Railway vehicle insulator anomaly detection method and system based on YOLOv8

Abstract

The application provides a track vehicle insulator anomaly detection method and system based on YOLOv8, and relates to the technical field of insulator detection. The method comprises the following steps: obtaining a scanning image of a track vehicle insulator; determining a plurality of primary selection abnormal risk insulators and a confidence distribution map of the primary selection abnormal risk insulators based on the scanning image of the track vehicle insulator by using a YOLOv8 model; determining a plurality of reference detection regions of each primary selection abnormal risk insulator based on the scanning image of the primary selection abnormal risk insulator; and determining a high-risk abnormal point of the track vehicle insulator based on infrared thermal imaging data of the plurality of reference detection regions in the plurality of secondary selection abnormal risk insulators and infrared thermal imaging data of a plurality of supplementary detection regions in the plurality of secondary selection abnormal risk insulators. The method can accurately and efficiently locate the abnormal point position of the track vehicle insulator.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of insulator detection, and particularly relates to a rail vehicle insulator anomaly detection method and system based on YOLOv8. BACKGROUND

[0002] With the development of rail transportation, the safe and stable operation of rail vehicles directly depends on the reliable work of the traction power supply system. As a core insulating component in the traction power supply system, the state integrity of the rail vehicle insulator is crucial to avoid safety accidents such as electric leakage and short circuit. With the increase of the running speed and the operation mileage of rail vehicles, the vehicle insulator is easily exposed to complex outdoor environments for a long time, and is prone to abnormal problems such as surface damage, internal cracks, and excessive dirt due to factors such as vibration, temperature difference, dirt adhesion, and mechanical wear. If these problems are not discovered and handled in time, they may cause power supply interruption and even train stoppage and other serious consequences. The traditional detection method of rail vehicle insulators mainly relies on manual inspection or fixed camera static monitoring. Manual inspection needs to be carried out during vehicle downtime, which is not only low in efficiency and limited in coverage, but also easily affected by factors such as the experience of the inspector, light conditions, and visual angle blind area, making it difficult to accurately identify hidden defects. Although fixed camera monitoring can achieve preliminary automation, the analysis method based on traditional image processing algorithms has poor adaptability to image shaking, strong light and shadow interference, and insulator shape differences in high-speed motion scenes, often resulting in false positives or false negatives. At the same time, the existing detection methods mainly focus on single-dimensional surface state judgment, and lack effective monitoring of internal defects of the insulator, which cannot meet the real-time and accurate detection requirements of rail vehicles.

[0003] Therefore, how to accurately and efficiently locate the abnormal point position of the rail vehicle insulator is a problem to be solved. SUMMARY

[0004] The technical problem solved by the present application is how to accurately and efficiently locate the abnormal point position of the rail vehicle insulator.

[0005] According to a first aspect, the present application provides a YOLOv8-based abnormality detection method for an insulator of a railway vehicle, comprising: obtaining a scanning image of the insulator of the railway vehicle; determining a plurality of primary abnormal risk insulators and a confidence distribution map of the primary abnormal risk insulators based on the scanning image of the insulator of the railway vehicle using a YOLOv8 model; determining a plurality of reference detection regions of each primary abnormal risk insulator based on the scanning image of the primary abnormal risk insulator; obtaining infrared thermal imaging data of the plurality of reference detection regions of each primary abnormal risk insulator; determining a plurality of supplemental detection regions of each primary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each primary abnormal risk insulator, and obtaining infrared thermal imaging data of the plurality of supplemental detection regions of each primary abnormal risk insulator; determining a plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each primary abnormal risk insulator and the infrared thermal imaging data of the plurality of supplemental detection regions of each primary abnormal risk insulator; and determining a high-risk abnormal point of the insulator of the railway vehicle based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselected abnormal risk insulators and the infrared thermal imaging data of the plurality of supplemental detection regions in the plurality of reselected abnormal risk insulators.

[0006] In a possible implementation, the determining of the plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each primary abnormal risk insulator and the infrared thermal imaging data of the plurality of supplemental detection regions of each primary abnormal risk insulator comprises: generating a second confidence distribution map of the primary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each primary abnormal risk insulator and the infrared thermal imaging data of the plurality of supplemental detection regions of each primary abnormal risk insulator; and determining the plurality of reselected abnormal risk insulators based on the confidence distribution map of the primary abnormal risk insulator and the second confidence distribution map of the primary abnormal risk insulator.

[0007] In a possible implementation, the determining the high-risk abnormal points of the rail vehicle insulator based on the infrared thermal imaging data of the plurality of benchmark detection regions in the plurality of rechecked abnormal risk insulators and the infrared thermal imaging data of the plurality of supplemented detection regions in the plurality of rechecked abnormal risk insulators includes: constructing an abnormal risk graph, the abnormal risk graph including a plurality of infrared thermal imaging detection region nodes and edges between the plurality of infrared thermal imaging detection region nodes, a node feature of each infrared thermal imaging detection region node being a position of an infrared thermal imaging detection region, infrared thermal imaging data of the detection region, and an edge between the infrared thermal imaging detection region nodes being a direction and distance between the detection region nodes, a similarity of the infrared thermal imaging data of the detection region; processing the abnormal risk graph based on a graph neural network to determine a plurality of core verification regions of each rechecked abnormal risk insulator; and determining the high-risk abnormal points of the rail vehicle insulator based on ultrasonic flaw detection data of the plurality of core verification regions of each rechecked abnormal risk insulator.

[0008] In a possible implementation, the input of the graph neural network is the abnormal risk graph, and the output of the graph neural network is the plurality of core verification regions of each rechecked abnormal risk insulator.

