Defect analysis method for electrified railway power supply safety detection monitoring system 5C

Abstract

This invention discloses a defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways, belonging to the field of computer vision detection technology. First, the mask of the sliding plate area and high-risk areas are obtained; then, multi-scale feature extraction is performed, and a fusion model is constructed using scale-adaptive fusion weights to obtain a defect fusion feature map; based on the physical mechanisms of three types of defects—foreign object attachment, sliding plate bending, and sliding plate breakage—decoupling discrimination is performed; finally, combining defect type, spatial location, and high-risk areas, a multi-factor coupled comprehensive risk assessment is conducted, outputting risk alarm levels and operation and maintenance decisions. This method utilizes attention mechanisms and mutual information adaptive enhancement to improve defect-sensitive scale features, solving the problem of missed detections caused by large scale spans. Furthermore, through structural prior constraints and physical mechanism decoupling, it effectively eliminates false detections, distinguishes true defects, and significantly improves detection accuracy and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of computer vision inspection technology, specifically a defect analysis method for the 5C power supply safety inspection and monitoring system of electrified railways. Background Technology

[0002] As a core component for high-speed trains to obtain traction power from the overhead contact line, the pantograph contactor's structural integrity and operational stability directly determine the train's power collection efficiency and are a crucial foundation for ensuring train safety. With the development of high-speed railways towards high speed, heavy load, and intelligent operation, the rail transit safety monitoring system places stringent requirements on pantograph contactor defect detection, demanding high reliability, low false alarm rates, strong interpretability, and engineering practicality. However, existing pantograph contactor defect detection methods are still insufficient to meet these industry needs.

[0003] Currently, pantograph sliding plate defect detection mainly relies on manual inspection and general machine vision defect detection methods. Manual inspection is not only inefficient and labor-intensive, but also prone to missed detections and misjudgments due to subjective human factors, and cannot meet the needs of real-time online monitoring of high-speed trains. General machine vision detection methods have many technical bottlenecks due to insufficient consideration of the structural characteristics and physical mechanisms of defect formation of pantograph sliding plates. Specifically, defects caused by foreign object attachment are easily confused with surface reflections and stains due to the strong randomness of their shape and color, and general feature extraction methods cannot effectively distinguish them. Sliding plate bending defects are characterized by continuous deformation of the overall geometric shape without obvious edge fracture features. Traditional target detection models can only capture local texture features and are unable to identify such overall structural anomalies. Sliding plate fracture defects have a huge scale range, with microcracks only a few pixels wide and large fractures reaching the centimeter level. Single-scale feature extraction methods are prone to missing small-scale defects and misjudging large-scale defects. Furthermore, the results of existing detection methods lack the support of an engineering-based risk assessment system. They can only achieve simple defect identification and cannot output operation and maintenance decisions with practical guidance. Overall, they suffer from problems such as high false alarm rate, poor scale adaptability, weak interpretability of results, and insufficient engineering practicality, making it difficult to meet the actual needs of pantograph condition monitoring in high-speed railways.

[0004] To address the aforementioned technical challenges, this invention starts with the structural prior features and physical mechanisms of defects in the pantograph sliding plate. Integrating technologies from multiple fields such as computer vision, deep learning, information theory, differential geometry, and statistical analysis, it proposes a defect analysis method for the 5C power supply safety monitoring system of electrified railways. Through structural perception modeling, multi-scale feature adaptive fusion, physical prior-driven defect decoupling and discrimination, and multi-factor coupled risk quantification assessment, it achieves high-precision detection and location of three typical defects. Simultaneously, it outputs tiered risk warning results and targeted operation and maintenance decisions, overcoming the shortcomings of existing technologies and improving the intelligence and engineering level of pantograph condition monitoring in high-speed railways. Summary of the Invention

