiments of the present invention will be described clearly and completely below. The described embodiments are only a part of the embodiments of the present invention, not all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] This embodiment is described by taking the detection of bolt loosening defects in steel structures as an example. According to the Figure 1 shown process, the bolt loosening detection based on the visual target of the conical spring washer and computer vision technology is completed. The detection is fast, and the detection results are accurate and reliable, effectively avoiding the problems of manpower, hardware consumption, and misjudgment or missed detection caused by human factors.

[0029] Step 1: Design the visual target of the conical spring washer for detecting bolt loosening.

[0030] As Figure 2 shown, the conical spring washer is in a circular disc shape axially. The visual target of the conical spring washer is a circular color band mark drawn on the convex surface of the conical spring washer and along the inner edge of the conical spring washer. The color of the color band is different from the color of the conical spring washer itself, and bright yellow or red can be used;

[0031] As Figure 3 、 Figure 4 shown, during the installation process of steel structure bolts, the connecting piece and the connected piece are fixed by bolts and nuts. When installing the bolts, the conical spring washer with the visual target is installed between the nut and the connecting piece and used as a spring washer on the nut side. A flat washer is used between the bolt head and the connected piece; the working mechanism of the visual target of the conical spring washer is as follows: when the bolt is tightened and in a tight state, the conical spring washer is flattened on the connecting piece by the nut, and the circular color band mark is blocked by the nut and cannot be observed; when the bolt is loose and in a loose state, the conical spring washer rebounds and lifts due to its own stiffness, and a gap appears between the surface of the conical spring washer and the nut, and the circular color band mark is then revealed and can be visually observed. Therefore, the tightness state of the bolt can be judged by whether the visual target of the conical spring washer is visible. To ensure that the visual target of the conical spring washer can be blocked on the tight state bolt and can be normally revealed on the loose state bolt, the following formula is needed to control the width of the circular color band mark:

[0032]

[0033] where w D represents the width of the visual target point of the conical spring washer, i.e., the color band mark; D represents the bolt diameter; R D represents the radius of the inscribed circle of the outer hexagon of the nut corresponding to the bolt with diameter D; d D represents the inner diameter of the conical spring washer corresponding to the bolt with diameter D; η is the width reduction coefficient. When no offset is required, η takes 1.0. When reduction is required, η can take 0.9 or be determined through experiments.

[0034] In this embodiment, the visual target point of the conical spring washer is used for the bolt connection of the steel structure nodes in actual engineering projects, and the bolt specification is M20. The diameter D of the M20 bolt is 20 mm, and the radius R D of the inscribed circle of the outer hexagon of the corresponding nut is 14 mm; the inner diameter d D of the conical spring washer of the corresponding model is 20.4 mm.

[0035] Draw a circular color band mark along the inner edge of the washer on the convex surface of the conical spring washer. The color of the circular color band mark is yellow, and the width of the circular color band mark is determined by the following formula, where the width reduction coefficient η takes 0.95:

[0036]

[0037] Step 2: Collect bolt images with the visual target points of the conical spring washers and manually annotate them to form an original data set.

[0038] Take pictures of the bolts with conical spring washers with visual target points in the project, including two states: the tightened state (the mark of the conical spring washer is invisible) and the loosened state (the mark of the conical spring washer is visible). Take pictures of the bolts from the nut side at different vertical perspectives (15°, 30°, 45°, 60°) and horizontal perspectives (left view 45°, front view, right view 45°) to obtain multiple bolt images. The bolt images include two states: the tightened state and the loosened state. In the tightened state, the mark of the conical spring washer is invisible, and in the loosened state, the mark of the conical spring washer is visible; use the LabelImg tool to manually annotate all bolt images, annotate the data in the YOLO format, and the categories include two categories: "tight" and "loose", which respectively point to the bolts in the tightened state and the loosened state, and generate a data file containing the position of the bounding box and category information. The images and the data file form the original data set.

[0039] Step 3: Build a virtual data derivation network model framework based on the generative adversarial network. First, perform pre-training, and then train on the original dataset to complete transfer learning. The virtual data derivation network model is used to generate more virtual image data based on the original dataset, thereby reducing the requirement of the bolt loosening detection deep neural network model for the quantity of the original dataset.

