CN117351320A - Bolt step-by-step detection method based on deep learning - Google Patents

Abstract

The invention discloses a bolt step-by-step detection method based on deep learning, which comprises the following steps: s1, selecting a YOLO v5 network as a deep learning network for identifying and dividing bolts in an image; s2, establishing a bridge bolt identification data set, and evaluating a bolt image to be detected; after the bolts are marked manually, training the selected different YOLO v5 network models by adopting the expanded data set, and selecting an optimal model as a bolt monitoring detection model according to the training result of each YOLO v5 network model; s3, identifying the bolt by the bolt detection model, recording the position of the bolt, and dividing the bolt from the image according to the identification result; the separated single bolt images are classified and judged through an efficientNet network. The invention not only can accurately and automatically detect the health condition of the bridge bolt, but also has high detection rate.

Description

Bolt step-by-step detection method based on deep learning

Technical Field

The invention relates to the field of automatic monitoring of structural construction monitoring, in particular to a bolt step-by-step detection method based on deep learning.

Background

The bolts are used as basic connecting elements, have high reliability, strong bearing capacity and convenient maintenance, and are widely applied. During bridge operation, bolts are often subjected to environmental factors such as vibration load, temperature effects, humidity changes and the like, resulting in damage (loosening and rusting) to the bolts. The loosening of bolts can cause potential partial damage to the structure, and if the structure is not found and maintained in time, the service life of the structure can be shortened, and even collapse is caused. Bolt corrosion can cause reduced fatigue strength performance and thin and weakened local stiffness, ultimately resulting in structural collapse. It is counted that the accidents caused by the loosening of bolts of the British railway are more than 1800; steel bridge structural defects report that korean expressway company managed that 33.3% of bridges had bolt defects. Compared with the research on bridge crack diseases, bridge deformation (deflection, displacement) and the like, the research on bolt disease detection is less concerned, so the research and development on the bolt disease detection are required to be enhanced.

Traditional bolt detection is judged visually by a detector and then further developed into a strain gauge, a torque wrench and the like, but the methods prove unreliable after verification. The bolts are affected by various dynamic loads and environmental factors in the service process and are in a dynamic change process. Its loosening process is mainly divided into two phases: firstly, the friction force of a bolt connection interface is reduced due to the loss of the pretightening force; the threads then deform plastically, eventually leading to a loosening phenomenon. Most early bolt detection is based on the use of a touch sensor to detect bolt pretension to determine if a bolt is loose.

The method based on the contact sensor mainly comprises the sensor methods based on acoustics, piezoelectric active induction, impedance, guided wave and the like. These can obtain detailed mechanical characteristics about the bolt, and even detect whether there is microcrack in the bolt, detection accuracy is high. And based on the mechanical property information of the sensor, the bolts can be modeled, and the Tao Yin (2019) establishes a full-period system dynamics model with bolt flange connection by using a spectral element method and a transmission matrix-based method. However, the touch sensor-based method is susceptible to environmental factors such as temperature and humidity, resulting in a large error in the final result. And be not suitable for the bolt detection of large-scale bridge because the bolt quantity is huge in the bridge, if use the method based on contact sensor, need to arrange a large amount of sensors, with high costs, the installation process is wasted time and energy.

In order to solve the problems of the contact sensor, a non-contact bolt detection method is further developed. Feblil Huda (2013) proposes a vibration detection and health monitoring system based on laser excitation impulse response for detecting bolt looseness. Jay Kumar Shaha (2019) uses ultrasonic pulse transmission technology to detect bolt connection, and researches the correlation among bolt torque, corrosion and mechanical properties of the connection. However, this method is poor in practicality because bolts need to be detected one by one.

