High-speed rail catenary anti-part falling method and system based on 4C detection

Abstract

This invention relates to the field of high-speed rail operation and maintenance technology, and proposes a method and system for preventing the detachment of high-speed rail catenary components based on 4C detection. The method includes: identifying raw image data through a target detection network model to obtain key component data of the high-speed rail catenary; extracting features from the key component data using a fine-grained detection network model to obtain key component feature data; judging the anti-loosening status of key components based on the key component feature data to obtain anti-loosening detection data; constructing a spatiotemporal feature matrix based on the anti-loosening detection data and historical status data, and calculating the detachment risk coefficient of key components using a risk assessment model; and generating an anti-loosening warning report containing location information, anomaly type, and maintenance suggestions based on the detachment risk coefficient, and sending it to management personnel. This invention achieves high-precision monitoring of the loosening status of key anti-loosening components of the high-speed rail catenary, enabling early warning of detachment risks.

Description

Technical Field

[0001] This invention relates to the field of high-speed rail operation and maintenance technology, and in particular to a method and system for preventing components from falling off high-speed rail catenary based on 4C detection. Background Technology

[0002] The overhead contact system of high-speed railways is a core component of the high-speed railway power supply system, and its safe and stable operation is directly related to the reliability and safety of railway operations. The contact system consists of numerous anti-loosening components, including bolt-nut assemblies, locking washers, and cotter pins. These components are subjected to enormous vibrations, temperature changes, and mechanical stresses during high-speed train operation, making them prone to loosening or even detachment. Component detachment not only affects the normal operation of the power supply system but can also, in severe cases, lead to contact wire breakage, equipment damage, or even major safety accidents.

[0003] In existing technologies, the monitoring of anti-loosening components in high-speed railway catenary mainly employs traditional image recognition technology combined with a simple threshold method for state detection, achieving basic functions such as determining the existence of components and identifying obvious anomalies. However, this method uses a static judgment mode based on a single time point, lacking the ability to dynamically monitor the loosening process of anti-loosening components and analyze its spatiotemporal characteristics. It cannot capture the patterns and spatial correlations of the component's loosening state over time, resulting in the ability to identify only obvious loosening that has already occurred, but failing to predict potential detachment risks. This leads to low early warning levels and high hazards in the monitoring of anti-loosening components in high-speed railway catenary, with a significant delay in early warning time. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for preventing the detachment of components in high-speed railway catenary based on 4C detection. This solves the problem that the existing technology lacks the ability to dynamically monitor and analyze the spatiotemporal characteristics of the loosening process of anti-loosening components, and cannot capture the laws and spatial correlation characteristics of the evolution of the loosening state of components over time. As a result, it can only identify obvious loosening that has already occurred, but cannot predict potential detachment risks, which leads to the problem of low early warning and high hazards in the monitoring of anti-loosening components of high-speed railway catenary, and serious delay in early warning time.

[0005] The technical solution of this invention is implemented as follows: In a first aspect, this invention provides a method for preventing the detachment of components from the overhead contact system of high-speed railways based on 4C detection, comprising the following steps:

[0006] Acquire multi-source detection data of high-speed railway catenary components, including raw image data and historical status data collected by the 4C detection device;

[0007] The original image data is identified by the target detection network model to obtain the key component data of the high-speed railway catenary. The key components include bolt-nut assemblies, locking washers and cotter pins.

[0008] Feature data of key components is obtained by extracting features from the key component data using a fine-grained detection network model.

[0009] Based on the key component feature data, the anti-loosening status of the key component is judged to obtain the anti-loosening detection data of the key component;

[0010] A spatiotemporal feature matrix is ​​constructed based on anti-loosening detection data and historical status data, and the detachment risk coefficient of the key component is calculated through a risk assessment model.

[0011] Based on the aforementioned detachment risk coefficient, a detachment prevention early warning is generated, which includes location information, anomaly type, and maintenance suggestions, and then sent to management personnel.

[0012] Based on the above technical solutions, preferably, the acquisition of multi-source detection data of high-speed railway catenary components includes raw image data and historical status data collected by the 4C detection device, including:

[0013] The original image data of the high-speed railway catenary components are collected by the 4C detection device. The original image data includes image data acquired at different angles and under different lighting conditions.

[0014] Historical status data of high-speed railway catenary components are read from the historical status database. The historical status data includes historical anti-loosening detection data and related maintenance records.

[0015] Based on the above technical solutions, preferably, the key component data of the high-speed railway catenary is obtained by recognizing the original image data through a target detection network model, including:

[0016] A target detection network model is constructed based on the YOLOv5 network. Candidate regions are generated through the target detection network model, and multi-scale features are extracted using an adaptive feature pyramid to obtain the coarse-grained positions of bolt-nut combinations, locking washers, and cotter pins.

[0017] The candidate regions are filtered using a dynamic confidence threshold, and the overlapping candidate regions are fused and optimized by combining the target size distribution characteristics to output the accurate bounding boxes of key components.

[0018] Based on the above technical solutions, preferably, the step of extracting features from the key component data using a fine-grained detection network model to obtain key component feature data includes:

[0019] The fine-grained detection network model includes a multi-branch feature extraction network and a multi-scale keypoint regression network;

[0020] Fine-grained features are extracted from the key components using a multi-branch feature extraction network, which includes morphological feature branches, geometric relationship feature branches, and texture feature branches.

[0021] The feature points of the key components are accurately located based on a multi-scale key point regression network. The feature points of the key components include the endpoints of the anti-loosening mark line, the junction of the long and short ends of the locking washer, and the root and bending points at both ends of the cotter pin.

[0022] Based on the above technical solutions, preferably, the step of judging the anti-loosening status of the key component based on the key component feature data to obtain the anti-loosening detection data of the key component includes:

[0023] For the characteristic data of different key components, calculate the corresponding anti-loosening state parameters. The anti-loosening state parameters include the linear offset of the anti-loosening mark line, the spatial angle between the locking washer and the clamp body, and the angle between the two bent sides of the cotter pin.