[0009] According to a second aspect, the present application provides a rail vehicle insulator anomaly detection system based on YOLOv8, including: an acquisition module configured to acquire a scan image of a rail vehicle insulator; a preliminary selection module configured to determine a plurality of preliminary abnormal risk insulators and a confidence distribution map of the preliminary abnormal risk insulators based on the scan image of the rail vehicle insulator using a YOLOv8 model; a benchmark region determination module configured to determine a plurality of benchmark detection regions of each preliminary abnormal risk insulator based on the scan image of the preliminary abnormal risk insulator; a first data acquisition module configured to acquire infrared thermal imaging data of the plurality of benchmark detection regions of each preliminary abnormal risk insulator; a supplemented region determination module configured to determine a plurality of supplemented detection regions of each preliminary abnormal risk insulator based on the infrared thermal imaging data of the plurality of benchmark detection regions of each preliminary abnormal risk insulator and acquire infrared thermal imaging data of the plurality of supplemented detection regions of each preliminary abnormal risk insulator; a rechecking module configured to determine a plurality of rechecked abnormal risk insulators based on the infrared thermal imaging data of the plurality of benchmark detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the plurality of supplemented detection regions of each preliminary abnormal risk insulator; and a high-risk point determination module configured to determine high-risk abnormal points of the rail vehicle insulator based on the infrared thermal imaging data of the plurality of benchmark detection regions in the plurality of rechecked abnormal risk insulators and the infrared thermal imaging data of the plurality of supplemented detection regions in the plurality of rechecked abnormal risk insulators.

[0010] In a possible implementation, the re-checking module is further configured to: generate a second confidence distribution map of the primary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection areas of each primary abnormal risk insulator and the infrared thermal imaging data of the plurality of supplementary detection areas of each primary abnormal risk insulator; and determine the plurality of re-checking abnormal risk insulators based on the confidence distribution map of the primary abnormal risk insulator and the second confidence distribution map of the primary abnormal risk insulator.

[0011] In a possible implementation, the high-risk point determination module is further configured to: construct an abnormal risk graph, the abnormal risk graph including a plurality of infrared thermal imaging detection area nodes and edges between the plurality of infrared thermal imaging detection area nodes, a node feature of each infrared thermal imaging detection area node being a location of an infrared thermal imaging detection area, infrared thermal imaging data of the detection area, and an edge between the infrared thermal imaging detection area nodes being a direction and distance between the detection area nodes, a similarity of the infrared thermal imaging data of the detection areas; process the abnormal risk graph based on a graph neural network to determine a plurality of core verification areas of each re-checking abnormal risk insulator; and determine the high-risk abnormal points of the rail vehicle insulator based on ultrasonic flaw detection data of the plurality of core verification areas of each re-checking abnormal risk insulator.

[0012] In a possible implementation, an input of the graph neural network is the abnormal risk graph, and an output of the graph neural network is the plurality of core verification areas of each re-checking abnormal risk insulator.

[0013] According to a third aspect, embodiments of the present application provide an electronic device, comprising: a processor; a memory; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement a method as previously described, the method comprising: obtaining a scanning image of a rail vehicle insulator; determining a plurality of primary abnormal risk insulators, a confidence distribution map of the primary abnormal risk insulators based on the scanning image of the rail vehicle insulator using a YOLOv8 model; determining a plurality of reference detection regions of each of the primary abnormal risk insulators based on the scanning image of the primary abnormal risk insulators; obtaining infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators; determining a plurality of supplemental detection regions of each of the primary abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators, and obtaining infrared thermal imaging data of the plurality of supplemental detection regions of each of the primary abnormal risk insulators; determining a plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators, and the infrared thermal imaging data of the plurality of supplemental detection regions of each of the primary abnormal risk insulators; and determining a high-risk abnormal point of the rail vehicle insulator based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselected abnormal risk insulators, and the infrared thermal imaging data of the plurality of supplemental detection regions in the plurality of reselected abnormal risk insulators.

[0014] According to a fourth aspect, embodiments of the present application provide a computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements a YOLOv8-based rail vehicle insulator anomaly detection method as previously described, the method comprising: obtaining a scanning image of a rail vehicle insulator; determining a plurality of primary abnormal risk insulators, a confidence distribution map of the primary abnormal risk insulators based on the scanning image of the rail vehicle insulator using a YOLOv8 model; determining a plurality of reference detection regions of each of the primary abnormal risk insulators based on the scanning image of the primary abnormal risk insulators; obtaining infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators; determining a plurality of supplemental detection regions of each of the primary abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators, and obtaining infrared thermal imaging data of the plurality of supplemental detection regions of each of the primary abnormal risk insulators; determining a plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators, and the infrared thermal imaging data of the plurality of supplemental detection regions of each of the primary abnormal risk insulators; and determining a high-risk abnormal point of the rail vehicle insulator based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselected abnormal risk insulators, and the infrared thermal imaging data of the plurality of supplemental detection regions in the plurality of reselected abnormal risk insulators.

[0015] The application provides a track vehicle insulator anomaly detection method and system based on YOLOv8, which comprises the following steps: acquiring a scanning image of a track vehicle insulator; determining a plurality of primary selection anomaly risk insulators and a confidence distribution map of the primary selection anomaly risk insulators based on the scanning image of the track vehicle insulator by using a YOLOv8 model; determining a plurality of reference detection regions of each primary selection anomaly risk insulator based on the scanning image of the primary selection anomaly risk insulator; acquiring infrared thermal imaging data of the plurality of reference detection regions of each primary selection anomaly risk insulator; determining a plurality of supplementary detection regions of each primary selection anomaly risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each primary selection anomaly risk insulator and acquiring infrared thermal imaging data of the plurality of supplementary detection regions of each primary selection anomaly risk insulator; determining a plurality of reselection anomaly risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each primary selection anomaly risk insulator and the infrared thermal imaging data of the plurality of supplementary detection regions of each primary selection anomaly risk insulator; and determining a high-risk anomaly point of the track vehicle insulator based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselection anomaly risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions in the plurality of reselection anomaly risk insulators, so that the anomaly point position of the track vehicle insulator can be accurately and efficiently located. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a track vehicle insulator anomaly detection method based on YOLOv8 is provided for the embodiments of the application.

[0017] Figure 2 A schematic diagram of a track vehicle insulator is provided for the embodiments of the application.

[0018] Figure 3 A flowchart of determining a plurality of reselection anomaly risk insulators is provided for the embodiments of the application.

