Method and device for detecting appearance defects of maglev train, electronic equipment and storage medium

Abstract

The application discloses a maglev train appearance defect detection method and device, electronic equipment and a storage medium. The method comprises the following steps: a deep convolution network and a graph neural network are used to respectively extract feature vectors of normal train images, to obtain a feature description vector set of the normal train images, and to perform normal fitting; an adaptive sliding window is used to detect key components after segmenting a to-be-detected train image; if the number of the detected key components is less than that in the normal train image, an area where the missing key components are located is extracted to obtain a missing area picture; a deep convolution network and a graph neural network are used to respectively process the extracted missing area picture to obtain semantic feature vectors and regional structure feature vectors, the semantic feature vectors and the regional structure feature vectors are weighted, a feature description vector set of the missing area picture is obtained, and normal fitting is performed; and the feature description vector set is compared to determine whether the missing area picture is abnormal. The application improves train appearance defect detection efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of maglev train inspection and maintenance technology, and in particular, to a method, device, electronic equipment and storage medium for detecting appearance defects in maglev trains. Background Technology

[0002] Currently, the maintenance methods for maglev trains are traditional manual inspections conducted in the depot during daily returns. The main approach involves using simple measuring tools and flashlights to inspect the trains according to the prescribed vehicle maintenance procedures. This method is extremely inefficient, and the effectiveness relies entirely on worker experience, wasting significant manpower and resources. To address and improve the vehicle department's maintenance procedures and operational models, intelligent and automated inspection methods are being adopted to reduce the workload of manual inspections after daily vehicle returns. This will enable early warning of the status of critical maglev train components. Through big data analysis and intelligent data mining, alarms will be issued for abnormal conditions, facilitating timely vehicle maintenance. This will allow for the detection of anomalies in maglev train components such as brake calipers, suspension frames, suspension power supplies, and high-voltage reactor boxes, providing a basis for inspection and maintenance.

[0003] Currently, there is no technology for automated detection of anomalies in the exterior of maglev trains. However, there are technologies for automated algorithms for detecting anomalies in the exterior of subway trains. The main methods for detecting anomalies in the exterior of subway trains are as follows:

[0004] 1. After identifying abnormal areas of the target component through 2D grayscale image comparison, the 3D data of the abnormal areas is then compared to determine whether there are problems such as missing or loose parts. This method is only applicable to static inspection of vehicles in a garage.

[0005] 2. A set of magnets installed on the rails and cameras installed between and outside the rails achieve accurate image correction through precise speed measurement, and anomaly detection is achieved through difference between the image and a standard image. This method is simple and effective, but it is only suitable for wheel-rail vehicle inspection.

[0006] 3. Feature point detection is performed on the current running vehicle image and the template image. Feature points from the current image and the template image are extracted and paired. Alignment of the current running vehicle image and the template image is achieved based on the positional features of the matched feature point pairs. By comparing the slider, the aligned current running vehicle image and template image are traversed, and the contrast difference features of each region are extracted. When the contrast difference feature value is greater than a preset threshold, the current region is judged to be abnormal. This method is time-consuming and not suitable for real-time detection.

[0007] However, maglev trains are non-wheel-rail vehicles and require the maglev track to be energized during operation. Therefore, it is impossible to install magnets on the track to obtain the accurate speed of the maglev train. As a result, the collected images of the train body will contain some stretching or compression, and the effect of correction and alignment with the template image is poor. Therefore, the accuracy of the train body anomaly detection achieved by differential detection is poor, and the positioning error of abnormal parts is large. Summary of the Invention

[0008] This application provides a method for detecting appearance defects in maglev trains to solve the technical problems of poor accuracy and low efficiency in the detection of abnormalities in train bodies in the prior art.

[0009] The technical solution adopted in this application is as follows:

[0010] A method for detecting appearance defects in maglev trains includes the following steps:

[0011] Deep convolutional networks and graph neural networks are used to extract feature vectors from normal train images to obtain a set of feature description vectors for normal train images and then perform normal fitting. The set of feature description vectors includes semantic feature vectors extracted by deep convolutional networks and regional structure feature vectors extracted by graph neural networks.