[0024] Based on the preset standard values ​​of anti-loosening state parameters, threshold judgment is performed on the calculated anti-loosening state parameters to determine the anti-loosening state level of each key component and generate standardized anti-loosening test data.

[0025] Based on the above technical solutions, preferably, the step of constructing a spatiotemporal feature matrix based on anti-loosening detection data and historical state data, and calculating the detachment risk coefficient of the key component through a risk assessment model, includes:

[0026] A spatiotemporal feature matrix is ​​constructed, aligning the loosening detection data with historical state data in both time and space dimensions to obtain a multidimensional feature matrix reflecting the changes in the loosening state of key components over time and their spatial distribution.

[0027] By deeply integrating the risk assessment model, the spatiotemporal feature matrix is ​​identified to obtain the detachment risk coefficient of each key component;

[0028] The deep integration risk assessment model adopts a three-layer architecture, wherein:

[0029] The first layer establishes a basic risk assessment model for component loosening based on a gradient boosting decision tree, generating initial risk coefficients;

[0030] The second layer combines a long short-term memory network to capture time series trend features and corrects the initial risk coefficient in the time dimension.

[0031] The third layer models the spatial topological relationships between components using a graph convolutional network, and performs spatial dimension risk correction on the initial risk coefficient.

[0032] Based on the above technical solutions, preferably, the step of generating an anti-detachment early warning report based on the detachment risk coefficient, including location information, anomaly type, and maintenance suggestions, and sending it to management personnel, includes:

[0033] Based on the risk coefficient of detachment, combined with the importance of components and the operating level of the line, a multi-level early warning mechanism is established to generate an anti-detachment early warning report that includes precise location information, anomaly type analysis and maintenance priority.

[0034] The multi-level early warning mechanism includes information level, prompt level, warning level, and emergency level, wherein:

[0035] For information level, the test results are recorded without triggering an alert; for prompt level, a low-priority alert is generated and will be monitored during the next routine maintenance; for warning level, a medium-priority alert is generated and a special maintenance is carried out within one week; for emergency level, a high-priority alert is generated and emergency handling is carried out within 24 hours.

[0036] The early warning report is sent to relevant management personnel, and maintenance suggestions and resource allocation plans are generated based on the risk characteristics of component loosening and historical maintenance experience.

[0037] Secondly, the present invention also provides a high-speed rail contact network anti-component detachment system based on 4C detection, the system comprising:

[0038] The data acquisition module is used to acquire multi-source detection data of high-speed railway catenary components. The multi-source detection data includes raw image data and historical status data collected by the 4C detection device.

[0039] The critical component identification module is used to identify the original image data through the target detection network model and obtain the key component data of the high-speed rail catenary. The key components include bolt-nut assemblies, locking washers and cotter pins.

[0040] The feature extraction module is used to extract features from the key component data using a fine-grained detection network model to obtain key component feature data.

[0041] The anti-loosening detection module is used to determine the anti-loosening status of the key component based on the key component feature data, and obtain the anti-loosening detection data of the key component:

[0042] The detachment risk calculation module is used to construct a spatiotemporal feature matrix based on anti-loosening detection data and historical status data, and calculate the detachment risk coefficient of the key component through a risk assessment model.

[0043] The anti-fall-off early warning module is used to generate an anti-fall-off early warning report based on the aforementioned fall-off risk coefficient, which includes location information, anomaly type, and maintenance suggestions, and then sends it to the management personnel.

[0044] Thirdly, the present invention also provides an electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus;

[0045] The processor, memory, and communication interface communicate with each other through the bus. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to implement the steps of a method for preventing components from falling off high-speed rail catenary based on 4C detection.

[0046] Fourthly, the present invention also provides a computer-readable storage medium storing computer instructions that enable a computer to implement steps such as those in a method for preventing components from falling off a high-speed rail contact network based on 4C detection.

[0047] The method and system for preventing the detachment of components from high-speed railway catenary based on 4C detection, as proposed in this invention, have the following advantages over existing technologies:

[0048] (1) By collecting data from multiple sources, a two-level detection model is constructed to judge the loosening status, assess the risk and provide multi-level early warning. Multi-scale feature extraction of target detection and fine-grained detection network is integrated. Based on the geometric features, three-dimensional attitude estimation and angle calculation, the state parameters are accurately quantified. Combined with the spatiotemporal feature matrix and the dynamic prediction of the deep integrated risk assessment model, high-precision monitoring of the loosening status of key anti-loosening components of the high-speed rail contact network is realized. It can provide early warning of the risk of falling off, build a closed-loop management mechanism, effectively prevent pantograph-catenary failure caused by loosening of parts, and ensure the power supply safety under high-speed operation conditions.

[0049] (2) By integrating adaptive feature pyramid, dynamic confidence threshold and multi-task loss function into the target detection network model, the ability to accurately locate and identify key components of high-speed rail contact network in complex environments is realized. The adaptive feature pyramid dynamically adjusts the feature weights of different levels through the channel attention mechanism, which enhances the detection sensitivity of small components such as cotter pins. The dynamic confidence threshold adaptively adjusts the detection threshold according to the component size characteristics, which significantly reduces the false recognition rate of bolt-nut combination and background, and improves the detection rate of small targets.

[0050] (3) By combining a multi-branch feature extraction network and a multi-scale key point regression network in the fine-grained detection network model, the accurate feature extraction and sub-pixel level key point positioning of the key anti-loosening components of the high-speed rail contact network were realized. The multi-branch feature extraction network enhances the physical characteristics of different types of components through three dedicated feature channels: morphology, geometric relationship and texture. The adaptive feature enhancement function is used to dynamically adjust the weights so that the linear features of the anti-loosening mark line, the planar geometric features of the stop shim and the curvature features of the cotter pin can be optimally expressed at the same time. The multi-scale key point regression network uses a collaborative regression mechanism to geometrically constrain the key points as a whole, which improves the positioning accuracy in harsh environments such as uneven lighting, partial occlusion and variable angles, reduces the positioning error of the anti-loosening mark line endpoints and improves the angle measurement accuracy of the stop shim. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a method for preventing components from falling off high-speed railway catenary based on 4C detection, according to the present invention.