[0005] The purpose of this invention is to provide a defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways, so as to solve the problems of high false alarm rate, poor scale adaptability, weak interpretability of results and insufficient engineering practicality in the existing pantograph sliding plate defect detection methods mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways includes the following steps: Step S1: Obtain the original image of the pantograph, perform skateboard area localization and structure extraction on the image to obtain the skateboard area mask and the high-risk area of ​​the pantograph skateboard working contact surface; Step S2: Under the dual constraints of the skateboard area mask and the high-risk area, multi-scale feature extraction is performed on the original pantograph image, and a multi-scale feature adaptive fusion model is constructed to obtain the defect fusion feature map through the model; Step S3: Based on the physical mechanism of pantograph sliding plate defects, the defect fusion feature map is decoupled and identified to distinguish three types of defects: foreign object attachment, sliding plate bending, and sliding plate breakage. Step S4: Combining the defect type, defect spatial location, and high-risk areas of the working contact surface, perform a multi-factor coupled comprehensive assessment of defect risk based on the defect identification results, and output the risk alarm level of the pantograph sliding plate defect and the corresponding operation and maintenance decision.

[0007] According to the above technical solution, the implementation method of skateboard region localization and structured extraction in step S1 is as follows: skateboard region localization is achieved through a deep learning model with a joint detection-segmentation structure; the skateboard geometric parameters are calculated based on the skateboard region mask, and a structural consistency constraint term is constructed to correct the localization result. The formula for the structural consistency constraint term is: in These are the proportional parameters of the skateboard structure. The standard aspect ratio for a predefined skateboard.

[0008] According to the above technical solution, the proportional parameter of the skateboard structure is the ratio of the projected length to the projected width of the skateboard in the image coordinate system. Both the projected length and projection width are calculated based on the extreme values ​​of the horizontal and vertical coordinates of the effective pixels in the skateboard region mask. If the result exceeds the geometric consistency tolerance threshold, the skateboard area positioning result will be discarded or corrected a second time.

[0009] According to the above technical solution, the construction method of the high-risk area of ​​the working contact surface in step S1 is as follows: extract the longitudinal center line of the skateboard based on the skateboard area mask, set the width control parameters on both sides of the center line, and determine the set of pixels that satisfy the absolute value of the difference between the vertical coordinate of the pixel and the corresponding center line position as the high-risk area.

[0010] According to the above technical solution, in step S2, multi-scale features are extracted using a feature pyramid network equipped with a channel-space-global attention module. These multi-scale features include at least high-resolution small-scale features representing minute foreign objects and microcracks, medium-scale features representing the initial characteristics of skateboard bending, and low-resolution large-scale features representing overall bending and large-size fractures. Scale-adaptive fusion weights are constructed based on mutual information. Where k is the scale index. The mutual information between scale features and defect category random variables is used to achieve adaptive fusion of multi-scale features through this weight.

[0011] According to the above technical solution, the calculation formula for obtaining the defect fusion feature map through multi-scale feature adaptive fusion in step S2 is as follows: Where s, m, and l are the small, medium, and large scale indices, respectively. The adaptive fusion weights are for each scale. Feature maps extracted at each scale.

[0012] According to the above technical solution, the discrimination logic for foreign object attachment defects in step S3 is as follows: perform saliency response analysis on the defect fusion feature map to generate candidate defect regions, calculate the Euclidean norm of the texture features of the candidate defect regions and the texture features of the normal skateboard region to obtain the texture deviation degree, and when the texture deviation degree is greater than the texture judgment threshold and the area of ​​the candidate defect region is within the preset minimum identifiable area and maximum coverage area range, it is judged as a foreign object attachment defect.

[0013] According to the above technical solution, the discrimination logic of skateboard bending defects in step S3 is as follows: perform a quadratic polynomial fitting on the longitudinal centerline of the skateboard, calculate the curvature function of the centerline through the curvature calculation formula to characterize the bending trend strength of the skateboard, construct the structural response weight function by combining low-resolution large-scale features, and then obtain the comprehensive energy function of skateboard bending. When the value of the bending comprehensive energy function is greater than the bending judgment threshold, it is judged as a bending defect of the skateboard. The judgment of the skateboard fracture defect must simultaneously meet the following conditions: the number of connected components of the skateboard region mask is greater than 1, the length of the maximum discontinuity gap between adjacent connected components is greater than the structural continuity allowable threshold, and the average gradient intensity of the fused feature map in the neighborhood of the fracture gap is greater than the upper bound of the normal skateboard feature gradient statistics.