[0012] An adaptive sliding window is used to segment the actual acquired images of the train under test and then detect key components.

[0013] If the number of key components detected is less than that in a normal train image, the area where the missing key components are located is extracted to obtain an image of the missing area.

[0014] The semantic feature vector and regional structure feature vector obtained after processing the extracted missing region image using deep convolutional network and graph neural network are weighted to obtain the feature description vector set of the missing region image and then normalized.

[0015] The feature description vector set of the missing region image is compared with the feature description vector set of the normal train image, and the presence of anomalies in the missing region image is determined by calculating using the Gaussian density estimation function.

[0016] Furthermore, the step of using deep convolutional networks and graph neural networks to extract feature vectors from normal train images to obtain a set of feature description vectors for normal train images and then performing normal fitting, wherein the set of feature description vectors includes semantic feature vectors extracted by deep convolutional networks and regional structure feature vectors extracted by graph neural networks, specifically including the following steps:

[0017] Images of normal trains are captured using a camera, and after speed correction, undistorted images of normal trains are obtained.

[0018] A fixed-size window was used to slide and segment the acquired normal train images into several image sequences, and key components were labeled.

[0019] Deep convolutional networks and graph neural networks are used to extract feature vectors from the segmented image sequences, respectively, to obtain semantic feature vectors and regional structure feature vectors of key components, forming a set of feature description vectors for normal train images.

[0020] The feature description vector set of the extracted normal train images is fitted using a normal distribution.

[0021] Furthermore, the step of segmenting the actual acquired train image using an adaptive sliding window and then detecting key components specifically includes the following steps:

[0022] Images of the train under test are captured using a camera, and after speed correction, undistorted images of the train under test are obtained.

[0023] The target detection software is used to segment the image of the train under test into an image sequence and perform key component detection in a window of a set size;

[0024] If no key components are detected in the image, the sliding segmentation will continue by sliding a distance of a set number of pixels and the detection of key components will continue.

[0025] If a critical component is detected, the image is segmented and the detection of critical components continues until the entire image of the train under test is detected.

[0026] The window size is between 1300 pixels * 600 pixels and 1600 pixels * 1000 pixels, and the set number of pixels is between 400 and 600.

[0027] Furthermore, if the number of detected key components is less than that in a normal train image, the area where the missing key components are located is extracted to obtain a missing area image, specifically including the following steps:

[0028] Based on the coordinates of the images, the same key components detected in adjacent images are merged to accurately locate the key components in the image of the train under test.

[0029] If the number of detected critical components is less than that in a normal train image, it is considered that a critical component is suspected to be missing. Then, the location of the missing critical component on the image is located by using the locations of other detected critical components, and the area where the missing critical component is located is extracted to obtain the image of the missing area.

[0030] Furthermore, the semantic feature vector and region structure feature vector obtained by processing the extracted missing region image using deep convolutional networks and graph neural networks are weighted to obtain a feature description vector set for the missing region image, specifically including the following steps:

[0031] The semantic feature vectors of key components in the missing region image are obtained by using a deep convolutional network with a kernel size of 3×3.

[0032] A graph neural network is used to aggregate features of key component locations in a missing region image to obtain the region structure feature vector of the image.

[0033] The set of feature description vectors h for the missing region image is obtained by weighting the semantic feature vectors and the region structure feature vectors. * And perform normal fit:

[0034] h * =αh i +βh j

[0035] In the formula, h i h is a semantic feature vector. j Let be the feature vector of the region structure, and α and β be the corresponding weights. The weight α is positively correlated with the detection confidence of the key component by the deep convolutional network, and the weight β is positively correlated with the detection confidence of the key component by the graph neural network.

[0036] Furthermore, the weights α and β are determined through the following steps:

[0037] Key components of normal train images were detected using both deep convolutional networks and graph neural networks. The detection confidence scores of each key component were recorded for both networks, and the average detection confidence score v for each key component was obtained for each network. t-1 ;

[0038] In the actual detection process of the two networks, the actual detection confidence θ of the key components was obtained for each. t Then, the target confidence level v of the two networks is calculated using the following formulas. t :

[0039] v t =av t-1 +(1-a)θ t

[0040] Where a is v t-1 The weights are set and a is set to 0.9;

[0041] The weights α and β are calculated based on the proportion of target confidence in the two networks.