[0053] Figure 2 This is a structural diagram of a high-speed rail contact network anti-component detachment system based on 4C detection according to the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides a method for preventing components from falling off high-speed railway overhead contact lines based on 4C detection, comprising the following steps:

[0056] Acquire multi-source detection data of high-speed railway catenary components, including raw image data and historical status data collected by the 4C detection device;

[0057] The original image data is identified by a target detection network model to obtain key component data of the high-speed railway catenary. The key components include bolt-nut assemblies, retaining washers, and cotter pins.

[0058] Feature extraction of the key component data is performed by using a fine-grained detection network model to locate the anti-loosening marking line, the long and short ends of the locking washer, and the bent end of the cotter pin, thereby obtaining the key component feature data.

[0059] Based on the key component feature data, the anti-loosening status of the key component is determined, and the anti-loosening detection data of the key component is obtained, including:

[0060] a. Linear offset judgment based on the geometric characteristics of the anti-loosening marking line

[0061] b. Analyze the spatial relationship between the stop washer and the clamp body using three-dimensional attitude estimation.

[0062] c. Calculate the included angle between the two bent sides of the cotter pin;

[0063] A spatiotemporal feature matrix is ​​constructed based on anti-loosening detection data and historical status data, and the detachment risk coefficient of the key component is calculated through a risk assessment model.

[0064] Based on the aforementioned detachment risk coefficient, a detachment prevention early warning is generated, which includes location information, anomaly type, and maintenance suggestions, and then sent to management personnel.

[0065] Specifically, this embodiment constructs a two-level detection model through multi-source data acquisition to determine the loosening status, assess risks, and provide multi-level early warnings. It integrates multi-scale feature extraction from target detection and fine-grained detection networks, and uses a precise quantification method for state parameters based on geometric features, three-dimensional attitude estimation, and angle calculation. Combined with the dynamic prediction of spatiotemporal feature matrices and a deeply integrated risk assessment model, it achieves high-precision monitoring of the loosening status of key anti-loosening components of the high-speed railway catenary. This enables early warning of detachment risks, establishes a closed-loop management mechanism, effectively prevents pantograph-catenary faults caused by loose components, and ensures power supply safety under high-speed operating conditions.

[0066] The acquisition of multi-source detection data for high-speed railway catenary components includes raw image data and historical status data collected by the 4C detection device, including:

[0067] The original image data of the high-speed railway catenary components are collected by the 4C detection device. The original image data includes image data acquired at different angles and under different lighting conditions.

[0068] In one specific embodiment, the acquired raw image data is preprocessed. The preprocessing steps include image denoising, image enhancement, image cropping, and distortion correction to improve image resolution and the accuracy of subsequent target detection.

[0069] Historical status data of high-speed railway catenary components are read from the historical status database. The historical status data includes historical anti-loosening detection data and related maintenance records.

[0070] In one specific embodiment, the read historical state data is normalized and anomaly information is marked. The normalization process is used to eliminate the magnitude difference in the data caused by environmental factors, and the anomaly information marking is used to provide reliable data support for constructing a spatiotemporal feature matrix and risk assessment.

[0071] Specifically, this embodiment achieves high-quality data acquisition and processing capabilities in complex high-speed rail catenary environments by employing a four-step multi-stage image preprocessing process and dual-channel data normalization processing.

[0072] The image acquisition strategy, tailored to different angles and lighting conditions, effectively solves the challenges of strong light reflection, limited angles, and variable environments in high-speed rail catenary inspection, improving image resolution and edge clarity of small components such as cotter pins. Simultaneously, by normalizing historical state data and implementing an anomaly labeling mechanism, data biases caused by different environments, equipment, and times are eliminated, establishing a standardized dataset with consistent data volume and balanced information density. This improves model training efficiency, reduces sensitivity to environmental changes, and ensures the stability of detection capabilities under various adverse weather and lighting conditions.

[0073] The process of identifying raw image data using a target detection network model to obtain key component data of high-speed railway catenary components includes:

[0074] A target detection network model is constructed based on the YOLOv5 network. Candidate regions are generated through the target detection network model, and multi-scale features are extracted using an adaptive feature pyramid to obtain the coarse-grained positions of bolt-nut combinations, locking washers, and cotter pins.

[0075] In one specific embodiment, the feature map weights of the adaptive feature pyramid are calculated as follows:

[0076] ;

[0077] in, For the first The fusion weights of the layer feature maps For channel attention module, For the first Layer feature map, As the first learnable parameter, This represents the number of feature layers.

[0078] The candidate regions are filtered using a dynamic confidence threshold, and the overlapping candidate regions are fused and optimized by combining the target size distribution characteristics to output the accurate bounding boxes of key components.

[0079] In one specific embodiment, the dynamic confidence threshold is calculated as follows:

[0080] ;

[0081] in, For key components Class confidence threshold To train the key components The average target size of the class. To train the key components The standard deviation of the target size of the class The first empirical coefficient, The second empirical coefficient, It is the first zero constant.

[0082] In one specific embodiment, the loss function of the target detection network model is calculated as follows:

[0083] In one specific embodiment, the loss function of the target detection network model is calculated as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] in, Let be the loss function of the object detection network model. Focal loss function Let CIoU be the loss function. Let be the angle loss function. , , These are the balance coefficients for the Focal loss function, CIoU loss function, and angle loss function, respectively. For the model to the first The probability value of a sample being predicted as the positive class. The weighting factor for easy and difficult samples. Image location Spatial attention score, The total number of samples, To predict the intersection-union ratio (IoU) between the bounding box and the ground truth bounding box, To predict the center point of the bounding box Center point of the actual bounding box The Euclidean distance between them The diagonal length of the minimum closure rectangle that simultaneously covers both the predicted and ground truth bounding boxes. As a weighting factor, A parameter for measuring aspect ratio consistency. To predict the target rotation angle, The rotation angle of the actual target.