[0014] According to the above technical solution, the risk quantification method of a single defect in step S4 is as follows: the key discrimination quantity corresponding to each defect is mapped to the basic risk value through a monotonically increasing normalization function, and the defect type weight coefficient and spatial location weight coefficient are introduced. If the defect area is in a high-risk area, the spatial location weight coefficient is taken as a value greater than 1. The basic risk value is multiplied by the spatial location weight coefficient and the defect type weight coefficient in sequence to obtain the comprehensive risk value of a single defect. If the defective area is not located in a high-risk area, the spatial location weight coefficient is set to 1, and the basic risk value is obtained by multiplying the defect type weight coefficient.

[0015] According to the above technical solution, the specific implementation of the multi-factor coupled defect risk comprehensive assessment in step S4 is as follows: The total risk score of the skateboard is obtained by summing the comprehensive risk values ​​of all individual defects; the defect spatial distribution entropy is introduced to adaptively adjust the basic alarm threshold to obtain the dynamic alarm threshold. The defect spatial distribution entropy is shown in the following formula: in, Let be the probability of the area proportion of the i-th defect region; The dynamic alarm threshold is shown in the following formula: in, Basic alarm threshold, This is the entropy adjustment coefficient; A hierarchical and orderly set of risk thresholds is constructed. The total risk score is compared with the dynamic alarm threshold and the hierarchical threshold, and mapped to low, medium and high risk levels. Low risk outputs the operation and maintenance decision of normal operation and only periodic monitoring. Medium risk outputs the operation and maintenance decision of key attention or planned maintenance. High risk outputs the operation and maintenance decision of emergency alarm and triggering immediate action.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Compared with existing pantograph sliding plate defect detection methods, this invention possesses significant technical advantages and engineering application value. The overall detection and evaluation system features high precision, low false alarms, strong interpretability, and high engineering practicality. This invention achieves multi-scale feature extraction through a feature pyramid network equipped with an attention module, and constructs scale-adaptive fusion weights $\alpha_k$ based on mutual information to adaptively enhance defect-sensitive scale features. This effectively solves the problem of missed detections and false judgments caused by the large scale span of sliding plate fracture defects. Simultaneously, through the dual spatial constraints of sliding plate area masks and high-risk areas, interference from complex backgrounds is reduced, significantly improving the detection accuracy and scale adaptability of various defects. This invention introduces prior structural constraints on pantograph sliding plates and effectively eliminates false detections in non-sliding plate areas through geometric proportion consistency correction. Simultaneously, it designs differentiated decoupled discrimination logic based on the physical mechanisms of three types of defects, deeply binding defect discrimination criteria to the structural and physical characteristics of the sliding plate. This effectively distinguishes between surface reflections, stains, and genuine defects, significantly reducing the false alarm rate and improving the robustness of the detection results. Furthermore, each discrimination index and evaluation parameter has a clear physical meaning and engineering interpretability, solving the problem of weak interpretability in existing methods.

[0017] This invention constructs a multi-factor coupled defect risk assessment system that comprehensively considers factors of high concern to on-site operation and maintenance, such as defect type, spatial location, and distribution characteristics. It achieves quantitative scoring of defect risk through weighting coefficients and adaptive threshold adjustment. Furthermore, the tiered risk alarm levels and targeted operation and maintenance decisions can directly connect to the intelligent operation and maintenance system of high-speed railways, realizing integration from defect detection to operation and maintenance decision-making, significantly improving the engineering practicality of the method. In addition, the overall technical solution of this invention is developed based on images acquired by the 5C device in rail transit, making it compatible with the hardware equipment of existing rail transit safety monitoring systems without large-scale hardware modifications. It features low cost and easy deployment. Moreover, each technical module of the method adopts a modular design, allowing for flexible adjustment and parameter optimization according to different line conditions, operating speeds, and maintenance strategies. It possesses good versatility and scalability, and can be widely applied to pantograph status monitoring of different types of high-speed trains, providing reliable technical support for high-speed railway operation safety. Attached Figure Description