[0040] The virtual data derivation network model includes a generator G and a discriminator D. The generator G is responsible for generating new data samples, which gradually approach the real data distribution during the training process to deceive the discriminator. The generator is a deep neural network, whose input is a low-dimensional vector and output is a high-dimensional vector. Its input layer receives a low-dimensional random vector to introduce randomness, so that each generated data sample is different; its hidden layer gradually extracts and constructs complex features of the data through non-linear transformation and combination of the input vector; the output layer maps the features of the hidden layer to the final high-dimensional data sample. The discriminator D is responsible for distinguishing whether the input data is real data or fake data generated by the generator. It is a deep neural network, whose input is the high-dimensional vector from the generator and output is a scalar, representing the authenticity probability of the input data; its hidden layer contains various neurons and activation functions to extract features of the input data and gradually abstract its representative information, and the output layer uses the sigmoid activation function to output a scalar between 0 and 1, representing the probability that the input data comes from the real dataset.

[0041] After the virtual data derivation network model framework is built, model training is carried out:

[0042] First, pre-train the virtual data derivation network model on the open-source dataset ImageNet. During the training process, the generator and the discriminator are alternately trained. The latent variable z is generated by the generator, and the discriminator is responsible for discriminating whether the input data is a virtual sample or a real sample. Through adversarial training, the generator tends to generate more real data to deceive the discriminator, while the discriminator tends to improve its discrimination ability to more accurately identify real and generated samples. The loss function used in the training process is:

[0043] V(D,G)=E x～μ [logD(x)]+E z～γ [log(1-D(G(z)))]

[0044] Iterate the parameters during the training process to perform the following optimization process:

[0045]

[0046] After the pre-training is completed, export the generator G init and the discriminator D init, then use the original dataset to train the generator G init and the discriminator D init , complete the transfer learning, and obtain the finished virtual data derivative network model, which can generate bolt image data with visual targets of disc spring gaskets.

[0047] Step 4: Use the virtual data derivative network model to generate a large number of virtual bolt images to achieve fast and large-scale data augmentation. Manually annotate all virtual bolt images, and the annotation process is the same as that of the original dataset; finally, merge the virtual dataset with the original dataset to form a training set.

[0048] Step 5: Build a BoltYOLO deep neural network for bolt loosening detection, and use the training set to train the BoltYOLO deep neural network for bolt loosening detection.

[0049] The BoltYOLO deep neural network for bolt loosening detection, abbreviated as the BoltYOLO model, is based on the YOLO11 architecture and is specifically improved according to the bolt loosening detection task implemented according to the application requirements of the present invention. The overall structure of the BoltYOLO model can be divided into three main parts: the improved backbone network, the improved neck, and three detection heads. Specifically, as Figure 5 shown:

[0050] The improved backbone network successively includes a Conv block, a Conv block, a C3k2 block (c3k is set to True), a Conv block, a C3k2 block (c3k is set to True), a Conv block, a C3k2 block (c3k is set to True), a Conv block, a C3k2 block (c3k is set to True), an SPPF block, and a C2PSA block.

[0051] The improved neck network adds a convolutional block attention module CBAM (Convolutional Block Attention Module) ( Figure 5 the module numbers in it are 17, 21, 25) at the end of the three-scale feature fusion paths of P3, P4, and P5 respectively. Each convolutional block attention module CBAM is connected to a detection head, where: the convolutional block attention module CBAM ( Figure 5 number 17 in it) at the end of the small-scale feature (P3) fusion path, the number of channels of its channel attention module is set to 256, and the convolution kernel size used by its spatial attention module is set to 7; the convolutional block attention module CBAM ( Figure 5 number 21 in it) at the end of the medium-scale feature (P4) fusion path, the number of channels of its channel attention module is set to 512, and the convolution kernel size used by its spatial attention module is set to 7; the convolutional block attention module CBAM (Figure 5 In number 25), the number of channels of its channel attention module is set to 1024, and the size of the convolutional kernel used in its spatial attention module is set to 7.

[0052] Specifically, as Figure 6 , the calculation paths of each module are as follows:

[0053] Conv block: Process the given input tensor (number of channels, height, width) through a 2D convolutional layer, a 2D batch normalization layer, and a SiLU activation function.