[0089] In one specific embodiment, the target detection network model adopts an improved YOLOv5 network architecture, and has been optimized in three aspects to meet the component detection requirements in the complex environment of high-speed rail catenary.

[0090] First, an adaptive feature pyramid structure is introduced to assign optimal weights to feature maps at different levels. In experimental verification, when detecting small cotter pins, the weight of shallow feature maps automatically increases to 0.65, while the weight of deep feature maps decreases to 0.20, effectively enhancing the detection capability for targets with side lengths less than 30 pixels. Second, a dynamic confidence threshold mechanism is designed, adaptively adjusting the threshold according to the target size distribution. The confidence threshold for bolt-nut combinations (average size 120 pixels) in the high-speed rail contact network environment is set to 0.42, while it is reduced to 0.32 for cotter pins (average size 25 pixels), reducing the false negative rate for small targets. Third, a composite loss function including Focal loss, CIoU loss, and angle loss is constructed. In particular, the introduction of angle loss enables the model to accurately capture the posture of components in rotation, reducing the angle prediction error from the original ±12° to ±3.5° during model training.

[0091] Specifically, this embodiment, through the improved target detection network model, achieves high-precision identification of key anti-loosening components of high-speed rail catenary in complex environments, effectively overcoming three major challenges of traditional target detection in high-speed rail scenarios: First, it solves the detection problem of small-sized components such as cotter pins (only 2-3mm in diameter), increasing the detection rate from 68.3% of the baseline model to 94.7%; second, it significantly improves the detection balance of components of different sizes, reducing the difference in F1-score among large, medium, and small targets from the original 0.18 to 0.05; finally, through the angle perception mechanism, the model can maintain stable detection performance under a wide range of viewing angle changes from 30° to 150°, providing accurate target localization and segmentation for fine-grained anti-loosening status analysis, reducing computational redundancy in the feature extraction stage, improving overall inference speed, and enhancing adaptability and stability in actual high-speed rail line inspections.

[0092] The step of extracting features from the key component data using a fine-grained detection network model to obtain key component feature data includes:

[0093] The fine-grained detection network model includes a multi-branch feature extraction network and a multi-scale keypoint regression network;

[0094] Fine-grained feature extraction is performed on the key components using a multi-branch feature extraction network, which includes morphological feature branches, geometric relationship feature branches, and texture feature branches.

[0095] In one specific embodiment, the multi-branch feature extraction network employs an adaptive feature enhancement function, dynamically adjusting the feature weight coefficients of each branch according to the characteristics of different key components, specifically as follows:

[0096] For the anti-loosening marking lines, increase the weight of texture feature branches;

[0097] For stop shims, increase the weight of geometric relationship feature branches;

[0098] For cotter pins, increase the weight of morphological feature branches;

[0099] Meanwhile, multi-scale response fusion is performed on each feature map to eliminate the feature imbalance problem caused by size differences.

[0100] In one specific embodiment, the adaptive feature enhancement function is calculated as follows:

[0101] In one specific embodiment, the target detection network model adopts an improved YOLOv5 network architecture, and has been optimized in three aspects to meet the component detection requirements in the complex environment of high-speed rail catenary.

[0102] First, an adaptive feature pyramid structure is introduced to assign optimal weights to feature maps at different levels. In experimental verification, when detecting small cotter pins, the weight of shallow feature maps automatically increases to 0.65, while the weight of deep feature maps decreases to 0.20, effectively enhancing the detection capability for targets with side lengths less than 30 pixels. Second, a dynamic confidence threshold mechanism is designed, adaptively adjusting the threshold according to the target size distribution. The confidence threshold for bolt-nut combinations (average size 120 pixels) in the high-speed rail contact network environment is set to 0.42, while it is reduced to 0.32 for cotter pins (average size 25 pixels), reducing the false negative rate for small targets. Third, a composite loss function including Focal loss, CIoU loss, and angle loss is constructed. In particular, the introduction of angle loss enables the model to accurately capture the posture of components in rotation, reducing the angle prediction error from the original ±12° to ±3.5° during model training.

[0103] Specifically, this embodiment, through the improved target detection network model, achieves high-precision identification of key anti-loosening components of high-speed rail catenary in complex environments, effectively overcoming three major challenges of traditional target detection in high-speed rail scenarios: First, it solves the detection problem of small-sized components such as cotter pins (only 2-3mm in diameter), increasing the detection rate from 68.3% of the baseline model to 94.7%; second, it significantly improves the detection balance of components of different sizes, reducing the difference in F1-score among large, medium, and small targets from the original 0.18 to 0.05; finally, through the angle perception mechanism, the model can maintain stable detection performance under a wide range of viewing angle changes from 30° to 150°, providing accurate target localization and segmentation for fine-grained anti-loosening status analysis, reducing computational redundancy in the feature extraction stage, improving overall inference speed, and enhancing adaptability and stability in actual high-speed rail line inspections.

[0104] The step of extracting features from the key component data using a fine-grained detection network model to obtain key component feature data includes:

[0105] The fine-grained detection network model includes a multi-branch feature extraction network and a multi-scale keypoint regression network;

[0106] Fine-grained feature extraction is performed on the key components using a multi-branch feature extraction network, which includes morphological feature branches, geometric relationship feature branches, and texture feature branches.

[0107] In one specific embodiment, the multi-branch feature extraction network employs an adaptive feature enhancement function, dynamically adjusting the feature weight coefficients of each branch according to the characteristics of different key components, specifically as follows:

[0108] For the anti-loosening marking lines, increase the weight of texture feature branches;

[0109] For stop shims, increase the weight of geometric relationship feature branches;

[0110] For cotter pins, increase the weight of morphological feature branches;

[0111] Meanwhile, multi-scale response fusion is performed on each feature map to eliminate the feature imbalance problem caused by size differences.