[0019] Figure 4 A flowchart of determining a high-risk anomaly point of a track vehicle insulator is provided for the embodiments of the application.

[0020] Figure 5 A schematic diagram of a track vehicle insulator anomaly detection system based on YOLOv8 is provided for the embodiments of the application. DETAILED DESCRIPTION

[0021] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0022] In this embodiment of the invention, the following are provided: Figure 1 The YOLOv8-based method for detecting insulator anomalies in railway vehicles includes steps S1 to S7:

[0023] Step S1: Obtain scanned images of the insulators of the rail vehicle.

[0024] A rail vehicle insulator is an insulating component installed in the power supply system of a rail vehicle. Rail vehicle insulators are made of materials with excellent insulation properties. Figure 2 This is a schematic diagram of a rail vehicle insulator according to an embodiment of the present invention.

[0025] The scanned image of the insulator of a rail vehicle is a high-definition image focused on the insulator of the rail vehicle body, obtained by scanning the insulator on the rail vehicle using a high-precision industrial high-definition camera imaging device.

[0026] Step S2: Based on the scanned images of the rail vehicle insulators, the YOLOv8 model is used to determine multiple initially selected abnormal risk insulators and the confidence distribution map of the initially selected abnormal risk insulators.

[0027] The YOLOv8 model is a real-time object detection and multi-task computer vision model. It employs an anchor-free design, and its overall architecture comprises a backbone network, a neck network, and a head network. The YOLOv8 model incorporates multi-scale feature extraction and adaptive anchor box generation, and supports both grayscale and color image inputs, demonstrating strong recognition capabilities for small targets. The input to the YOLOv8 model is a scanned image of the rail vehicle insulator, and the output consists of multiple initially selected anomalous risk insulators and their confidence distribution maps.

[0028] The plurality of primary abnormal risk insulators is a plurality of sets of rail vehicle insulators with abnormality possibilities determined by the YOLOv8 model. Each primary abnormal risk insulator corresponds to a specific insulator target in the scan image.

[0029] The confidence distribution map of the primary abnormal risk insulators is generated based on the analysis results of the scan image of the rail vehicle insulator by the YOLOv8 model, and presents the distribution map of the possibility of the existence of faults at each insulator position in the plurality of primary abnormal risk insulators in a visual form.

[0030] The confidence distribution map of the primary abnormal risk insulators visually presents the possibility of the existence of abnormal risks at each position of all the primary abnormal risk insulators through color gradient and numerical labeling.

[0031] The scan image of the rail vehicle insulator contains visual information such as the appearance details, surface state, and structural features of the insulator, which can directly reflect whether the insulator has potential abnormal signs such as surface damage, stain coverage, and structural deformation. Different types of abnormalities will present unique visual features, such as the irregularity of the edges of the damaged part and the gray difference of the stain.

[0032] The YOLOv8 model can extract features of the scan image of the rail vehicle insulator through the Backbone network. The cross-stage local network structure can retain key features while reducing computational complexity through residual connection and cross-stage partial connection, so that the image can be down-sampled multiple times to obtain feature maps of different scales. The shallow feature map can focus on the edges and textures of the insulator to capture traces of minor surface damage. The deep feature map can integrate global information to identify overall abnormalities such as structural deformation. Then, the PAFPN in the Neck part can fuse feature maps of different scales through top-down and bottom-up paths to solve the problem of small target feature loss, so as to ensure that the model can detect abnormalities of insulators of different sizes. Subsequently, the decoupled head in the Head part can separate the classification task and the regression task, and calculate the abnormal class probability and the bounding box coordinates of each insulator target, respectively. Then, in combination with the preset confidence threshold, insulators with a confidence higher than the threshold are selected as the plurality of primary abnormal risk insulators. At the same time, the model can associate the confidence value of each primary abnormal risk insulator with the corresponding position, and then generate the confidence distribution map of the primary abnormal risk insulators, so as to intuitively display the spatial distribution of the abnormal risks of all the primary abnormal risk insulators through the confidence distribution map of the primary abnormal risk insulators.

[0033] In step S3, based on the scan image of the primary abnormal risk insulator, a plurality of reference detection regions of each primary abnormal risk insulator are determined.

[0034] The scan image of the primary abnormal risk insulator is a local area image corresponding to a part of the primary abnormal risk insulator in the scan image of the rail vehicle insulator.

[0035] In some embodiments, a first region determination model can be used to determine the multiple baseline detection regions of each primary abnormal risk insulator. The first region determination model is a deep neural network model. The input of the first region determination model is the scan image of the primary abnormal risk insulator, and the output of the first region determination model is the multiple baseline detection regions of each primary abnormal risk insulator.

[0036] The deep neural network model includes a deep neural network (DNN). The deep neural network is a neural network architecture inspired by the connection mechanism of biological brain neurons and contains multiple hidden layers. The deep neural network model can receive data through the input layer and then perform feature extraction and conversion by artificial neurons in multiple hidden layers step by step, and finally output the prediction result through the output layer. The neurons of each layer of the deep neural network model transmit information through adjustable weight parameters and continuously optimize the parameters through the backpropagation algorithm, which can achieve accurate learning and fitting of data rules. The deep neural network can be applied to complex tasks such as image recognition and speech processing.

[0037] The multiple baseline detection regions of each primary abnormal risk insulator are a set of local areas that need to be detected in detail, which are drawn by analyzing the scan image of the primary abnormal risk insulator through the first region determination model. Each baseline detection region corresponds to a specific location on the primary abnormal risk insulator where an anomaly may exist, and has a clear boundary range.

[0038] The scan image of the primary abnormal risk insulator can present the detailed features of the primary abnormal risk insulator, including surface contour, gray scale distribution, structural morphology, and other accurate information. These information can reflect the potential abnormal hidden dangers in the local area of the insulator, such as contour mutation and gray scale abnormal area, thereby providing accurate data basis for the model to locate the baseline detection region.