[0042] Another preferred embodiment of this application provides a magnetic levitation train appearance defect detection device, including:

[0043] The module for obtaining the feature description vector set of normal train images is used to extract feature vectors from normal train images using a deep convolutional network and a graph neural network, respectively, to obtain the feature description vector set of normal train images and to perform normal fitting. The feature description vector set includes semantic feature vectors extracted by the deep convolutional network and regional structure feature vectors extracted by the graph neural network.

[0044] The critical component detection module is used to detect critical components after segmenting the actual acquired images of the train under test using an adaptive sliding window.

[0045] The missing area image extraction module is used to extract the area where the missing key components are located and obtain the missing area image if the number of detected key components is less than that in the normal train image.

[0046] The module for obtaining the feature description vector set of the missing region image is used to weight the semantic feature vector and the region structure feature vector obtained after processing the extracted missing region image using a deep convolutional network and a graph neural network, respectively, to obtain the feature description vector set of the missing region image and perform normal fitting.

[0047] The defect confirmation module is used to compare the feature description vector set of the missing region image with the feature description vector set of the normal train image, and to determine whether there is an anomaly in the missing region image by calculating the Gaussian density estimation function.

[0048] Another preferred embodiment of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the maglev train appearance defect detection method.

[0049] Another preferred embodiment of this application also provides a storage medium including a stored program that, when the program is executed, controls the device where the storage medium is located to perform the steps of the maglev train appearance defect detection method.

[0050] Compared with the prior art, this application has the following beneficial effects:

[0051] 1. This application uses a defect detection model based on deep convolution and graph neural networks to accurately detect anomalies in key components. By weighting and summing the semantic feature vector and regional structure feature vector of the input image, the feature description vector of the image is obtained. Deep convolution and graph neural networks can verify each other, eliminating one-sided missed detections and false detections, and can obtain the feature description of the image more accurately, thereby improving the detection rate of key components.

[0052] 2. This application uses an adaptive sliding window algorithm to segment train images. After detecting the location of key components, the sliding distance can be adaptively changed, thereby reducing the repeated detection of the same key component, improving detection efficiency and reducing detection time.

[0053] 3. This application uses a normal distribution to fit the feature description vector set of the extracted normal train images, and uses a Gaussian density estimation function to calculate and determine whether the input image is abnormal, thereby improving detection efficiency and feature matching accuracy.

[0054] 4. The detection method of this application does not require the installation of expensive high-precision speed sensors or other equipment on the maglev train track, resulting in low equipment cost and ensuring the safety of maglev operation;

[0055] In addition to the purposes, features, and advantages described above, this application provides other purposes, features, and advantages. The application will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0056] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0057] Figure 1 This is a schematic diagram of the process for detecting appearance defects in a maglev train according to a preferred embodiment of this application.

[0058] Figure 2 A flowchart illustrating the sub-steps of step S1 in the preferred embodiment of this application.

[0059] Figure 3 A flowchart illustrating the sub-steps of step S2 in the preferred embodiment of this application.

[0060] Figure 4 A flowchart illustrating the sub-steps of step S3 in the preferred embodiment of this application.

[0061] Figure 5 A flowchart illustrating the sub-steps of step S4 in the preferred embodiment of this application.

[0062] Figure 6 A flowchart illustrating the sub-steps of step S43 in the preferred embodiment of this application.

[0063] Figure 7 This is a schematic diagram of a method for detecting appearance defects in maglev trains according to another preferred embodiment of this application.

[0064] Figure 8This is a schematic diagram of a maglev train appearance defect detection device module according to a preferred embodiment of this application.

[0065] Figure 9 This is a schematic diagram of a magnetic levitation train appearance defect detection device module according to another preferred embodiment of this application.

[0066] Figure 10 This is a schematic block diagram of an electronic device according to a preferred embodiment of this application.