[0112] In one specific embodiment, the adaptive feature enhancement function is calculated as follows:

[0113] ;

[0114] in, For the first Enhanced features for critical components, For the first The basic characteristics of key components, For the first The thermal response of key components For the first Average heatmap values ​​for key components. For the first Standard deviation of heatmap for key components For the first Enhancement coefficient of key components, It is the hyperbolic tangent activation function.

[0115] The feature points of the key components are accurately located based on a multi-scale key point regression network. The feature points of the key components include the endpoints of the anti-loosening mark line, the junction of the long and short ends of the locking washer, and the root and bending points at both ends of the cotter pin.

[0116] In one specific embodiment, the multi-scale key point regression network adopts a collaborative regression mechanism, constraining key feature points as a whole geometric structure. By introducing a loss function of geometric relationship between points, the geometric consistency of feature point positioning is improved. Specific methods include: for anti-loosening marker lines, enhancing the accuracy of endpoint positioning through linearity constraints; for locking shims, improving the accuracy of the junction points of long and short ends through planar orthogonal constraints; and for cotter pins, ensuring the geometric rationality of bending points through angle constraints.

[0117] In one specific embodiment, the fine-grained detection network model adopts a dual-core architecture based on a multi-branch feature extraction network and a multi-scale keypoint regression network using an improved ResNeXt-50. First, the multi-branch feature extraction network achieves accurate feature extraction for different types of anti-loosening components through three dedicated feature channels: the morphological feature branch uses a deformable convolutional network, which is particularly suitable for extracting the curvature features of cotter pins. In experiments, the offset field of the deformable convolutional network can adaptively adjust to the curvature variation region of the cotter pin edge, effectively improving edge accuracy; the geometric relationship feature branch is constructed based on a graph convolutional network. By establishing planar geometric relationship constraints for the locking gasket, the feature points at the gasket edge are treated as a graph structure, accurately capturing the relative positional relationship between the long and short ends of the gasket, reducing the spatial pose estimation error to 2.8°; the texture feature branch uses depthwise separable convolution to enhance the subtle texture variations of the anti-loosening marking lines, extracting marking line offsets that differ by only 0.15mm under low contrast.

[0118] Secondly, in this embodiment, the adaptive feature enhancement function dynamically allocates weights according to the characteristics of different components. Through a heatmap response mechanism, the texture branch weight is automatically increased to 0.58 when detecting anti-loosening markings, the geometric branch weight is increased to 0.62 when detecting locking shims, and the morphological branch weight is increased to 0.65 when detecting cotter pins. The dynamic weight adjustment strategy in this embodiment improves feature representation capability by 18.7% compared to fixed weights, and maintains high feature extraction quality even under partial occlusion conditions.

[0119] The multi-scale keypoint regression network, by introducing a collaborative regression mechanism and geometric constraint loss, applies a linearity constraint coefficient of 0.35 to the anti-loosening marker line, ensuring that the endpoint positioning accuracy remains at 0.18 mm even under vibration conditions at 350 km / h. Planar orthogonal constraints are introduced for the locking shims to ensure that the angle measurement error between the shims and the clamp contact surfaces is controlled within ±2.5°. Curvature-based angle constraints are applied to the cotter pins, establishing a loosening warning criterion by calculating the angle formed by the root of the cotter pin and the bending points at both ends. In actual line testing, this model maintains a 98.3% keypoint detection rate and sub-millimeter positioning accuracy within a temperature range of -10°C to 40°C and under various severe weather conditions such as rain, fog, and snow.

[0120] Specifically, the fine-grained detection network model in this embodiment achieves high-precision feature extraction and key point localization for critical anti-loosening components of high-speed railway catenary. The multi-branch feature extraction strategy accurately captures the key physical features of different anti-loosening devices, enabling the model to simultaneously process components of varying shapes, increasing accuracy from 87.2% for standard convolutional networks to 96.8%. The adaptive feature enhancement function reduces feature instability under different lighting and angle conditions, improving feature quality by 42% in shadow areas and feature detection capability by 37% in overexposed edge areas. The geometric constraint mechanism of the multi-scale key point regression network ensures the physical rationality of key point positions, solving the key point drift problem in traditional methods, and increasing the average accuracy of overall anti-loosening device identification from 76.5% of the baseline model to 94.2%.

[0121] The step of determining the anti-loosening status of the key component based on its feature data, and obtaining the anti-loosening detection data of the key component, includes:

[0122] For the characteristic data of different key components, the corresponding anti-loosening state parameters are calculated. The anti-loosening state parameters include the linear offset of the anti-loosening mark line, the spatial angle between the locking washer and the clamp body, and the angle between the two bent sides of the cotter pin.

[0123] In one specific embodiment, for the anti-loosening marking line, straight line segments are extracted using Hough transform, and the included angle deviation and endpoint offset between adjacent marking lines are calculated to characterize the looseness of the bolt-nut assembly; for the locking washer, the spatial position of the locking washer is reconstructed using a three-dimensional attitude estimation algorithm, and the contact angle between its long end and the side end face of the clamp body and the contact angle between its short end and the nut in six aspects are calculated; for the cotter pin, the skeleton lines of the two curved sides are extracted, the included angle value between the two sides is calculated, and it is compared with the standard installation angle to analyze its deformation degree.

[0124] Based on the preset standard values ​​of anti-loosening state parameters, threshold judgment is performed on the calculated anti-loosening state parameters to determine the anti-loosening state level of each key component and generate standardized anti-loosening test data.