For bridge structures, the method can not meet the requirements of rapid and large-scale detection. The development of machine vision technology brings hopes for breaking through the problem. Because the rust and apparent looseness of the bolts can be judged visually, a large number of sensors are not required to be arranged to detect the details of each bolt, and an objective and engineering-required result can be obtained visually. And the method based on machine vision and image processing does not need to carry out a complex equipment installation process, and has the advantages of low equipment cost, high detection speed and wide range. The main equipment only needs to: image data collection devices (CCD cameras, CMOS cameras, unmanned aerial vehicles and other devices with photographing and video recording functions) and computers with image processing capabilities. Jae-Hyung Park (2015) uses a Canny edge detector and a circular Hough transform to segment the image of each nut, then uses Hough transform nut features to estimate the rotation angle of each nut, and finally detects loosening of the bolt by comparing the reference image with the current image. The Young-Jin Cha (2016) uses a smart phone as an image mobile phone device, then uses Hough transformation to extract characteristic features of bolts in the image, uses the extracted characteristics for SVM training, and finally forms a bolt tightness classifier. Through experimental analysis, the method has a certain limit on the angle and distance of the camera. Lovedeep Ramana (2018) uses a Viola-Jones algorithm to locate bolts in an image, and then uses a trained Support Vector Machine (SVM) to classify the tightness of the located bolts.

Most of the bolt detection techniques based on machine vision are based on laboratory studies and have not been validated in field experiments. The method is not very adaptive to the site, and the image processing technology is basically in a machine learning stage, so that an independent and intelligent learning and training method cannot be realized. Xuefang Zhao (2018) combines deep learning and machine vision, uses SSD model for target detection, and then uses image processing technology for angle analysis to judge bolt looseness. The Thank-Canh Huynh (2019) uses RCNN neural network to identify and cut out the target image, and then uses HLC algorithm to extract the characteristics of the cut out image (bolt rotation angle) to judge the tightness. Further improvements are needed because of the need for human intervention during the image correction phase. Hai Chien Pham (2020) uses the bolt-resultant image generated from the graphical model to train a deep-learning model for loose bolt detection. Bolt detection is performed using deep learning (RCNN) based, and then bolt loosening is evaluated by estimating the rotation angle of the detected bolt based on image processing of hough transform. The bolt is identified, positioned and classified by using the fast R-CNN by Yang Zhang (2020), and finally the real-time detection of the bolt in the transmission video by using the network shooting can be realized, the classification precision can be influenced by the illumination condition and the shooting angle, and the method is only verified in a laboratory.

Most of the studies described above have focused on judging bolt loosening by the rotation angle of the bolt, which is not essential in practice, because both rust and loosening of the bolt are damages in the development stage and are not formed suddenly, and the bolt having a slight rust or loosening tendency does not substantially adversely affect the bridge structure. Bolt detection of light rust and loosening is not a major goal based on visual detection techniques. Most of the bolt detection technologies based on vision and deep learning adopt neural networks, and the same model is adopted for training in the 2 steps of bolt detection and classification, so that the accuracy is low.

Disclosure of Invention

The invention aims to: the invention aims to provide a bolt step-by-step detection method based on deep learning, which can accurately and automatically detect the health condition of a bridge bolt.

The technical scheme is as follows: the invention relates to a step-by-step detection method for bolts, which comprises the following steps:

s1, selecting a YOLO v5 network as a deep learning network for identifying and dividing bolts in an image;

s2, establishing a bridge bolt identification data set, and evaluating a bolt image to be detected; if the to-be-detected image is different from the image in the data set in the bolt type, shape and background, adding 5% -10% of the to-be-detected image in the total data set, manually marking the bolt, training the selected different YOLO v5 network models by using the expanded data set, and selecting an optimal model as a bolt monitoring detection model according to the training result of each YOLO v5 network model;

s3, identifying the bolt by the bolt detection model, recording the position of the bolt, and dividing the bolt from the image according to the identification result; the separated single bolt images are classified and judged through an efficientNet network.

In step S2, the depth and width of the YOLO v5 network are controlled by adjusting the depth multiple and the width multiple, respectively, and four models S, m, l, x are selected;

four models are trained s, m, l, x by establishing a bridge bolt identification data set, and the model with the highest target detection precision index is selected as the bolt detection model.

Further, the implementation steps for establishing the bridge bolt identification data set are as follows:

s21, selecting bolt images of three in-service bridges as images of a bolt detection part, extracting 1000 bolts from each bridge bolt image, amplifying and dividing the images by adopting a self-adaptive scale unification method, and obtaining 24000 processed bolt images;

s22, marking the bolt position by labelimg, wherein a normal bolt is marked as a bolt, a rusted bolt is marked as a corrobot, and a bolt loosening or missing is marked as a loosbolt;

s23, according to 8:1:1, dividing the processed bolt image into a training image, a test image and a verification image, and calculating the sizes of 12 groups of anchors by adopting a k-means clustering method for the sizes of the markers.