[0067] Figure 11 This is an internal structural diagram of a computer device according to a preferred embodiment of this application. Detailed Implementation

[0068] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0069] The overall detection system adopted in this application mainly consists of a gantry, line-scan cameras, a laser light source, a speed measurement unit, a central control unit, and an analysis and service unit. It enables 360-degree image detection of the maglev train's body to identify surface faults. Multiple line-scan cameras and laser light sources are mounted on the gantry, which stands above the maglev track. When the train passes the gantry, the line-scan cameras and laser light sources activate, acquiring real-time 360-degree images of the train's roof, sides, and undercarriage. Simultaneously, the speed measurement unit acquires the train's speed as it passes the gantry, providing speed data for the analysis and service unit to reconstruct the train images.

[0070] Reference Figure 1 A preferred embodiment of this application provides a method for detecting appearance defects in maglev trains, including the following steps:

[0071] S1. Feature vectors are extracted from normal train images using deep convolutional networks and graph neural networks respectively to obtain a set of feature description vectors for normal train images, and then normality is fitted. The set of feature description vectors includes semantic feature vectors extracted by deep convolutional networks and regional structure feature vectors extracted by graph neural networks. Normal train images are acquired by line scan cameras mounted on gantry frames in the detection system.

[0072] S2. After segmenting the actual acquired train image, an adaptive sliding window is used to detect key components.

[0073] S3. If the number of key components detected is less than that in a normal train image, extract the area where the missing key components are located to obtain a missing area image.

[0074] S4. After processing the extracted missing region images using deep convolutional networks and graph neural networks respectively, the semantic feature vectors and regional structure feature vectors are weighted to obtain a set of feature description vectors for the missing region images and then normalized.

[0075] S5. Compare the feature description vector set of the missing region image with the feature description vector set of the normal train image, and use the Gaussian density estimation function to determine whether the missing region image has any anomalies.

[0076] The detection method in this embodiment is based on a defect detection model using deep convolution and graph neural networks to accurately detect anomalies in key components. It obtains the image's feature description vector by weighting the semantic feature vector and regional structure feature vector of the input image. Deep convolution and graph neural networks can mutually verify each other, eliminating one-sided missed detections and false detections, thus obtaining more accurate feature descriptions and improving the detection rate of key components. This embodiment uses an adaptive sliding window algorithm to segment the train image. After detecting the location of a key component, the sliding distance can be adaptively adjusted, thereby reducing repeated detection of the same key component, improving detection efficiency, and reducing detection time. This embodiment uses a normal distribution to fit the feature description vector set of the extracted normal train image, and calculates whether the input image is abnormal using a Gaussian density estimation function, improving detection efficiency and feature matching accuracy. The detection method in this embodiment does not require the installation of expensive high-precision speed sensors or other equipment on the maglev train track, resulting in low equipment cost and ensuring the safety of maglev operation.

[0077] Specifically, such as Figure 2 As shown, the method employs deep convolutional networks and graph neural networks to extract feature vectors from normal train images to obtain a set of feature description vectors for the normal train images, and then performs normal fitting. The set of feature description vectors includes semantic feature vectors extracted by the deep convolutional network and regional structure feature vectors extracted by the graph neural network. Specifically, the method includes the following steps:

[0078] S11. Use a camera to capture images of normal trains, and obtain undistorted images of normal trains after speed correction.

[0079] S12. Using a fixed-size window, the acquired normal train images are divided into several image sequences by sliding segmentation, and key components are labeled.

[0080] S13. Deep convolutional networks and graph neural networks are used to extract feature vectors from the segmented image sequences to obtain semantic feature vectors and regional structure feature vectors of key components, forming a set of feature description vectors for normal train images.

[0081] S14. Fit the feature description vector set of the extracted normal train images using a normal distribution.

[0082] Specifically, such as Figure 3 As shown, the step of segmenting the actual acquired train image using an adaptive sliding window and then detecting key components specifically includes the following steps:

[0083] S21. Use a camera to acquire images of the train under test, and obtain undistorted images of the train under test after speed correction. When the train passes the vehicle arrival detection unit, the vehicle arrival detection unit transmits the train arrival signal to the control center. The control center controls the camera on the detection gantry to take pictures of the passing vehicle body to obtain images of the train under test.