[0054] C3k2 block: Control the calculation path executed by the module through the c3k parameter. Specifically: when c3k = False, execute the lightweight module path. The input first passes through the Conv block, then the feature layer is divided into two halves, processed through 2 bottleneck layers, the outputs of each layer are merged, and then connected to the final Conv block; when c3k = True, the calculation process is the same as that of C2f but the bottleneck layer is replaced with a C3k layer.

[0055] SPPF block: Generate multiple fixed-length feature vectors by performing pooling operations on the input feature map at different scales, thereby capturing context information at different scales.

[0056] C2PSA block: Use two PSA modules to operate on different parts of the feature map, then perform feature merging, and then input to the Conv block.

[0057] CBAM block: Sequentially enhance the attention of the BoltYOLO bolt loosening detection deep neural network to the important information related to the visual target point of the disc spring gasket in the input feature map through channel attention and spatial attention.

[0058] The loss function adopted by the BoltYOLO bolt loosening detection deep neural network includes bounding box loss, classification loss, and distribution loss. Among them:

[0059]

[0060] In the formula, S is the number of grids in each row (or column); B represents the number of bounding boxes detected in each grid; (x, y) are the coordinates of the center of the bounding box; (w, h) are the width and height of the bounding box; Indicates whether the bounding box j in grid i is detecting the target.

[0061]

[0062] In the formula, S is the number of grids in each row (or column); y i (c) is the probability that the target detected by the model in grid i belongs to c class; is the ground truth value indicating whether the target in grid i belongs to c class (0 or 1). indicates whether the target is included in bounding box i.

[0063]

[0064] In the formula, N is the number of samples; C is the number of classes; y ic is the ground truth value of sample i; p ic is the detection probability that sample i belongs to class c.

[0065] The training set is composed of the original dataset and the virtual dataset, and then the training set is split proportionally for training, validation, and testing. In the embodiment, it is split according to the ratio of training:validation:testing = 8:1:1. The BoltYOLO bolt loosening detection deep neural network is trained, and the training parameters are as follows in the table:

[0066] Training parameters Value Input image size 640×640 Number of training epochs 300 Number of samples per batch 16 Initial learning rate 0.01 Final learning rate 0.0001 Optimizer SGD

[0067] Step 6: Input the image to be detected into the BoltYOLO model (BoltYOLO bolt loosening deep neural network), and the BoltYOLO model infers and outputs a result image with detection bounding boxes, detection classes, and confidence levels. The test result accuracy and detection speed are as follows in the table.

[0068] Parameter Value mAP50 0.986 mAP50-95 0.974 Speed (ms / image) 12.2

[0069] Step 7: Input the detection result into the model interpretation and visualization module, and finally output and visualize the class activation heatmap through backpropagation and gradient calculation, and use the class activation heatmap to evaluate the model detection result.

[0070] To verify the capture and positioning ability of the BoltYOLO bolt loosening deep neural network for visual targets, and to analyze the relevance between the decision basis of the BoltYOLO bolt loosening deep neural network and visual targets, the present invention proposes a model interpretation and visualization module. Based on the Grad-CAM algorithm, the class activation heatmap is finally output through backpropagation and gradient calculation. The class activation heatmap shows the degree of attention of the neural network to different regions in the input image when making classification decisions. The color depth on the heatmap reflects the degree of attention of the model to different regions, and the darker the color, the greater the contribution of the region to the classification result. Specifically, it is divided into the following steps:

[0071] Step 7.1: Obtain the class score y c for the class c in the forward inference result;

[0072] Step 7.2: Calculate the activation A c of yk Gradient which represents A k The influence degree of each neuron in c on y;

[0073] Step 7.3, calculate the average gradient to obtain the weight α of each feature map k kc :

[0074]

[0075] Step 7.4, use as the weight to linearly combine the feature maps, and activate with ReLU to obtain the class activation heat map:

[0076]

[0077] Step 7.5, match the class activation heat map to the size of the input image and overlay it on the input image for visualization;

[0078] Step 7.6, compare the class activation heat map results of the BoltYOLO model with the YOLO11 model trained using the same dataset and parameters. The results are as Figure 7 . It can be seen that when the BoltYOLO model detects loose bolts, its attention is completely concentrated on the visual target area, and there are no redundant activation areas, proving that other areas cannot interfere with the decision-making of the BoltYOLO model; when detecting tight bolts, the BoltYOLO model transfers its attention to the bolt itself, but still pays attention to the gap between the nut and the disc spring washer, indicating that the absence of the visual target is also the basis for it to detect tight bolts. On the contrary, the YOLO11 model makes decisions completely based on the bolt itself without paying attention to the visual target, verifying the superiority and reliability of the BoltYOLO model.