[0039] The deep neural network can capture local information such as contour edges, gray scale changes, etc. in the scan image of the primary abnormal risk insulator through a feature extraction mechanism, while compressing the data dimension and retaining key features in the process. Then through a feature integration mechanism, the association between local features and abnormal risks is established, for example, the contour mutation feature is associated with the structural damage risk, and the gray scale unevenness feature is associated with the internal defect risk. The model can also combine the standard structure template of the insulator to calculate the difference degree of each region in the scan image and the standard template, and the region with a higher difference degree is determined to have a higher abnormal risk. Finally, the model can determine a plurality of local regions with clear boundaries according to the feature association result and the difference degree ranking, which are the plurality of reference detection regions of each primary abnormal risk insulator, to ensure that the detection can focus on the high-risk local position.

[0040] In some embodiments, the determining, based on the scan image of the primary abnormal risk insulator, the plurality of reference detection regions of each primary abnormal risk insulator comprises steps S31-S33:

[0041] Step S31, determining the insulator surface contour abnormal position, gray scale distribution difference sequence, and structural deformation region based on the scan image of the primary abnormal risk insulator.

[0042] In some embodiments, a convolutional neural network can be used to determine the insulator surface contour abnormal position and the gray scale distribution difference sequence.

[0043] The convolutional neural network model includes a convolutional neural network (CNN). The convolutional neural network is a deep learning model that can process data with a grid structure. The convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional neural network model can perform sliding calculation on the input data through multiple convolution kernels in the convolutional layer to extract local features, the pooling layer can further compress the feature data to retain key information, and the fully connected layer can integrate the extracted multi-level features to achieve classification and detection tasks.

[0044] The insulator surface contour abnormal position is the specific position where the contour deviates from the standard insulator contour determined by the convolutional neural network after contour analysis of the scan image of the primary abnormal risk insulator.

[0045] The insulator surface contour abnormal position can be manifested as contour breakage, protrusion or depression.

[0046] The gray scale distribution difference sequence is the ordered data reflecting the deviation of the gray scale values at different positions in the image from the standard gray scale distribution generated by the convolutional neural network after gray scale value analysis of the scan image of the primary abnormal risk insulator.

[0047] The gray scale distribution difference sequence is arranged in the order of spatial positions of the image, and each data point represents a gray scale deviation degree of a corresponding position. The gray scale distribution difference sequence can be used to identify abnormal gray scale areas caused by surface stains, internal defects, and the like of the insulator.

[0048] The convolutional neural network can perform gray scale processing on the scan image of the preliminary abnormal risk insulator to convert the colored scan image of the preliminary abnormal risk insulator into a gray scale image to simplify data calculation. Subsequently, the profile feature is extracted through the first group of convolutional layers and pooling layers, the convolution kernel adopts an edge detection operator, and can respond to areas with sudden changes in pixel values in the image, thereby obtaining a profile feature map. Then, the profile feature map is compared with a standard insulator profile template, and a profile deviation value is calculated. The position with a deviation value exceeding a threshold value is an abnormal profile position on the surface of the insulator. Then, the model can perform multi-scale feature extraction on the gray scale image through the second group of convolutional layers. Then, the difference between the gray scale value of each pixel point and the average gray scale value of the surrounding pixels and the standard gray scale is calculated, and these difference values are arranged in the order of rows and columns of the image to generate a gray scale distribution difference sequence. The position with a larger absolute value in the sequence corresponds to an abnormal gray scale area.

[0049] In step S32, a plurality of risk target areas, an abnormal risk level of each risk target area, a feature correlation degree between the risk target areas, and a priority ranking of the risk target areas are determined based on the abnormal profile position on the surface of the insulator and the gray scale distribution difference sequence.

[0050] In some embodiments, a deep neural network can be used to determine the plurality of risk target areas, the abnormal risk level of each risk target area, the feature correlation degree between the risk target areas, and the priority ranking of the risk target areas.

[0051] The plurality of risk target areas are small-scale spaces on the preliminary abnormal risk insulator determined by the deep neural network, which are significantly abnormal and have accurately locatable boundaries, and are used for subsequent targeted defect verification and risk level determination.

[0052] The abnormal risk level of each risk target area is a possibility level of the existence of a real abnormality in the region determined by the deep neural network after evaluating the features of each risk target area. The abnormal risk level can be divided into three levels: high, medium, and low.

[0053] The feature correlation degree between the risk target areas is a correlation degree of abnormal features between different risk target areas calculated by the deep neural network.

[0054] The higher the feature correlation degree between the risk target areas, the more likely it is that the abnormalities of the two risk target areas are caused by the same reason, such as multiple profile abnormal areas caused by the same external force impact.

[0055] The risk target region priority ranking is the ranking of multiple risk target regions by a deep neural network in combination with the abnormal risk level of each risk target region and the feature correlation degree between risk target regions. In the ranking result, the region with high risk and low correlation degree with other regions has a higher priority, facilitating subsequent priority detection.

[0056] The deep neural network can input the coordinate information of the insulator surface profile abnormal position, the quantized data of the gray scale distribution difference sequence, and the features of the quantized data into the fusion layer, then focus on the regions with the three overlapping or strong correlation through the attention mechanism, and further generate multiple risk target regions through boundary regression. The deep neural network uses a fully connected layer to perform weighted calculation on the features of each region and outputs the abnormal risk level of each risk target region in combination with a pre-trained abnormal sample library, and obtains the feature correlation degree between risk target regions by comparing the feature patterns of different regions through a feature similarity calculation module. The decision layer can comprehensively consider the risk level, correlation degree, and area of the region, and then output the risk target region priority ranking through a normalization index layer.

[0057] Step S33, determining multiple reference detection regions of each preliminary abnormal risk insulator based on the multiple risk target regions, the abnormal risk level of each risk target region, the feature correlation degree between risk target regions, and the risk target region priority ranking.

[0058] In some embodiments, a deep neural network can be used to determine the multiple reference detection regions of each preliminary abnormal risk insulator.