[0018] Figure 1 This is a flowchart of the analysis method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1 As shown, a defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways is presented. This method uses high-definition pantograph images acquired by the 5C device as input and achieves the detection, location, and risk assessment of pantograph sliding plate defects through four core steps. The overall technical solution forms an integrated technical process from effective area extraction, multi-scale feature fusion, accurate defect identification, to risk quantification assessment. Specifically, it includes: First, acquiring the original pantograph image, locating and structurally extracting the sliding plate area from the original image, and obtaining the sliding plate area mask and high-risk areas of the pantograph sliding plate working contact surface through precise region modeling, thus defining the effective area for subsequent feature extraction and defect identification and reducing interference from complex backgrounds; Second, under the dual spatial constraints of the sliding plate area mask and high-risk areas, multi-scale feature extraction is performed on the original pantograph image, combined with information theory to construct... A multi-scale feature adaptive fusion model is employed to weight and enhance defect features at different scales, resulting in a unified defect fusion feature map that comprehensively represents defects of different scales and types. Next, based on the physical mechanisms of three typical defects in the pantograph sliding plate, the defect fusion feature map is decoupled and discriminated. Differentiated discrimination logic is designed for the different physical characteristics of the three types of defects: foreign object attachment, sliding plate bending, and sliding plate breakage, enabling accurate identification and location of these three types of defects. Finally, combining defect type, defect spatial location, and high-risk area features of the working contact surface, a multi-factor coupled comprehensive defect risk assessment is performed on the defect identification results. By introducing weight coefficients and an adaptive threshold adjustment mechanism, a quantitative score for defect risk is achieved, and the pantograph sliding plate defect risk alarm level and corresponding maintenance decisions that match actual operation and maintenance needs are output.

[0021] Based on the aforementioned core technical solutions, this invention optimizes and refines the technical implementation of each step. Specifically, the skateboard region localization and structure extraction achieve precise localization of the skateboard region through a deep learning model with a detection-segmentation joint structure. This model can simultaneously output the bounding box parameters and pixel-level mask of the skateboard region, balancing localization efficiency and accuracy. Furthermore, based on the skateboard region mask, the geometric parameters of the skateboard are calculated, and a structural consistency constraint term is constructed to correct the localization results. The formula for the structural consistency constraint term is: in These are the proportional parameters of the skateboard structure. To establish a predefined standard aspect ratio based on the standard pantograph sliding plate structure, this constraint effectively eliminates false detections in non-sliding plate areas, improving the accuracy of area localization. The sliding plate structure ratio parameter is the ratio of the projected length to the projected width of the sliding plate in the image coordinate system. Both the projected length and projection width are calculated based on the extreme values ​​of the horizontal and vertical coordinates of the effective pixels in the sliding plate area mask. When the structural consistency constraint term... When the result exceeds the preset geometric consistency tolerance threshold, the location result of the skateboard area is discarded or corrected a second time to further ensure the effectiveness of the location result.

[0022] The construction of high-risk areas on the working contact surface is based on the actual working characteristics of the skateboard. This area is the direct contact area between the skateboard and the overhead contact line, and is a high-risk area for defects. Specifically, the longitudinal centerline of the skateboard is extracted based on the skateboard area mask. Set width control parameters on both sides of the centerline that match the actual working contact surface width of the slide. The absolute value of the difference between the ordinate of a pixel and its corresponding centerline position is not greater than 1. The set of pixels is identified as a high-risk area. This method can accurately delineate the core area of ​​defect detection, improving the targeting and efficiency of subsequent feature extraction and defect identification.