[0079] Specific implementation cases show that the disc spring gasket marking provided by the present invention has the advantages of being easy to observe, easy to perceive, convenient, stable, and beautiful. The bolt loosening detection method based on it and computer vision has small data volume requirements, can complete the evaluation of the loosening state of a large number of bolts in a short time, ensures high-precision detection while reducing the input cost of hardware equipment and reducing the dependence on human resources, making the detection work more economical and efficient.

[0080] This implementation case proves that the visual target of the disc spring gasket can accurately reflect the loosening condition of the bolt and can be accurately recognized by the computer, demonstrating its effectiveness and reliability in engineering applications. The application case of bolt loosening detection in real engineering proves that under the constraint of a limited data set, the present invention can still complete the evaluation of the loosening state of a large number of bolts in a short time. Through the application of model structure optimization and model interpretation technology, the high accuracy and reliability of the detection results are ensured, effectively avoiding misjudgment or missed detection.

[0081] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application in any way. Any change or modification made by any person skilled in the art based on the disclosed technical content shall be regarded as an equivalent effective embodiment and fall within the scope of protection of the technical solution of the present application.

Claims

1. A bolt loosening detection method based on disc spring washer visual target and computer vision technology, characterized in that: include: Step 1: Design visual targets for disc spring washers used to detect bolt loosening; Step 2: collect bolt images with disc spring gasket visual targets and manually annotate them to form an original data set; Step 3: Build a virtual data derived network model based on the generative adversarial network, first pre-train it, then train it on the original data set to complete transfer learning; Step 4: Generate a large number of virtual bolt images using the virtual data derived network model, manually annotate all images to form a virtual data set, and then merge the virtual data set with the original data set to form a training set; Step 5: Build the BoltYOLO bolt loosening detection deep neural network and train the BoltYOLO bolt loosening detection deep neural network on the training set; Step 6: Input the image to be detected into the BoltYOLO bolt loosening detection deep neural network, and output the result image with the detection bounding box, detection category and confidence after the BoltYOLO bolt loosening detection deep neural network inference.

2. A bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1, characterized in that: In step 1: The disc spring gasket is in the shape of a circular disc in the axial direction, and the visual target of the disc spring gasket is an annular color band mark drawn on the convex surface of the disc spring gasket and along the inner ring edge of the disc spring gasket, and the color of the color band is different from the color of the gasket itself; when installing the bolt, the disc spring gasket with the visual target is installed between the nut and the connecting piece, and is used as a spring washer on the nut side, and a flat washer is used between the bolt head and the connecting piece; The working mechanism of the disc spring washer visual target is as follows: when the bolt is tightened and in a tight state, the disc spring washer is flattened on the connector by the nut, and the annular color band mark is blocked by the nut and cannot be observed; when the bolt is loose, the disc spring washer is lifted up due to its own stiffness, and a gap appears between the surface of the disc spring washer and the nut, and the annular color band mark is then revealed and can be visually observed; in order to ensure that the visual target can be blocked on the bolt in the tight state and can be normally revealed on the bolt in the loose state, the width of the annular color band mark is controlled by the following formula: Where w D Indicates the width of the color band marking on the disc spring washer; D indicates the bolt diameter; R D The radius of the inscribed circle of the outer hexagon of the nut corresponding to the bolt with a diameter of D; d D It indicates the inner diameter of the disc spring washer corresponding to the bolt with diameter D; η is the width reduction factor.

3. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1 is characterized in that: In step 2: The bolt is photographed from the nut side at different vertical and horizontal angles to obtain multiple bolt images, wherein the bolt images include two states: a tight state and a loose state. In the tight state, the disc spring gasket mark is invisible, and in the loose state, the disc spring gasket mark is visible; All bolt images are manually annotated and the data is annotated in YOLO format. The categories include "tight" and "loose", which refer to bolts in tight and loose states respectively. A data file containing the bounding box position and category information is generated. The image and data file constitute the original data set.

4. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1 is characterized in that: In step 3: The virtual data derived network model includes a generator G and a discriminator D. The generator G is obtained after pre-training on the ImageNet dataset. init and the discriminator D init , and then use the original data set to train the generator G init and the discriminator D init Complete transfer learning.

5. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1 is characterized in that: In step 5: The BoltYOLO bolt loosening detection deep neural network is an improved model based on YOLO11, and its structure includes: an improved backbone network, an improved neck and three detection heads, specifically: The improved backbone network includes Conv block, Conv block, C3k2 block (c3k is set to True), Conv block, C3k2 block (c3k is set to True), Conv block, C3k2 block (c3k is set to True), Conv block, C3k2 block (c3k is set to True), SPPF block, and C2PSA block in sequence; The improved neck network adds a convolutional block attention module (CBAM) at the end of the three scale feature fusion paths of P3, P4, and P5. Each convolutional block attention module (CBAM) is connected to a detection head, where: The convolutional attention module CBAM at the end of the small-scale feature P3 fusion path has the number of channels of its channel attention module set to 256 and the convolution kernel size used by its spatial attention module set to 7; The convolutional attention module CBAM at the end of the medium-scale feature P4 fusion path has the number of channels of the channel attention module set to 512 and the convolution kernel size used by the spatial attention module set to 7; The convolutional attention module CBAM at the end of the large-scale feature P5 fusion path has the number of channels of the channel attention module set to 1024 and the convolution kernel size used by the spatial attention module set to 7; The calculation path of each module is: Conv block: processes the given input tensor, including the number of channels, height, and width, through a 2D convolution layer, a 2D batch normalization layer, and a SiLU activation function; C3k2 block: The computational path of the module execution is controlled by the c3k parameter. Specifically, when c3k=False, the lightweight module path is executed. The input first passes through the Conv block, then the feature layer is split into two halves, processed by two bottleneck layers, and the output of each layer is merged and then connected to the last Conv block. When c3k=True, the computational process is the same as C2f, but the bottleneck layer is replaced by the C3k layer. SPPF block: It generates multiple fixed-length feature vectors by performing pooling operations of different scales on the input feature map, thereby capturing contextual information of different scales; C2PSA block: Use two PSA modules to operate on different parts of the feature map, then merge the features and input them into the Conv block; CBAM block: Enhances the BoltYOLO bolt loosening detection deep neural network’s attention to important information related to the disc spring washer visual target in the input feature map through channel attention and spatial attention in turn.

6. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1 is characterized in that: The loss functions used by the BoltYOLO bolt loosening detection deep neural network include bounding box loss, classification loss, and distribution loss, among which: Where S is the number of grids in each row; B is the number of bounding boxes detected in each grid; (x, y) is the coordinate of the center of the bounding box; (w, h) is the width and height of the bounding box; Indicates whether the bounding box j in grid i is detecting the object; Where S is the number of grids in each row; y i (c) The model detects that the target in grid i belongs to c Probability of class; Is the true value, indicating whether the target in grid i belongs to c class (0 or 1); Indicates whether the target is contained in the bounding box i; Where N is the number of samples; C is the number of categories; y ic is the true value of sample i; p ic is the detection probability that sample i belongs to class c.

7. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 1 is characterized in that: Also includes: Step 7: Input the detection results into the model interpretation and visualization module, output the class activation heat map and visualize it; The model interpretation and visualization module is based on the Grad-CAM algorithm. The module input is the category score result for the selected category in the forward reasoning result of the BoltYOLO bolt loosening detection deep neural network. Through back propagation, gradient calculation, weight calculation and activation function, the module finally outputs the class activation heat map.

8. The bolt loosening detection method based on disc spring washer visual target and computer vision technology according to claim 7 is characterized in that: The step 7 comprises the following steps: Step 7.1: Get the forward reasoning results for the category c The category score y c ; Step 7.2, calculate y c Relative to the activation A of the selected convolutional layer k Gradient It means A k Each neuron in y c the extent of the impact; Step 7.3: Calculate the average gradient to get the weight of each feature map k Step 7.4: Use The feature maps are linearly combined as weights and activated with ReLU to obtain the class activation heat map: Step 7.5: Match the class activation heatmap to the size of the input image and overlay it on the input image for visualization.

Images