[0059] The deep neural network can select the risk target regions with high priority in sequence according to the risk target region priority ranking for analysis. For multiple risk target regions with high feature correlation degree, the model can determine whether they belong to the correlation regions caused by the same abnormal source, and if so, merge them into a comprehensive region to avoid repeated detection. If they belong to different sources, the independent regions are retained. Then the model can combine the abnormal risk level of each region to optimize the boundary of the merged region, and appropriately expand the boundary of the high-risk level region to cover possible extended abnormalities, and accurately focus on the core abnormal points for the medium and low-risk level regions. Subsequently, the model can check whether the optimized regions overlap, and adjust the boundary of the overlapping regions to ensure that each region is independent and completely covers the abnormal features. Finally, the model can filter out multiple regions that meet the requirements according to the preset detection quantity threshold and region size standard, and these regions are the multiple reference detection regions of each preliminary abnormal risk insulator.

[0060] Step S4, acquiring infrared thermal imaging data of the multiple reference detection regions of each preliminary abnormal risk insulator.

[0061] The infrared thermal imaging data of the plurality of reference detection areas of each preliminary abnormal risk insulator is data obtained by detecting the plurality of reference detection areas of each preliminary abnormal risk insulator by the infrared thermal imaging device.

[0062] The infrared thermal imaging data exists in the form of temperature values and thermal image pixel values, and can be used to record the temperature distribution of the reference detection area. Insulation abnormalities can be accompanied by abnormal temperature rise, so the infrared thermal imaging data can reflect whether there is a problem such as a decrease in insulation performance inside the insulator.

[0063] Step S5, determining a plurality of supplemental detection areas of each preliminary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection areas of each preliminary abnormal risk insulator, and obtaining infrared thermal imaging data of the plurality of supplemental detection areas of each preliminary abnormal risk insulator.

[0064] In some embodiments, a second region determination model can be used to determine the plurality of supplemental detection areas of each preliminary abnormal risk insulator. The second region determination model is a deep neural network model. The input of the second region determination model is the infrared thermal imaging data of the plurality of reference detection areas of each preliminary abnormal risk insulator, and the output of the second region determination model is the plurality of supplemental detection areas of each preliminary abnormal risk insulator.

[0065] The plurality of supplemental detection areas of each preliminary abnormal risk insulator is a set of local areas that need to be detected for each preliminary abnormal risk insulator, which is supplemented by the second region determination model after analyzing the infrared thermal imaging data of the reference detection area.

[0066] The supplemental detection area is a position that is not covered by the reference detection area but has temperature abnormality signs, and can be used to make up for possible omissions of the reference detection.

[0067] The infrared thermal imaging data of the plurality of supplemental detection areas of each preliminary abnormal risk insulator is data obtained by detecting the plurality of supplemental detection areas of each preliminary abnormal risk insulator by the infrared thermal imaging device.

[0068] The infrared thermal imaging data of the plurality of reference detection areas of each preliminary abnormal risk insulator provides thermal characteristic samples of the reference area, and determines abnormal thermal characteristics by analyzing information such as temperature values, temperature distribution gradients, and abnormal temperature rise ranges in the data, including local temperature rise exceeding a normal threshold or temperature gradient mutation. The infrared thermal imaging data of the plurality of reference detection areas of each preliminary abnormal risk insulator can expose the detection blind area of the reference area, and some small temperature abnormalities may not be completely captured in the reference detection. The model can find these potential abnormal areas by deep analysis of the infrared thermal imaging data of the reference detection area.

[0069] The deep neural network can extract spatial features of temperature distribution, such as the shape of hot spots, the rate of change of temperature gradient, etc., from the infrared thermal imaging data of multiple baseline detection regions of each preliminary abnormal risk insulator, and then analyze the potential extension trend of the temperature features through the cycle layer, and determine whether there are signs of hot spots spreading outside the baseline region. The model can also calculate the temperature prediction value of the surrounding area of the baseline region, and compare the prediction value with the actual environmental temperature. If the prediction value is significantly higher than the environmental temperature, it is marked as a potential abnormal region. Then the model can perform cluster analysis on the potential abnormal region, and then merge the regions with close distance and similar temperature abnormal features into a single supplementary detection region, and determine the boundary range of each region. Then the model can preliminarily evaluate the abnormality possibility of the supplementary detection region, eliminate the misjudged region, and finally determine multiple supplementary detection regions of each preliminary abnormal risk insulator.

[0070] Step S6, determining multiple reselected abnormal risk insulators based on the infrared thermal imaging data of the multiple baseline detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the multiple supplementary detection regions of each preliminary abnormal risk insulator.

[0071] In some embodiments, Figure 3 A flowchart for determining multiple reselected abnormal risk insulators is provided for embodiments of the present application. The determination of multiple reselected abnormal risk insulators includes steps S61-S62:

[0072] Step S61, generating a second confidence distribution map of the preliminary abnormal risk insulator based on the infrared thermal imaging data of the multiple baseline detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the multiple supplementary detection regions of each preliminary abnormal risk insulator.

[0073] In some embodiments, a risk prediction model can be used to generate the second confidence distribution map of the preliminary abnormal risk insulator. The risk prediction model is a Transformer model. The input of the risk prediction model is the infrared thermal imaging data of the multiple baseline detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the multiple supplementary detection regions of each preliminary abnormal risk insulator. The output of the risk prediction model is the second confidence distribution map of the preliminary abnormal risk insulator.

[0074] The Transformer model is a deep learning model based on a self-attention mechanism. The core architecture of the Transformer model includes an encoder and a decoder. The encoder is composed of multiple layers of multi-head self-attention mechanisms and feedforward neural networks. The encoder can extract deep features from input sequences and can simultaneously focus on the correlation information of different positions in the input data through the self-attention mechanism. The encoder can efficiently capture long-range dependencies. The decoder adds an attention mechanism to the encoder. The decoder can generate target sequences based on the output of the encoder. The Transformer model uses parallel computing and does not rely on a loop structure. The Transformer model has high efficiency and accuracy when processing time-series or spatial data.

[0075] The second confidence distribution map of the primary abnormal risk insulator is a visual analysis image generated by analyzing and calculating the infrared thermal imaging data of the multiple baseline detection regions of each primary abnormal risk insulator and the infrared thermal imaging data of the multiple supplemental detection regions of each primary abnormal risk insulator through the risk prediction model, and is used to accurately present the failure possibility of each position of all primary abnormal risk insulators.