[0023] Multi-scale feature extraction is achieved through a feature pyramid network equipped with a channel-spatial-global attention module. This network can extract multi-scale features at different downsampling rates, and the attention module can enhance key features of the skateboard region and suppress background interference. The extracted multi-scale features include at least high-resolution small-scale features representing minute foreign objects and microcracks. Mesoscale features characterizing the initial bending characteristics of a skateboard Low-resolution, large-scale features characterizing overall bending and large-size fractures This enables comprehensive capture of defects at different scales; simultaneously, it constructs scale-adaptive fusion weights based on mutual information in information theory. in For scale indexing, Random variables for characteristics at each scale and defect categories The mutual information is used to quantify the correlation between features at each scale and the defect category. The stronger the correlation, the higher the corresponding fusion weight, thus achieving adaptive enhancement of features at defect-sensitive scales. Finally, a defect fusion feature map is obtained through adaptive fusion of multi-scale features. The calculation formula is as follows: Where s, m, and l are small-scale, medium-scale, and large-scale indices, respectively. The adaptive fusion weights are for each scale. Feature maps extracted at each scale.

[0024] Defect decoupling discrimination is based on differentiated discrimination logic designed according to the different physical mechanisms of three types of defects. For foreign object attachment defects, discrimination is achieved by generating candidate defect regions from the fused feature map through saliency response analysis. Then, the Euclidean norm of the texture features of the candidate defect region and the texture features of the normal skateboard region is calculated to obtain the texture deviation. When the texture deviation exceeds a preset texture judgment threshold, and the area of ​​the candidate defect region is within the preset minimum identifiable area and maximum coverage area range based on the imaging resolution and the physical size of the skateboard, it is determined to be a foreign object attachment defect. This discrimination logic is based on the difference in texture features between foreign object attachment and normal skateboard, and has a clear physical meaning. Skateboard bending defects are discerned by analyzing the longitudinal centerline of the skateboard. Perform a quadratic polynomial fitting and combine it with the curvature calculation formula to calculate the curvature function of the centerline. Characterizes the bending tendency intensity of the skateboard while incorporating low-resolution, large-scale features. Constructing the structural response weight function This leads to the derive of a comprehensive energy function for skateboard bending that reflects both the degree of skateboard bending and structural anomalies. When the combined bending energy function value exceeds the preset bending judgment threshold When a skateboard is bent, the resulting defect is identified. This discrimination logic closely follows the physical characteristics of the geometric deformation of a skateboard bend, effectively identifying continuous bend defects without obvious edge breaks. The discrimination of skateboard fracture defects is verified from two dimensions: topological structure and feature continuity. It requires that three conditions be met simultaneously: the number of connected components in the skateboard region mask is greater than 1, the length of the maximum discontinuity gap between adjacent connected components is greater than the allowable threshold of structural continuity, and the average gradient intensity of the fused feature map in the neighborhood of the fracture gap is greater than the upper bound of the statistical gradient of normal skateboard features. This discrimination logic is based on the physical characteristics of the structural continuity disruption caused by skateboard fracture, effectively distinguishing between real fractures and false defects caused by image noise, thus improving the accuracy of fracture defect discrimination.