Further, in step S3, the implementation steps of classifying and discriminating the segmented single bolt image through the efficientNet network are as follows:

s31, processing a bridge bolt identification data set, and dividing bolts according to the result of manual marking to form three types of bolt images, namely normal bolts, rusted bolts and loose bolts; training the efficientNet network by taking the three bolt images as training sets;

s32, changing the marked names of the marked rust bolts and the marked loose bolts into common bolts, and training the bolt detection model and the efficientNet network again by using the data set.

Compared with the prior art, the invention has the following remarkable effects:

1. the bolt detection is divided into two stages of bolt identification segmentation and bolt disease classification, two networks are adopted for segmentation and classification operation respectively, and the existing method is to directly adopt a target detection network for segmentation and classification of detection targets. Compared with the existing method, the method has higher detection precision; when the YOLO v5 network is applied, the network is further divided into s, m, l, x four sizes, the four sizes are respectively tested, and the optimal network is selected to execute the bolt detection segmentation operation, but the existing method does not change the segmentation operation;

2. aiming at the problems that the number of bolts at bridge nodes is large, the size of an image obtained by shooting by an unmanned aerial vehicle is limited, so that the target size of the bolts is small and is difficult to completely identify and divide, a small target detection layer is added on the basis of a utilized YOLO v5 network, and the sizes of the small target detection layer are set into three groups of [5,6], [8,14], [11,15] according to the sizes of small-scale bolts in a bolt data set, so that the problems that the downsampling multiple of the YOLO v5 network is large and a deeper feature map is difficult to learn small target feature information are solved;

3. aiming at the problem that the traditional bolt detection method based on deep learning directly uses a target detection network to identify bolts and judges the diseases simultaneously, the method has low disease judging accuracy, the method of judging the diseases step by adopting two networks is provided, and the bolts are firstly identified and split and then the bolt diseases are judged; the method is characterized in that the bolts are rapidly identified and automatically segmented by comparing with a latest YOLO v5 series network, and the bolt diseases are judged by adopting an improved EFicientNet network; compared with the traditional method which only adopts single model detection, the method has higher precision;

4. aiming at the problem that the common classification network possibly has difficulty in learning the characteristics with enough classification distinction because the difference of image texture characteristics among different types of diseases (such as looseness and normal bolts) is small when the bolt diseases are classified, in order to realize more accurate bolt disease classification, the independent heat codes are modified into smooth tag codes on the basis of the original efficientNet network, and the classification capability of the model on the characteristics of micro-distinction is enhanced by adding a probability prediction function into a prediction tag;

5. the invention is based on in-service bridge engineering in terms of data acquisition, training set production and method verification, but not laboratory components in most of the existing researches; in the process of manufacturing the data set, a bolt target detection data set and a bolt disease classification data set are manufactured respectively and are used for training a bolt identification segmentation model based on improved YOLO v5 and a bolt disease classification model based on improved efficientNet respectively. In order to make the data set have stronger generalization capability, after the manual marking of the data set is finished, scaling and random splicing processing are carried out on the images in the data set to increase the number of the data set, and generating an anti-network to add random shadow shielding, rain and fog noise, random exposure and random stains to the images, so that various interference factors possibly encountered in an engineering scene are simulated, and the generalization capability of the model can be improved; the verification of the suspension bridge shows that the method can accurately and automatically detect the health condition of the bridge bolts, achieves the same precision as manual detection, takes only half a day for detecting the field and the inner industry of hundreds of thousands of bolts, and has great engineering significance as proved by the fact that the speed is far beyond the manual detection.

Drawings

FIG. 1 is a schematic diagram of a frame of a method for rapidly detecting a bridge bolt based on an unmanned aerial vehicle;

FIG. 2 is a schematic diagram of a YOLO v5s network structure employed in the present invention;

FIG. 3 is a schematic diagram of the P-R curve of the YOLO v5s model used in the present invention;

FIG. 4 is a schematic diagram of the P-R curve of the YOLO v5m model used in the present invention;

FIG. 5 is a schematic diagram of the P-R curve of the YOLO v5l model used in the present invention;

FIG. 6 is a schematic diagram of the P-R curve of the YOLO v5x model used in the present invention;

FIG. 7 is a schematic diagram of a training process of the YOLO v5x model of the present invention;

fig. 8 is a schematic diagram of a training process of the efficientNet network of the present invention.