[0076] The second confidence distribution map of the primary abnormal risk insulator is different from the confidence distribution map of the primary abnormal risk insulator in that it is based on the temperature characteristics measured by the infrared thermal imaging data, and can quantify the failure possibility distribution of each position of all primary abnormal risk insulators from the perspective of the operating thermal state of the insulator.

[0077] The infrared thermal imaging data of the multiple baseline detection regions and the multiple supplemental detection regions of each primary abnormal risk insulator together constitute the complete temperature distribution data set of the primary abnormal risk insulator. The temperature distribution data set can reflect the insulation performance state of the insulator from the perspective of thermal science. Temperature abnormalities are directly related to internal defects, surface flashover, and other failures of the insulator. The temperature characteristics of different regions can accurately indicate the potential location and severity of the failure.

[0078] The transformer model can convert the infrared thermal imaging data of multiple baseline detection regions and the infrared thermal imaging data of multiple supplementary detection regions of each preliminary abnormal risk insulator into sequence data, where each data point contains region coordinates, temperature values, temperature gradients, and other features. Subsequently, the encoder part can process the sequence data through a multi-head self-attention mechanism, and calculate the attention weight of each data point with other data points, while focusing on the correlation between temperature abnormal data points, such as the distance and temperature difference between multiple hot spots. A feedforward neural network can perform nonlinear transformation on the features output by the attention mechanism to further extract the deep correlation between temperature features and fault risk. Then the model can map the features output by the encoder into the spatial coordinate system of the insulator, so as to assign each coordinate point with an abnormal risk probability value based on temperature features. After that, the probability values of undetected regions are filled in through an interpolation algorithm to ensure the integrity of the distribution map. Finally, the model can convert the probability values into visual color and numerical scale, and generate a second confidence distribution map of the preliminary abnormal risk insulator.

[0079] In step S62, a plurality of reselected abnormal risk insulators are determined based on the confidence distribution map of the preliminary abnormal risk insulator and the second confidence distribution map of the preliminary abnormal risk insulator.

[0080] In some embodiments, a risk analysis model can be used to determine the plurality of reselected abnormal risk insulators. The risk analysis model is a deep neural network. The input of the risk analysis model is the confidence distribution map of the preliminary abnormal risk insulator and the second confidence distribution map of the preliminary abnormal risk insulator. The output of the risk analysis model is the plurality of reselected abnormal risk insulators.

[0081] The plurality of reselected abnormal risk insulators are a set of insulators with higher abnormal risk possibility selected from the plurality of preliminary abnormal risk insulators through fusion analysis of the confidence distribution map of the preliminary abnormal risk insulator and the second confidence distribution map of the preliminary abnormal risk insulator by the risk analysis model.

[0082] The confidence distribution map of the preliminary abnormal risk insulator is generated based on the visual features of the scanned image to reflect the abnormal possibility at the appearance level of the insulator. The second confidence distribution map of the preliminary abnormal risk insulator is generated based on the infrared thermal imaging data, thereby reflecting the abnormal possibility at the thermal level of the insulator. The confidence distribution map and the second confidence distribution map of the preliminary abnormal risk insulator provide abnormal risk information from different dimensions, and the consistency and complementary performance of the two significantly improve the accuracy of abnormal judgment, thereby providing double data support for the model to select reselected abnormal risk insulators.

[0083] The deep neural network can convert the confidence values of the confidence distribution map of the primary abnormal risk insulator and the second confidence distribution map into a unified scale between 0 and 1, then extract the spatial features of the two distribution maps, such as the shape and distribution density of the high confidence area, and then perform point-by-point fusion of the two features through a fusion layer to calculate the comprehensive confidence value of each position, which is the weighted average of the confidence values of the corresponding positions of the confidence distribution map of the primary abnormal risk insulator and the second confidence distribution map of the primary abnormal risk insulator. Then the model can analyze the comprehensive confidence distribution corresponding to each primary abnormal risk insulator, calculate the average comprehensive confidence and the high confidence area ratio of the insulator. Then set double screening thresholds, namely the average comprehensive confidence threshold and the high confidence area ratio threshold, only the primary abnormal risk insulator that meets the requirements of both thresholds is retained, thereby serving as a reselected abnormal risk insulator.

[0084] Step S7, determining the high-risk abnormal point of the insulator of the railway vehicle based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselected abnormal risk insulators and the infrared thermal imaging data of the plurality of supplemented detection regions in the plurality of reselected abnormal risk insulators.

[0085] In some embodiments, Figure 4 A flowchart for determining a high-risk abnormal point of an insulator of a railway vehicle is provided for an embodiment of the present application, and the determination of the high-risk abnormal point of the insulator of the railway vehicle includes steps S71-S73.

[0086] Step S71, constructing an abnormal risk graph, the abnormal risk graph including a plurality of infrared thermal imaging detection region nodes and edges between the plurality of infrared thermal imaging detection region nodes, the node features of each infrared thermal imaging detection region node being the position of the infrared thermal imaging detection region and the infrared thermal imaging data of the detection region, and the edges between the infrared thermal imaging detection region nodes being the orientation and distance between the detection region nodes and the similarity of the infrared thermal imaging data of the detection region.

[0087] The abnormal risk graph is a data structure for associating infrared thermal imaging detection region parameters and insulator abnormal features, and the abnormal risk graph is composed of two parts of nodes (vertices) and edges (edges). The abnormal risk graph includes a plurality of infrared thermal imaging detection region nodes, and the node features of each infrared thermal imaging detection region node are the position of the infrared thermal imaging detection region and the infrared thermal imaging data of the detection region. The edges between the nodes represent the association information, and the edges are the orientation and distance between the detection region nodes and the similarity of the infrared thermal imaging data of the detection region. The spatial correlation and thermal feature correlation of different detection regions can be established through the edges, and then the abnormal risk correlation network of the whole insulator can be constructed.

[0088] In some embodiments, a deep neural network can be used to determine the similarity of the infrared thermography data of the detection regions.