[0084] S22. Using target detection software such as YOLOv4-tiny, the image of the train under test is segmented into an image sequence in a window of 1500 pixels * 800 pixels and key components are detected.

[0085] S23. If no key components are detected in the image, slide segmentation is performed by sliding a distance of 500 pixels forward and key component detection continues.

[0086] S24. If a critical component is detected, the image segmentation is performed by sliding the window to the next target position using the coordinate position of the critical component as the window, and the detection of critical components continues until the entire image of the train to be tested is detected.

[0087] Specifically, such as Figure 4 As shown, if the number of detected key components is less than that in a normal train image, the area where the missing key components are located is extracted to obtain a missing area image. This specifically includes the following steps:

[0088] S31. Merge the same key components detected in adjacent images according to the coordinate position of the images to complete the accurate positioning of the key components in the image of the train under test.

[0089] If the number of key components detected by S32 is less than that in a normal train image, it is considered that the key component is suspected to be missing. Then, the location of the missing key component on the image is located by the location of other detected key components, and the area where the missing key component is located is extracted to obtain the missing area image.

[0090] Specifically, such as Figure 5 As shown, the semantic feature vector and region structure feature vector obtained by processing the extracted missing region image using a deep convolutional network and a graph neural network are weighted to obtain a feature description vector set for the missing region image. The specific steps include:

[0091] S41. Use a deep convolutional network to obtain the semantic feature vectors of key components in the image of the missing region. The size of the convolutional kernel is 3×3.

[0092] S42. Use graph neural networks to aggregate features of key component locations in the missing region image to obtain the regional structure feature vector of the image.

[0093] S43. After weighting the semantic feature vector and the region structure feature vector, a set of feature description vectors h for the missing region image is obtained. * And perform normal fit:

[0094] h * =αh i +βh j

[0095] In the formula, h i h is a semantic feature vector. j Let be the feature vector of the region structure, and α and β be the corresponding weights. The weight α is positively correlated with the detection confidence of the key component by the deep convolutional network, and the weight β is positively correlated with the detection confidence of the key component by the graph neural network.

[0096] Specifically, such as Figure 6 As shown, the weights α and β are determined through the following steps:

[0097] S431. Use deep convolutional networks and graph neural networks respectively to detect key components in normal train images. Record the detection confidence of each key component under the two networks, and then obtain the average detection confidence v of each key component for each network. t-1 ;

[0098] S432. Obtain the actual detection confidence θ of key components in the actual detection process of the two networks. t Then, the target confidence level v of the two networks is calculated using the following formulas. t :

[0099] v t =av t-1 +(1-a)θ t

[0100] Where a is v t-1 The weights are set and a is set to 0.9;

[0101] S433. Calculate the weights α and β based on the proportion of the target confidence scores of the two networks. For example, the target confidence score v using the deep convolutional network can be calculated using the above formula. t1 The target confidence v is calculated using a graph neural network. t2 Finally, using v t1 / (vt1 +v t2 α can be obtained from this, and at this time, β = 1 - α.

[0102] Specifically, such as Figure 7 As shown, the method for detecting appearance defects in maglev trains also includes the following steps:

[0103] S6. Finally, analyze the abnormal image information, save the detected abnormal image information as disease information format to the database, and upload it to the host computer interface, while generating alarm log information.

[0104] like Figure 8 As shown, another preferred embodiment of this application also provides a magnetic levitation train appearance defect detection device, including:

[0105] The module for obtaining the feature description vector set of normal train images is used to extract feature vectors from normal train images using a deep convolutional network and a graph neural network, respectively, to obtain the feature description vector set of normal train images and to perform normal fitting. The feature description vector set includes semantic feature vectors extracted by the deep convolutional network and regional structure feature vectors extracted by the graph neural network.

[0106] The critical component detection module is used to detect critical components after segmenting the actual acquired images of the train under test using an adaptive sliding window.