Detailed Description

The invention is described in further detail below with reference to the drawings and the detailed description.

According to the invention, the YOLO v5 network is adopted for bolt detection, and then the efficientNet network is used for bolt classification, so that the method has pertinence, and the speed and the accuracy of detection and classification are improved. Simultaneously, the detected blurred bolt image is identified, the image is optimized by using a deep learning method based on super resolution, and then the image is input into a classification model for classification

According to the invention, the latest YOLO v5 algorithm is applied to automatically mark and divide the bolts into single units, the single unit bolt images are input, and then the single unit bolt images are automatically classified into four types of normal bolts, rusted bolts, loose bolts and missing bolts by using an efficientNet algorithm, so that the automatic analysis method flow from data acquisition to bolt detection is completed. The method framework is shown in fig. 1.

A bolt step-by-step detection method based on deep learning comprises the following steps:

and 1, selecting a model series. Since the industrial application requires as fast online detection as possible and the computing platform on which the model depends as inexpensive as possible, the development direction of deep learning in image detection in recent years has been focused on the high frame rate, low computational performance requirement direction, which are also important for unmanned aerial vehicle-based bridge detection. The YOLO series model is one of the models with optimal performance and the most extensive application, and the verification of a large number of databases shows that the model meets engineering requirements by considering that the latest YOLO v5 series model already has very high recognition rate and detection precision, and the model of the YOLO v5 series has strong multi-platform portability, so that the invention directly applies the latest YOLO v5 series model. FIG. 2 is a network structure of Yolo v5s, similar to other network structures of the Yolo v5 series, except that CSP structures using a different number of residual components and Focus structures of a different number of convolution kernels are employed. The network is divided into four parts, namely an input end and a Backbone, neck, prediction.

And 2, training the model and selecting an optimal model.

The YOLO v5 series network has s, m, l, x models by adjusting depth multiple and width multiple to control the depth and width of the network respectively. And establishing a bridge bolt identification data set, training four models, and comparing the effects of the four models. After training, mAPs of four models of YOLO v5s, YOLO v5m, YOLO v5l and YOLO v5x are respectively 0.841, 0.934, 0.936 and 0.940, and the accuracy of YOLO v5x with the largest network depth and width is the highest, and the model can be selected as a bolt detection model under the condition of not considering real-time detection.

The YOLO v5s network is the network with the smallest depth and the smallest width of the feature map in the YOLO v5 series. The YOLO v5 series network controls the depth and width of the network by adjusting the depth multiple and width multiple, respectively, which are [0.33,0.50], [0.67,0.75], [1.0,1.0], [1.33,1.25] for the s, m, l, x four models, respectively, with a consequent increase in model size. The increased depth and width give the model an increasingly higher AP, but the computational efficiency decreases. Because bolt identification is a simpler detection task than multi-type target identification, even though APs of four models should be distinguished in theory, differences may not be present in the data set of bolt detection, and thus it is necessary to build a bridge bolt identification data set, train the four models, and compare the effects of the four models.

The implementation steps for establishing the bridge bolt identification data set are as follows:

and 21, taking images of the bolt detection parts from three in-service bridges, respectively extracting 1000 bolt images from the three bridge bolt images, and expanding the images by adopting a method of selecting, randomly splicing and generating countermeasure network enhancement to obtain 24000 bolt images.

Step 22, marking the bolt position by labelimg, wherein a normal bolt is marked as a bolt, a rusted bolt is marked as a corrobot, and a bolt loosening or missing is marked as a loosbolt. After the manual marking is completed, a 3-category bridge bolt database required by training is formed, and marking results show that the database contains 237600 normal bolts, 21583 rusted bolts and 195 loose bolts. The bolts of the small-span suspension bridge are mainly cable clamp bolts and bridge bottom steel truss connecting bolts, the large-span cable-stayed bridge is mainly a connecting bolt on the side face of the steel truss, and the large-span suspension bridge is a cable clamp connecting bolt.