[0125] In one specific embodiment, a three-level anti-loosening condition judgment standard is established based on the mechanical characteristics of the anti-loosening components and historical fault data. The standard is as follows: Normal state: anti-loosening mark offset < 0.3 mm, stop shim position deviation angle < 5°, cotter pin bending angle between 120° and 130°; Slightly loose state: anti-loosening mark offset between 0.3 mm and 0.5 mm, stop shim position deviation angle between 5° and 10°, cotter pin bending angle deviating from the standard range by no more than ±3°; Severely loose state: anti-loosening mark offset ≥ 0.5 mm, stop shim position deviation angle ≥ 10°, cotter pin bending angle deviating from the standard range by more than ±5°. Simultaneously, the judgment results are corrected by comprehensively considering the component's stress state, installation location, and environmental factors, thereby improving the adaptability and reliability of the anti-loosening condition judgment.

[0126] Specifically, this embodiment establishes a precise anti-loosening state calculation and classification judgment mechanism based on physical characteristics for key components of the high-speed rail catenary, achieving millimeter-level precision in quantifying loosening state and early risk identification capabilities.

[0127] For anti-loosening marking lines, the method of extracting straight line segments using Hough transform and calculating offset is used to improve the accuracy of bolt-nut loosening detection. For locking washers, the spatial position is reconstructed and the fitting angle is calculated using a three-dimensional attitude estimation algorithm, which solves the problem that traditional methods cannot accurately judge the spatial attitude changes of washers in the complex environment of high-speed rail, and improves the accuracy of position deviation angle measurement. For cotter pins, the skeleton line extraction and included angle calculation method is used to improve the deformation detection sensitivity.

[0128] Meanwhile, the three-level anti-loosening status judgment standard (normal state, slightly loose state and severely loose state) established based on the mechanical characteristics of anti-loosening components and historical fault data, combined with the comprehensive correction mechanism of component stress state, installation position and environmental factors, greatly improves the reliability of anti-loosening status judgment, so that it can still maintain an accuracy rate of 96.3% in the complex high-speed rail operating environment.

[0129] The process involves constructing a spatiotemporal feature matrix based on anti-loosening detection data and historical status data, and calculating the detachment risk coefficient of the key components using a risk assessment model, including:

[0130] A spatiotemporal feature matrix is ​​constructed by aligning the loosening detection data with historical state data in both time and space dimensions to obtain a multidimensional feature matrix that reflects the changes in the loosening state of key components over time and their spatial distribution.

[0131] In one specific embodiment, the spatiotemporal feature matrix is ​​composed of the following static features, time series features, and spatially correlated features, wherein:

[0132] Static features include component type, installation location, load characteristics, and line grade; time-series features include historical trends, rates of change, periodic fluctuations, and anomalous abrupt changes in component anti-loosening status parameters; spatial correlation features include the correlation of loosening status between adjacent components, the correlation of loosening status among different components within the same component assembly, and regional environmental influencing factors. By aligning and standardizing these time-space dual-dimensional features, inconsistencies in detection frequencies and interference from environmental factors are eliminated, enhancing the consistency and representativeness of the feature matrix.

[0133] By deeply integrating the risk assessment model, the spatiotemporal feature matrix is ​​identified, and the risk coefficient of detachment of each key component is obtained by comprehensively considering the component type, operating environment, stress condition and historical change trend.

[0134] The deep integration risk assessment model adopts a three-layer architecture, wherein:

[0135] The first layer establishes a basic risk assessment model for component loosening based on a gradient boosting decision tree, generating initial risk coefficients;

[0136] The second layer combines a long short-term memory network to capture time series trend features and corrects the initial risk coefficient in the time dimension.

[0137] The third layer models the spatial topological relationship between components through a graph convolutional network, comprehensively analyzes the impact of the loosening state of adjacent components on the target component, and performs spatial dimension risk correction on the initial risk coefficient.

[0138] In one specific embodiment, the initial risk coefficient is normalized and mapped to the [0,1] interval, and divided into three levels according to risk level: low risk (0-0.3), medium risk (0.3-0.7) and high risk (0.7-1.0).

[0139] Specifically, this embodiment constructs a three-dimensional spatiotemporal feature matrix that includes static features, time series features, and spatial correlation features, and combines it with a three-layer architecture of a deeply integrated risk assessment model to achieve a high-precision prediction capability for the risk of loosening components falling off high-speed railway catenary.

[0140] This embodiment's spatiotemporal feature matrix organically integrates the gradual changes in component loosening status over time with their spatial correlation, solving the problem that traditional single-point-of-time detection methods cannot capture the evolutionary patterns of loosening. In the deep integrated risk assessment model, the first-layer gradient boosting decision tree can identify key factors of basic component loosening risk with an accuracy of 92.5%; the second-layer long short-term memory network can capture historical trend features over 30 days, effectively predicting the acceleration period of loosening, extending the lead time for early warning from 1-2 days to 3-7 days; the third-layer graph convolutional network, by modeling the spatial topological relationships between components, uncovers the loosening transmission chain within the same connection group, improving the spatial correlation identification accuracy by 41%.

[0141] The process involves generating a fall-off prevention warning report based on the fall-off risk coefficient, which includes location information, anomaly type, and maintenance suggestions. This report is then sent to management personnel.

[0142] Based on the risk coefficient of detachment, combined with the importance of components and the operating level of the line, a multi-level early warning mechanism is established to generate an anti-detachment early warning report that includes precise location information, anomaly type analysis and maintenance priority.

[0143] The multi-level early warning mechanism includes information level, prompt level, warning level, and emergency level, wherein:

[0144] If the information level has a risk coefficient of detachment below 0.3, the detection results will be recorded and no warning will be triggered.

[0145] The risk coefficient for the detachment at the warning level is between 0.3 and 0.5, generating a low-priority warning for monitoring during the next routine maintenance.

[0146] The risk coefficient for detachment at the warning level is between 0.5 and 0.7, generating a medium-priority warning and requiring special maintenance within one week;

[0147] The emergency level has a risk coefficient of detachment of 0.7 or higher, generating a high-priority warning that requires emergency handling within 24 hours.

[0148] In one specific embodiment, the cluster risk of multiple components simultaneously exhibiting abnormalities within the same segment is assessed. When the density of abnormal components in a certain segment exceeds a preset threshold, the overall early warning level of that segment is raised to prevent systemic risks.