[0107] The missing area image extraction module is used to extract the area where the missing key components are located and obtain the missing area image if the number of detected key components is less than that in the normal train image.

[0108] The module for obtaining the feature description vector set of the missing region image is used to weight the semantic feature vector and the region structure feature vector obtained after processing the extracted missing region image using a deep convolutional network and a graph neural network, respectively, to obtain the feature description vector set of the missing region image and perform normal fitting.

[0109] The defect confirmation module is used to compare the feature description vector set of the missing region image with the feature description vector set of the normal train image, and to determine whether there is an anomaly in the missing region image by calculating the Gaussian density estimation function.

[0110] like Figure 9 As shown, the maglev train appearance defect detection device also includes:

[0111] The anomaly statistics and alarm module is used to finally collect anomaly image information, save the detected anomaly image information as disease information format to be stored in the database, and upload it to the host computer interface, while generating alarm log information.

[0112] like Figure 10 As shown, a preferred embodiment of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the maglev train appearance defect detection method in the above embodiments.

[0113] like Figure 11 As shown, a preferred embodiment of this application also provides a computer device, which may be a terminal or a liveness detection server, and its internal structure diagram may be as follows. Figure 11 As shown. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with other external computer devices via a network connection. When the computer program is executed by the processor, it implements the steps of the aforementioned method for detecting external defects in maglev trains.

[0114] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0115] A preferred embodiment of this application also provides a storage medium, the storage medium including a stored program, which, when the program is executed, controls the device where the storage medium is located to perform the steps of the maglev train appearance defect detection method in the above embodiments.

[0116] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0117] If the functions described in this embodiment are implemented as software functional units and sold or used as independent products, they can be stored in one or more computing device-readable storage media. Based on this understanding, the parts of this application's embodiments that contribute to the prior art or the technical solutions can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computing device (which may be a personal computer, server, mobile computing device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage media include: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0118] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0119] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0120] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0121] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0122] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0123] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A method for detecting appearance defects in maglev trains, characterized in that, Including the following steps: Deep convolutional networks and graph neural networks are used to extract feature vectors from normal train images to obtain a set of feature description vectors for normal train images and then perform normal fitting. The set of feature description vectors includes semantic feature vectors extracted by deep convolutional networks and regional structure feature vectors extracted by graph neural networks. An adaptive sliding window is used to segment the actual acquired images of the train under test and then detect key components. If the number of key components detected is less than that in a normal train image, the area where the missing key components are located is extracted to obtain an image of the missing area. The semantic feature vector and regional structure feature vector obtained after processing the extracted missing region image using deep convolutional network and graph neural network are weighted to obtain the feature description vector set of the missing region image and then normalized. The feature description vector set of the missing region image is compared with the feature description vector set of the normal train image, and the presence of anomalies in the missing region image is determined by calculating using the Gaussian density estimation function.

2. The method for detecting appearance defects in maglev trains according to claim 1, characterized in that, The method involves using deep convolutional networks and graph neural networks to extract feature vectors from normal train images to obtain a set of feature description vectors for the normal train images, and then performing normal fitting. The set of feature description vectors includes semantic feature vectors extracted by the deep convolutional network and regional structure feature vectors extracted by the graph neural network. Specifically, the method includes the following steps: Images of normal trains are captured using a camera, and after speed correction, undistorted images of normal trains are obtained. A fixed-size window was used to slide and segment the acquired normal train images into several image sequences, and key components were labeled. Deep convolutional networks and graph neural networks are used to extract feature vectors from the segmented image sequences, respectively, to obtain semantic feature vectors and regional structure feature vectors of key components, forming a set of feature description vectors for normal train images. The feature description vector set of the extracted normal train images is fitted using a normal distribution.