Step 23, training a model by using a pytorch frame, wherein a computer used for training is i7 9700k CPU (Central processing Unit) and 32G memory, 1 NVIDIA RTX3090 display card, the training steps are 50000 steps, and before training, the training steps are firstly according to 8:1:1, dividing the image into a training image, a test image and a verification image, statistically analyzing bolt size data in all training set images, clustering the sizes of 12 groups of anchors from the size data by adopting a k-means clustering method, so that the predicted size of a network accords with the bolt size in the data set, and setting the size, namely, a model is provided with a predicted layer with four scales of micro scale, small scale, medium scale and large scale, wherein the calculation result is [5,6,8,14, 11,15, 12,32,40,68,52,100,60,80,44,72,44,44,80,80,160,200,120,140].

After training, the mAP of four models of YOLO v5s, YOLO v5m, YOLO v5l and YOLO v5x are respectively 0.841, 0.934, 0.936 and 0.940, the accuracy of YOLO v5x with the largest network depth and width is the highest, and the model can be selected as a bolt detection model under the condition of not considering real-time detection, and the P-R curves in the training process of the four models are respectively shown in fig. 3, 4, 5 and 6.

And 3, customizing model parameters on a bolt database, and further optimizing a bolt detection method based on a YOLO v5 series model. Because the method for guaranteeing the correct identification of the bolts is the most important purpose, a method for adopting a target detection model and a classification model to detect step by step is proposed based on the method, firstly, the bolts are identified by a YOLO v5x model, the positions of the bolts are recorded, and the bolts are segmented from an image according to the identification result; the separated single bolt images are classified through an efficientNet network, and normal bolts, rusted bolts and loose bolts are distinguished.

From the P-R curve, the identification accuracy of the three types of bolts is different. The classification accuracy of the test after the training is finished is 0.883, which shows that although the recognition rate of the three types of bolts is very high, the classification of a part of bolts is wrongly recognized, a part of images are extracted for testing after the training is finished, and the test result shows that a part of normal bolts are recognized as rusted bolts.

The detection method for the further optimized bolt can be obtained by the following steps:

firstly, processing a bolt database established in the front, dividing bolts according to the result of manual marking to form three types of bolt images, and training an efficientNet network by taking the divided images as a training set;

and then processing the marked result, changing the marked names of the marked rust bolt and the marked loose bolt into a common bolt, namely, the modified data set is a data set only comprising a bolt type, and training the YOLO v5x model again by using the data set, wherein under the condition that other parameters are consistent, the mAP of the trained model reaches 0.997, and almost all bolts can be correctly identified. The pytorch frame is adopted in training of the efficientNet network, the training steps are 5000 steps, the learning rate is 0.001, the test precision after training is 0.993, and the precision is far higher than the precision of identifying and classifying simultaneously by using only the YOLO v5x model. The parameters of the YOLO v5x model and the efficientNet network training process are varied as shown in fig. 7 and 8. The training is completed, and the model identification and classification precision tested in the training are over 99 percent, so that the trained model can be used for bolt detection work.

In order to fully verify the practicability and the detection efficiency of the invention, the in-service bridge is taken as a test object, and whether the proposed method meets the engineering detection requirement is checked. The test results of the invention on an in-service bridge are introduced below, and the invention mainly detects the bottom of the bridge and the cable clamp bolts.

(1) Bridge profile

The test bridge is a pedestrian suspension bridge which is arranged in a seawater channel engineering, the total length of the bridge is 197.7 meters, the middle span is 115.7 meters, and the bridge deck is 2.7 meters wide. The bridge deck system is of a longitudinal and transverse channel steel structure, wherein the total number of the transverse short channel steel of the bridge is 53, the number of the longitudinal channel steel is 5, and the number of cable clamps is 106. The bolts of interest in the detection are bridge bottom channel steel connecting bolts and sling clamp fixing bolts, wherein the channel steel connecting bolts 371 and the sling clamp bolts 742. The bridge has been in service for 18 years, a part of bolts are loosened and rusted, but the bridge is located above a sea water channel, so that the quick detection is difficult for people.

According to the shooting method, as for the inhaul cables on the side surface of the bridge, the unmanned aerial vehicle is suspended at three positions close to the two ends and the middle of the bridge in a distributed manner, and the camera is controlled to move to shoot the cable clamp bolts; for the bridge bottom, hover at both ends and middle position of the bottom as well, shoot upwards. A total of about 94 minutes of bolt video was taken during the test, covering all bolts required to be tested. The resolution of the video was 1920×1080pixel, the frame rate was 30fps, and the shutter speed was set to 1/100s and the sensitivity was automatic in order to match the parameters such as image exposure at the time of shooting.