[0149] The system sends early warning reports to relevant managers through an intelligent distribution system. At the same time, it automatically generates maintenance suggestions and resource allocation plans based on the risk characteristics of loose components and historical maintenance experience.

[0150] In one specific embodiment, the intelligent distribution system adopts a hierarchical, multi-channel information transmission method, including:

[0151] The system automatically selects notification channels based on the warning level, including in-system messages, emails, SMS, and emergency voice calls; it intelligently identifies warning recipients based on organizational structure and job responsibilities, ensuring that critical information is simultaneously delivered to on-site maintenance, technical support, and management decision-making personnel; for different types of loosening anomalies, it automatically generates standardized maintenance suggestions containing anomaly descriptions, possible causes, recommended tools, spare parts requirements, and time estimates, based on historical maintenance databases; it provides closed-loop management of the maintenance process, requiring maintenance personnel to provide feedback on actual handling results and transmit the handling information back to the historical status database, improving the data samples for loosening pattern recognition and continuously enhancing the accuracy of the risk assessment model.

[0152] Specifically, this embodiment achieves fully automated management of the anti-loosening components of high-speed rail contact networks from risk assessment to early warning processing by establishing a multi-level early warning mechanism and intelligent distribution system based on the risk coefficient of detachment.

[0153] The multi-level early warning mechanism divides the risk coefficient of component detachment into four levels: information level, alert level, warning level, and emergency level. It dynamically adjusts the level based on the importance of the component and the operational status of the line, overcoming the crudeness and blindness of the traditional binary early warning model and improving the accuracy of early warnings. For the intelligent identification algorithm targeting clustered risks, when the density of abnormal components within a 5-kilometer section exceeds a set threshold, the overall early warning level is automatically raised.

[0154] Meanwhile, the hierarchical, multi-channel intelligent distribution system automatically selects the optimal notification method based on the warning level, reducing the emergency warning response time from 120 minutes to less than 15 minutes. Intelligent warning routing based on job responsibilities ensures that information is simultaneously transmitted to the maintenance, technical, and management lines, eliminating information silos. Standardized maintenance suggestions automatically generated for different loosening types, including anomaly descriptions, possible causes, and resource requirement estimates, improve maintenance efficiency and shorten average processing time. The closed-loop management mechanism continuously optimizes the risk assessment model by collecting maintenance feedback data, increasing the warning accuracy rate from 87.5% to 96.8% within 6 months, and reducing the false alarm rate by 71%.

[0155] Please see Figure 2 The present invention also provides a high-speed rail contact network anti-component detachment system based on 4C detection, the system comprising:

[0156] The data acquisition module is used to acquire multi-source detection data of high-speed railway catenary components. The multi-source detection data includes raw image data and historical status data collected by the 4C detection device.

[0157] The key component identification module is used to identify raw image data through a target detection network model and obtain key component data of high-speed railway catenary components.

[0158] The feature extraction module is used to extract features from the key component data using a fine-grained detection network model to obtain key component feature data.

[0159] The anti-loosening detection module is used to determine the anti-loosening status of the key component based on the key component feature data, and obtain the anti-loosening detection data of the key component:

[0160] The detachment risk calculation module is used to construct a spatiotemporal feature matrix based on anti-loosening detection data and historical status data, and calculate the detachment risk coefficient of the key component through a risk assessment model.

[0161] The anti-fall-off early warning module is used to generate an anti-fall-off early warning report based on the aforementioned fall-off risk coefficient, which includes location information, anomaly type, and maintenance suggestions, and then sends it to the management personnel.

[0162] Specifically, this embodiment presents a high-speed rail catenary anti-component detachment system based on 4C detection. Through six functional modules—data acquisition, key component identification, feature extraction, anti-loosening detection, detachment risk calculation, and anti-loosening early warning—it achieves intelligent closed-loop monitoring and early warning of high-speed rail catenary anti-loosening components, improving system integration and coordination. By adopting a modular design and distributed computing architecture, seamless data flow and processing are achieved between functional modules, reducing system response latency and improving data processing consistency.

[0163] Especially in the high-speed rail operating environment, the high-speed rail contact network anti-component detachment system based on 4C detection in this embodiment exhibits three major advantages: First, through the deep integration of the data acquisition module and the key component identification module, adaptive acquisition and preprocessing of images in the moving environment are achieved, enabling the system to maintain extremely high recognition accuracy even under high-speed operation conditions of 350km / h; Second, the linkage mechanism between the feature extraction module and the anti-loosening detection module transforms maintenance decisions from traditional experience-based judgment to data-driven intelligent analysis, improving resource allocation efficiency; Third, the detachment risk calculation module and the anti-detachment early warning module achieve dynamic allocation of computing load based on a cloud-edge collaborative architecture, increasing processing capacity by 68% under high load conditions while ensuring the real-time nature of emergency early warnings.

[0164] The present invention also discloses an electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus; wherein the processor, memory, and communication interface communicate with each other through the bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement a method for preventing components from falling off high-speed rail contact wires based on 4C detection.