3. The method for detecting appearance defects in maglev trains according to claim 1, characterized in that, The process of segmenting the acquired train image using an adaptive sliding window and then detecting key components includes the following steps: Images of the train under test are captured using a camera, and after speed correction, undistorted images of the train under test are obtained. The target detection software is used to segment the image of the train under test into an image sequence and perform key component detection in a window of a set size; If no key components are detected in the image, the sliding segmentation will continue by sliding a distance of a set number of pixels and the detection of key components will continue. If a critical component is detected, the image is segmented by sliding the window to the next target position using the coordinates of the critical component as the window, and the detection of critical components continues until the entire image of the train under test has been detected.

4. The method for detecting appearance defects in maglev trains according to claim 3, characterized in that, The window size is between 1300 pixels * 600 pixels and 1600 pixels * 1000 pixels, and the set number of pixels is between 400 and 600.

5. The method for detecting appearance defects in maglev trains according to claim 1, characterized in that, If the number of detected critical components is less than that in a normal train image, the area containing the missing critical components is extracted to obtain a missing area image. This process includes the following steps: Based on the coordinates of the images, the same key components detected in adjacent images are merged to accurately locate the key components in the image of the train under test. If the number of detected critical components is less than that in a normal train image, it is considered that a critical component is suspected to be missing. Then, the location of the missing critical component on the image is located by using the locations of other detected critical components, and the area where the missing critical component is located is extracted to obtain the image of the missing area.

6. The method for detecting appearance defects in maglev trains according to claim 1, characterized in that, The semantic feature vector and region structure feature vector obtained by processing the extracted missing region image using a deep convolutional network and a graph neural network are weighted to obtain a set of feature description vectors for the missing region image, and then normalized. The specific steps include: The semantic feature vectors of key components in the missing region image are obtained by using a deep convolutional network with a kernel size of 3×3. A graph neural network is used to aggregate features of key component locations in a missing region image to obtain the region structure feature vector of the image. The set of feature description vectors h for the missing region image is obtained by weighting the semantic feature vectors and the region structure feature vectors. * And perform normal fit: h * =ah i +βh j In the formula, h i h is a semantic feature vector. j Let be the feature vector of the region structure, and α and β be the corresponding weights. The weight α is positively correlated with the detection confidence of the key component by the deep convolutional network, and the weight β is positively correlated with the detection confidence of the key component by the graph neural network.

7. The method for detecting appearance defects in maglev trains according to claim 6, characterized in that, The weights α and β are determined through the following steps: Key components of normal train images were detected using both deep convolutional networks and graph neural networks. The detection confidence scores of each key component were recorded for both networks, and the average detection confidence score v for each key component was obtained for each network. t-1 ; In the actual detection process of the two networks, the actual detection confidence θ of the key components was obtained for each. t Then, the target confidence level v of the two networks is calculated using the following formulas. t : v t =off t-1 +(1-a)θ t Where a is v t-1 The weights are set and a is set to 0.9; The weights α and β are calculated based on the proportion of target confidence in the two networks.

8. A device for detecting appearance defects in maglev trains, characterized in that, include: The module for obtaining the feature description vector set of normal train images is used for... Deep convolutional networks and graph neural networks are used to extract feature vectors from normal train images to obtain a set of feature description vectors for normal train images and then perform normal fitting. The set of feature description vectors includes semantic feature vectors extracted by deep convolutional networks and regional structure feature vectors extracted by graph neural networks. The critical component detection module is used to detect critical components after segmenting the actual acquired images of the train under test using an adaptive sliding window. The missing area image extraction module is used to extract the area where the missing key components are located and obtain the missing area image if the number of detected key components is less than that in the normal train image. The module for obtaining the feature description vector set of the missing region image is used to weight the semantic feature vector and the region structure feature vector obtained after processing the extracted missing region image using a deep convolutional network and a graph neural network, respectively, to obtain the feature description vector set of the missing region image and perform normal fitting. The defect confirmation module is used to compare the feature description vector set of the missing region image with the feature description vector set of the normal train image, and to determine whether there is an anomaly in the missing region image by calculating the Gaussian density estimation function.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the maglev train appearance defect detection method as described in any one of claims 1 to 7.

10. A storage medium comprising a stored program, characterized in that, When the program is running, it controls the device containing the storage medium to perform the steps of the magnetic levitation train appearance defect detection method as described in any one of claims 1 to 7.

Images