(2) Bolt identification and disease discrimination results

For the detection of large-batch bridge bolts, ensuring that all bolt diseases can be accurately identified is the most important requirement in engineering, so the invention provides a series of methods for improving the automatic identification accuracy and stability of the bolts. In the invention, in the automatic identification and disease discrimination of the bridge bolt, three methods of a method for directly identifying and discriminating the bolt disease, a method for identifying the bolt firstly and then discriminating the bolt and a method for discriminating the bolt manually are adopted at the same time, and the accuracy of the three methods is compared to verify the accuracy of the proposed method. The specific implementation mode of the method provided by the invention is as follows: firstly, deriving a bolt video acquired in unmanned aerial vehicle detection frame by frame to obtain a bolt image, carrying out bolt identification by adopting a single-class YOLO v5x model of bolt target detection, and dividing the identified bolt into single bolt images; and then outputting the single bolt image to a trained efficientNet network, and automatically judging whether the bolt is a normal bolt, a rusted bolt or a loose and falling bolt. The prior method for comparison is implemented by directly outputting the unified scale segmentation image to a trained three-category YOLO v5x model and directly segmenting the image into three-category bolt images. The manual judging method is to observe whether the bolt in the image shot by the unmanned aerial vehicle has diseases or not by a manual visual detection method, and the result of manual detection is taken as a group trunk. The speed of the two automatic identification methods on a computer adopted in training is very high, wherein the identification speed of the YOLO v5x model is about 0.08 s/sheet, the classification speed of the efficientNet model is about 0.02 s/sheet, and the time required for automatically judging the full-bridge bolt is not more than 20 minutes. However, the method provided by the invention can achieve the bolt detection result consistent with manual discrimination, and the method of direct identification and classification is adopted to have a certain degree of misjudgment on bolt identification and automatic discrimination, wherein 1.2% of bolts are not identified, and 13.11% of bolt category judgment errors exist.

Claims (4)

1. The bolt step-by-step detection method based on deep learning is characterized by comprising the following steps of:

s1, selecting a YOLO v5 network as a deep learning network for identifying and dividing bolts in an image;

s2, establishing a bridge bolt identification data set, and evaluating a bolt image to be detected; if the to-be-detected image is different from the image in the data set in the bolt type, shape and background, adding 5% -10% of the to-be-detected image in the total data set, manually marking the bolt, training the selected different YOLO v5 network models by using the expanded data set, and selecting an optimal model as a bolt monitoring detection model according to the training result of each YOLO v5 network model;

s3, identifying the bolt by the bolt detection model, recording the position of the bolt, and dividing the bolt from the image according to the identification result; the separated single bolt images are classified and judged through an efficientNet network.

2. The bolt step detection method based on deep learning according to claim 1, wherein in step S2, four models S, m, l, x are selected by adjusting depth multiple and width multiple to control the depth and width of the network respectively for the YOLO v5 series network;

four models are trained s, m, l, x by establishing a bridge bolt identification data set, and the model with the highest target detection precision index is selected as the bolt detection model.

3. The deep learning-based bolt step detection method according to claim 1, wherein the implementation steps of establishing the bridge bolt identification dataset are as follows:

s21, selecting bolt images of three in-service bridges as images of a bolt detection part, extracting 1000 bolts from each bridge bolt image, amplifying and dividing the images by adopting a self-adaptive scale unification method, and obtaining 24000 processed bolt images;

s22, marking the bolt position by labelimg, wherein a normal bolt is marked as a bolt, a rusted bolt is marked as a corrobot, and a bolt loosening or missing is marked as a loosbolt;

s23, according to 8:1:1, dividing the processed bolt image into a training image, a test image and a verification image, and calculating the sizes of 12 groups of anchors by adopting a k-means clustering method for the sizes of the markers.

4. The bolt step detection method based on deep learning as claimed in claim 3, wherein in step S3, the implementation steps of classifying and discriminating the segmented single bolt image through an efficientNet network are as follows:

s31, processing a bridge bolt identification data set, and dividing bolts according to the result of manual marking to form three types of bolt images, namely normal bolts, rusted bolts and loose bolts; training the efficientNet network by taking the three bolt images as training sets;

s32, changing the marked names of the marked rust bolts and the marked loose bolts into common bolts, and training the bolt detection model and the efficientNet network again by using the data set.