[0165] This invention also discloses a computer-readable storage medium storing computer instructions that enable the computer to implement all or part of the steps of the method for preventing component detachment from high-speed rail contact wire based on 4C detection, as described in this embodiment of the invention. The storage medium includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-speed rail overhead line system anti-part falling method based on 4C detection, characterized in that, The method comprises the following steps: obtaining multi-source detection data of high-speed rail overhead line system components, the multi-source detection data comprising original image data collected by a 4C detection device and historical state data; identifying the original image data through a target detection network model to obtain key component data of the high-speed rail overhead line system components, the key components comprising a bolt-nut combination, a stop washer and a split pin; identifying the original image data through the target detection network model to obtain key component data of the high-speed rail overhead line system components, comprising: constructing a target detection network model based on a YOLOv5 network, generating a candidate region through the target detection network model, extracting multi-scale features by using an adaptive feature pyramid, and obtaining coarse-grained positions of the bolt-nut combination, the stop washer and the split pin; screening the candidate region by using a dynamic confidence threshold, fusing and optimizing overlapping candidate regions in combination with target size distribution characteristics, and outputting accurate bounding boxes of the key components; extracting features of the key component data through a fine-grained detection network model to obtain key component feature data; judging the loosening state of the key components based on the key component feature data to obtain loosening detection data of the key components; constructing a spatio-temporal feature matrix based on the loosening detection data and the historical state data, and calculating a shedding risk coefficient of the key components through a risk assessment model; the method of constructing a spatio-temporal feature matrix based on the loosening detection data and the historical state data, and calculating a shedding risk coefficient of the key components through a risk assessment model, comprising: constructing a spatio-temporal feature matrix, aligning the loosening detection data and the historical state data in the time dimension and the space dimension, and obtaining a multi-dimensional feature matrix reflecting the loosening state of the key components changing with time and the spatial distribution; identifying the spatio-temporal feature matrix through a deep integrated risk assessment model to obtain the shedding risk coefficient of each key component; the deep integrated risk assessment model adopts a three-layer architecture, wherein: the first layer establishes a component loosening basic risk assessment model based on gradient boosting decision trees to generate an initial risk coefficient; the second layer combines a long short-term memory network to capture time series trend features to correct the initial risk coefficient in the time dimension; the third layer models the spatial topological relationship between components through a graph convolution network to correct the initial risk coefficient in the space dimension; based on the shedding risk coefficient, a shedding prevention early warning is performed to generate a shedding prevention early warning report containing positioning information, abnormal type and maintenance suggestion, and the report is sent to management personnel.

2. A method for preventing the falling of components of a high-speed rail overhead line system according to claim 1, characterized in that, the method of obtaining multi-source detection data of high-speed rail overhead line system components, the multi-source detection data comprising original image data collected by a 4C detection device and historical state data, comprising: collecting original image data of high-speed rail overhead line system components through a 4C detection device, the original image data comprising image data obtained under different angles and different lighting conditions; reading historical state data of high-speed rail overhead line system components from a historical state database, the historical state data comprising historical loosening detection data and related maintenance records.

3. A method for preventing the falling of components of a high-speed rail overhead line system according to claim 1, characterized in that, the method of extracting features of the key component data through a fine-grained detection network model to obtain key component feature data, comprising: The fine-grained detection network model comprises a multi-branch feature extraction network and a multi-scale key point regression network; The multi-branch feature extraction network comprises a morphological feature branch, a geometric relationship feature branch and a texture feature branch; The multi-scale key point regression network is used for accurately positioning feature points of the key component, and the feature points of the key component comprise end points of a loosening mark line, intersection points of long and short ends of a stop washer, and curved points at two ends of a split pin.

4. The method for preventing the falling of components of a high-speed rail overhead line system according to claim 1, wherein, The anti-loosening state of the key component is judged based on the key component feature data, and anti-loosening detection data of the key component is obtained, comprising: For feature data of different key components, corresponding anti-loosening state parameters are calculated, and the anti-loosening state parameters comprise a linear offset of the loosening mark line, a spatial included angle between the stop washer and the clamp body, and an included angle value of two curved edges of the split pin; Based on a preset anti-loosening state parameter standard value, threshold judgment is performed on the calculated anti-loosening state parameters to determine the anti-loosening state level of each key component, and standardized anti-loosening detection data is generated.

5. The method for preventing the falling of components of a high-speed rail overhead line system according to claim 1, wherein, Based on the falling risk coefficient, a multi-level early warning mechanism is established to generate an anti-falling early warning report containing accurate positioning information, abnormal type analysis and maintenance priority, and is sent to the management personnel, comprising: Based on the falling risk coefficient, a multi-level early warning mechanism is established to generate an anti-falling early warning report containing accurate positioning information, abnormal type analysis and maintenance priority, and is sent to the management personnel, comprising: The multi-level early warning mechanism comprises information level, prompt level, warning level and emergency level, wherein: For the information level, the detection result is recorded, and no early warning is triggered; for the prompt level, a low-priority early warning is generated, and attention is paid during the next routine maintenance; for the warning level, a medium-priority early warning is generated, and special maintenance is performed within one week; for the emergency level, a high-priority early warning is generated, and emergency treatment is performed within 24 hours; The early warning report is sent to the relevant management personnel, and maintenance suggestions and resource allocation schemes are generated according to the component loosening risk characteristics and historical maintenance experience.

6. A system for preventing parts from falling off a high-speed rail overhead contact system based on 4C detection, used to perform a method for preventing parts from falling off a high-speed rail overhead contact system based on 4C detection according to any one of claims 1-5, characterized in that, The system comprises: A data acquisition module is configured to acquire multi-source detection data of the high-speed rail catenary component, wherein the multi-source detection data comprises original image data collected by a 4C detection device and historical state data; A key position identification module is configured to identify the original image data by a target detection network model to acquire key component data of the high-speed rail catenary component, wherein the key component comprises a bolt-nut combination, a stop washer and a split pin; A feature extraction module is configured to extract features of the key component data by a fine-grained detection network model to obtain key component feature data; An anti-loosening detection module is configured to judge an anti-loosening state of the key component based on the key component feature data to obtain anti-loosening detection data of the key component; A falling risk calculation module is configured to construct a space-time feature matrix based on the anti-loosening detection data and the historical state data, and calculate a falling risk coefficient of the key component by a risk evaluation model. The anti-falling early warning module is configured to perform anti-falling early warning based on the falling risk coefficient, generate an anti-falling early warning report containing positioning information, an abnormal type and a maintenance suggestion, and send the anti-falling early warning report to a manager.

7. An electronic device, comprising: The method comprises the following steps: at least one processor, at least one memory, a communication interface and a bus; wherein the processor, the memory and the communication interface communicate with each other through the bus, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to implement the method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions enable a computer to implement the method of any one of claims 1-5.

Images