[0025] The multi-factor coupled defect risk comprehensive assessment is carried out from two levels: single defect risk quantification and multi-defect fusion assessment. The risk quantification of single defects maps the key discrimination parameters corresponding to each defect to a basic risk value with a unified scoring range through a monotonically increasing normalization function. Then, defect type weight coefficients that can reflect the degree of danger of different defects are introduced. Spatial location weighting coefficient that reflects the risk of defect location The basic risk value is weighted and adjusted. If the defective area is located in a high-risk area, the spatial location weight coefficient is adjusted. Take a value greater than 1, and multiply the base risk value by the spatial location weighting coefficient in sequence. Defect type weighting coefficient Obtain the comprehensive risk value of a single defect If the defective area is not located in a high-risk area, the spatial location weighting coefficient... Set to 1, the base risk value is multiplied only by the defect type weighting coefficient. The single-defect comprehensive risk value is obtained, enabling precise quantification of single-defect risk; the multi-defect fusion assessment, on the other hand, sums up the comprehensive risk values ​​of all single defects on the same skateboard to obtain the total risk score of the skateboard. At the same time, the spatial distribution entropy of defects is introduced. The basic alarm threshold is adaptively adjusted, whereby... The probability of the area proportion of the i-th defect region is given. The spatial distribution entropy of defects can effectively characterize the spatial distribution characteristics of defects. The more dispersed the distribution, the larger the entropy value, reflecting a higher risk of overall structural deterioration of the skateboard. The dynamic alarm threshold is obtained by adjusting the entropy value. in Basic alarm threshold, As the entropy adjustment coefficient, a hierarchical ordered risk threshold set based on historical fault statistics, industry standards and specifications, and on-site operation and maintenance experience is constructed, and the total risk score is then adjusted accordingly. With dynamic alarm threshold The system compares risk levels with thresholds and maps them to low, medium, and high risk levels. It then outputs matching maintenance decisions for each risk level. For low-risk levels, it outputs maintenance decisions for normal operation and only periodic monitoring. For medium-risk levels, it outputs maintenance decisions for key monitoring or planned maintenance. For high-risk levels, it outputs maintenance decisions for emergency alarms and triggers immediate action. This achieves integration from defect detection to maintenance decision-making.

[0026] Example 2 The defect analysis method of the present invention for the 5C safety detection and monitoring system for electrified railway power supply uses the original pantograph image captured by a high-definition industrial camera in the 5C device of rail transit as input. The image has a resolution of 2048×1080 pixels and a frame rate of 25fps, which can clearly capture the detailed features and structural morphology of the pantograph sliding plate. The specific implementation process of the method revolves around four core links: sliding plate area localization and structured extraction, multi-scale feature collaborative fusion model construction, defect decoupling and discrimination based on physical priors, and comprehensive defect risk assessment based on multi-factor coupling. Each link is progressive and synergistic, forming an integrated defect detection and risk assessment process. The specific implementation process is as follows: The first step involves locating and extracting the pantograph area. The core objective of this step is to accurately extract the effective area of ​​the pantograph from the complex raw pantograph images and delineate high-risk areas on the working contact surface where defects are frequent, thus limiting the effective analysis scope for subsequent steps. The raw pantograph images acquired by the 5C device are first obtained, denoted as [image 1]. Where H and W are the height and width of the image, respectively, and 3 represents the RGB three channels of the image. Subsequently, a deep learning localization model with a joint detection-segmentation structure is constructed. This model uses an object detection model as the base network and combines it with a U-Net segmentation network. It is trained using a labeled pantograph skateboard image dataset, which contains at least 10,000 images of normal skateboards and at least 10,000 images of each of three types of defects: foreign object attachment, skateboard bending, and skateboard breakage. The training process uses the AdamW optimizer, and the loss function is a weighted sum of the detection loss and the segmentation loss. After training, the model can simultaneously output the bounding box parameters of the skateboard region. With pixel-level skateboard area mask ,in The coordinates of the center of the bounding box are: The bounding box width and height, and the mask. The region with a pixel value of 1 is the effective area of ​​the skateboard, and the region with a pixel value of 0 is the background area. Therefore, the model output satisfies... .