[0089] At step S72, the graph neural network is used to process the abnormal risk graph to determine a plurality of core verification regions of each suspected abnormal risk insulator.

[0090] The input of the graph neural network is the abnormal risk graph, and the output of the graph neural network is the plurality of core verification regions of each suspected abnormal risk insulator.

[0091] A graph neural network (GNN) is a type of deep learning model that can be used to process graphs. Graph neural networks can learn node, edge, and global structure information in a graph to achieve feature representation and prediction tasks in nodes, edges, and entire graphs.

[0092] The plurality of core verification regions of each suspected abnormal risk insulator is a set of core local regions most likely to have real abnormalities on each suspected abnormal risk insulator, determined by analyzing the abnormal risk graph using a graph neural network.

[0093] By constructing an abnormal risk graph, the association network between nodes of each infrared thermography detection region can be clearly reflected. This relationship information is very important for determining core verification regions, because the abnormal risk of a single insulator often has an associated impact on other insulators. By using the properties and characteristics of each suspected abnormal risk insulator as node features and the association strength between insulators as edge features, data information can be more fully utilized. This helps the model better understand the state of each insulator and the interaction mechanism between them, thereby improving the accuracy of core verification region identification. Processing abnormal risk graph data based on a graph neural network model can effectively learn the complex relationships and information transmission between nodes, thereby more accurately identifying the regions of each insulator that need to be focused on. Compared to traditional analysis methods, graph neural networks have better representation and learning capabilities when processing graph data, and can more comprehensively capture the distribution patterns of abnormal risks.

[0094] The graph neural network can accurately locate the core verification regions by analyzing the node and edge features of the abnormal risk graph. After inputting the features of each infrared thermography detection region node and edge features in the graph into the graph neural network, the graph neural network can aggregate neighbor information for each node through a message passing mechanism. The graph neural network can identify nodes with high thermal data abnormality and strong association with surrounding abnormal regions through multi-layer iterative learning, and then score the importance of these nodes. The regions corresponding to the highest scoring nodes are then selected as the core verification regions of each suspected abnormal risk insulator.

[0095] Step S73, based on the ultrasonic flaw detection data of the multiple core verification areas of each suspected abnormal risk insulator, determining the high-risk abnormal point of the insulator of the rail vehicle.

[0096] The ultrasonic flaw detection data of the multiple core verification areas of each suspected abnormal risk insulator is a quantifiable data set formed by emitting ultrasonic waves to the multiple core verification areas through an ultrasonic flaw detection device, receiving and converting the reflection, refraction and other signals generated by structural differences when the ultrasonic waves propagate inside the areas.

[0097] In some embodiments, an abnormality positioning model can be used to determine the high-risk abnormal point of the insulator of the rail vehicle. The abnormality positioning model is a deep neural network. The input of the abnormality positioning model is the ultrasonic flaw detection data of the multiple core verification areas of each suspected abnormal risk insulator, and the output of the abnormality positioning model is the high-risk abnormal point of the insulator of the rail vehicle.

[0098] The high-risk abnormal point of the insulator of the rail vehicle is a specific point on the insulator where a real abnormality exists, which is determined by analyzing the ultrasonic flaw detection data of the multiple core verification areas of each suspected abnormal risk insulator through the abnormality positioning model.

[0099] The high-risk abnormal point corresponds to the defect position inside or on the surface of the insulator, such as cracks, bubbles, damage, etc., which is a high-risk position that needs to be handled first.

[0100] Deep neural networks have strong feature extraction and pattern recognition capabilities. Ultrasonic flaw detection data contains parameters such as sound propagation time, amplitude, frequency, and other parameters that reflect the internal structure of the insulator. These parameters contain subtle features related to defects. Through progressive learning of multiple neurons, deep neural networks can automatically mine deep correlations from basic signals to defect features in ultrasonic flaw detection data. After a large amount of labeled ultrasonic flaw detection data is trained, the deep neural network can accurately remember the ultrasonic signal patterns corresponding to different types of defects. When the ultrasonic flaw detection data of the core verification area is input, the model can quickly match the learned defect patterns. The nonlinear fitting capability of the deep neural network can handle the complex mapping relationship between the ultrasonic flaw detection signal and the defect, and can effectively distinguish between normal structure signals and abnormal defect signals, and thus locate the specific point on the insulator where a high-risk abnormality exists.

[0101] Based on the same inventive concept, Figure 5 The YOLOv8-based rail vehicle insulator anomaly detection system provided by the embodiments of the present application is shown in the schematic diagram, and the YOLOv8-based rail vehicle insulator anomaly detection system comprises:

[0102] The acquisition module 81 is configured to acquire a scanning image of the insulator of the rail vehicle.

[0103] The preliminary selection module 82 is configured to determine a plurality of preliminary abnormal risk insulators and a confidence distribution of the preliminary abnormal risk insulators based on the scanning image of the insulator of the rail vehicle using the YOLOv8 model.

[0104] The reference region determination module 83 is configured to determine a plurality of reference detection regions of each of the preliminary abnormal risk insulators based on the scanning image of the preliminary abnormal risk insulator.

[0105] The first data acquisition module 84 is configured to acquire infrared thermal imaging data of the plurality of reference detection regions of each of the preliminary abnormal risk insulators.

[0106] The supplementary region determination module 85 is configured to determine a plurality of supplementary detection regions of each of the preliminary abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the preliminary abnormal risk insulators, and acquire infrared thermal imaging data of the plurality of supplementary detection regions of each of the preliminary abnormal risk insulators.

[0107] The recheck module 86 is configured to determine a plurality of recheck abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the preliminary abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions of each of the preliminary abnormal risk insulators.

[0108] The high risk point determination module 87 is configured to determine a high risk abnormal point of the insulator of the rail vehicle based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of recheck abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions in the plurality of recheck abnormal risk insulators.

[0109] It should be noted that, in the interest of simplifying the presentation of the disclosure made herein to aid in the understanding of one or more inventive embodiments, various features can be grouped together in a single embodiment, figure or description of an embodiment, in the foregoing description of embodiments of the disclosure. However, this method of disclosure is not meant to imply that the features of the disclosure claimed are more than the features mentioned in the claims. In fact, the features of the embodiments are less than all the features disclosed in the foregoing description of a single embodiment.

[0110] Finally, it should be understood that the embodiments described herein are merely illustrative of the principles of the embodiments of the disclosure. Other variations can also be possible within the scope of the disclosure. Thus, alternative configurations of the embodiments of the disclosure can be considered as consistent with the teachings of the disclosure. Accordingly, the embodiments of the disclosure are not limited to the embodiments explicitly introduced and described in the disclosure.

Claims

1. A method for detecting abnormalities of insulators of rolling stock based on YOLOv8, characterized in that, The method comprises: acquiring a scanning image of an insulator of a railway vehicle; determining a plurality of primary abnormal risk insulators and a confidence distribution map of the primary abnormal risk insulators based on the scanning image of the insulator of the railway vehicle using a YOLOv8 model; determining a plurality of reference detection regions of each of the primary abnormal risk insulators based on the scanning image of the primary abnormal risk insulator; acquiring infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators; determining a plurality of supplementary detection regions of each of the primary abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators and acquiring infrared thermal imaging data of the plurality of supplementary detection regions of each of the primary abnormal risk insulators; determining a plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions of each of the primary abnormal risk insulators; determining a high-risk abnormal point of the insulator of the railway vehicle based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of reselected abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions in the plurality of reselected abnormal risk insulators, the determining comprising: constructing an abnormal risk graph, the abnormal risk graph comprising a plurality of infrared thermal imaging detection region nodes and edges between the plurality of infrared thermal imaging detection region nodes, a node feature of each of the infrared thermal imaging detection region nodes being a position of an infrared thermal imaging detection region, infrared thermal imaging data of the detection region, and the edges between the infrared thermal imaging detection region nodes being a direction and a distance between the detection region nodes, a similarity of the infrared thermal imaging data of the detection region; processing the abnormal risk graph based on a graph neural network to determine a plurality of core verification regions of each of the reselected abnormal risk insulators, an input of the graph neural network being the abnormal risk graph, and an output of the graph neural network being the plurality of core verification regions of each of the reselected abnormal risk insulators; determining the high-risk abnormal point of the insulator of the railway vehicle based on ultrasonic flaw detection data of the plurality of core verification regions of each of the reselected abnormal risk insulators.

2. The YOLOv8-based rolling stock insulator anomaly detection method of claim 1, wherein, The determining the plurality of reselected abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions of each of the primary abnormal risk insulators comprises: generating a second confidence distribution map of the primary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each of the primary abnormal risk insulators and the infrared thermal imaging data of the plurality of supplementary detection regions of each of the primary abnormal risk insulators; determining the plurality of reselected abnormal risk insulators based on the confidence distribution map of the primary abnormal risk insulator and the second confidence distribution map of the primary abnormal risk insulator.

3. A YOLOv8-based rail vehicle insulator anomaly detection system, characterized in that, The method comprises: An acquisition module is configured to acquire a scanning image of a rail vehicle insulator; A preliminary selection module is configured to determine a plurality of preliminary abnormal risk insulators and a confidence distribution map of the preliminary abnormal risk insulators based on the scanning image of the rail vehicle insulator using a YOLOv8 model; A reference region determination module is configured to determine a plurality of reference detection regions of each preliminary abnormal risk insulator based on the scanning image of the preliminary abnormal risk insulator; A first data acquisition module is configured to acquire infrared thermal imaging data of the plurality of reference detection regions of each preliminary abnormal risk insulator; An additional region determination module is configured to determine a plurality of additional detection regions of each preliminary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each preliminary abnormal risk insulator, and acquire infrared thermal imaging data of the plurality of additional detection regions of each preliminary abnormal risk insulator; A recheck module is configured to determine a plurality of recheck abnormal risk insulators based on the infrared thermal imaging data of the plurality of reference detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the plurality of additional detection regions of each preliminary abnormal risk insulator; A high-risk point determination module is configured to determine a high-risk abnormal point of the rail vehicle insulator based on the infrared thermal imaging data of the plurality of reference detection regions in the plurality of recheck abnormal risk insulators and the infrared thermal imaging data of the plurality of additional detection regions in the plurality of recheck abnormal risk insulators, and the high-risk point determination module is further configured to: construct an abnormal risk graph, the abnormal risk graph comprising a plurality of infrared thermal imaging detection region nodes and edges between the plurality of infrared thermal imaging detection region nodes, a node feature of each infrared thermal imaging detection region node being a position of an infrared thermal imaging detection region, infrared thermal imaging data of the detection region, and an edge between infrared thermal imaging detection region nodes being a direction and distance between the detection region nodes, a similarity of the infrared thermal imaging data of the detection region; process the abnormal risk graph based on a graph neural network to determine a plurality of core verification regions of each recheck abnormal risk insulator, an input of the graph neural network being the abnormal risk graph, and an output of the graph neural network being the plurality of core verification regions of each recheck abnormal risk insulator; determine a high-risk abnormal point of the rail vehicle insulator based on ultrasonic flaw detection data of the plurality of core verification regions of each recheck abnormal risk insulator.

4. The YOLOv8-based rolling stock insulator anomaly detection system of claim 3, wherein, The recheck module is further configured to: generate a second confidence distribution map of the preliminary abnormal risk insulator based on the infrared thermal imaging data of the plurality of reference detection regions of each preliminary abnormal risk insulator and the infrared thermal imaging data of the plurality of additional detection regions of each preliminary abnormal risk insulator; determine the plurality of recheck abnormal risk insulators based on the confidence distribution map of the preliminary abnormal risk insulator and the second confidence distribution map of the preliminary abnormal risk insulator.

5. An electronic device, comprising: comprise: a processor; a memory; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the YOLOv8-based rail vehicle insulator anomaly detection method of any one of claims 1 to 2.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the YOLOv8-based rail vehicle insulator anomaly detection method according to any one of claims 1-2.

Images