[0027] Based on the obtained skateboard area mask Calculate the extreme values ​​of the horizontal and vertical coordinates of all valid pixels in the mask to obtain the projection length of the skateboard in the image coordinate system. With projection width Define the proportional parameters of the skateboard structure. At the same time, based on the actual structural parameters of the standard pantograph slider, a standard aspect ratio is predefined. Construct structural consistency constraints Set a geometric consistency tolerance threshold This threshold is determined using the 3σ principle based on the statistical results of the structural proportions of a large number of normal skateboard samples. If the location result is determined to be a false detection caused by background interference, it is discarded and a secondary location correction is triggered until a location result that meets geometric consistency constraints is obtained, thereby ensuring the accuracy of skateboard area location. After completing skateboard area location, based on the skateboard area mask... Extract the longitudinal centerline of the skateboard The center line is a mask. Each x-coordinate The corresponding ordinates of all valid pixels The average value can accurately characterize the overall longitudinal structure of the skateboard, and then width control parameters are set on both sides of the center line. Pixels, this parameter matches the actual working contact surface width of the pantograph slider, will meet The set of all pixels was identified as the high-risk area on the working contact surface of the skateboard. This area is the direct contact area between the skateboard and the overhead contact line, and it is a high-risk area for defects. This method enables the precise delineation of the core area for defect detection.

[0028] Next, a multi-scale feature collaborative fusion model is constructed. This step, under the dual spatial constraints of the skateboard region mask and high-risk regions, performs multi-scale feature extraction and adaptive fusion on the original image to obtain a fused feature map that comprehensively represents defects at different scales. First, the original pantograph image... With skateboard area mask Pixel-by-pixel multiplication is performed to obtain a masked image of the skateboard region, eliminating background interference. This image is then input into a feature pyramid network equipped with a channel-spatial-global attention module. The downsampling rates of this network are 2, 4, and 8, respectively, to extract high-resolution small-scale features. Mesoscale features Low-resolution large-scale features ,Right now: in Used to characterize small-scale defects such as tiny foreign objects and microcracks. Used for preliminary assessment of bending defects in skateboards. The attention module is used to characterize large-scale defects such as overall bending and large-size fractures. It enhances the network's ability to extract skateboard defect features by weighting and strengthening channel features, spatial features and global features.

[0029] To investigate the response characteristics of different defects across various scales, random variables relating each scale feature to the defect category are statistically analyzed based on training samples and historical normal samples. Mutual information between ,in Mutual information is used to quantify the correlation between features at each scale and defect categories. The stronger the correlation, the larger the mutual information value. Subsequently, scale-adaptive fusion weights are constructed based on the mutual information magnitude. The sum of the fusion weights for each scale is 1, achieving adaptive enhancement of features at defect-sensitive scales. Finally, based on the constructed scale-adaptive fusion weights, the feature maps of the three scales are weighted and fused to obtain the defect fusion feature map. This fused feature map can simultaneously represent defect features of different scales and types, providing comprehensive and accurate feature support for subsequent defect identification.

[0030] Then, a physical prior-based defect decoupling and discrimination process is performed. This step is based on the different formation mechanisms of three types of defects: foreign object adhesion, skateboard bending, and skateboard breakage, and analyzes the fused feature map. Differentiated decoupling discrimination is performed to achieve accurate identification and localization of three types of defects. First, saliency response analysis is conducted on the fused feature map to calculate the feature energy map. This energy map represents the spatial location of each channel in the fused feature map. The sum of the absolute values ​​of the response values ​​at the defective region indicates that, due to texture or structural abnormalities, the characteristic energy value is significantly higher than that of the normal region. Subsequently, based on the statistical results of the characteristic energy of a large number of historical normal skateboard samples, a dynamic threshold is set. Abnormal response regions are obtained through adaptive threshold segmentation. Connectivity analysis was performed to obtain a set of candidate defect regions. This completes the initial screening of candidate defect areas.

[0031] To identify foreign object attachment defects, all candidate defect areas in the skateboard area mask are first removed to obtain the normal skateboard area. The mean of the fused features within this region is calculated as the normal skateboard texture feature. Then for each candidate defect region The mean of their fusion features is calculated as the texture features of the candidate region. Then calculate and The Euclidean norm is used to obtain the texture deviation. Set texture detection threshold Based on the imaging resolution and the physical size of the skateboard, a minimum identifiable area for candidate defect regions is preset. Pixels, maximum coverage area Pixels, when and When the candidate region is identified as having a foreign object attachment defect, it is determined that the region is a foreign object attachment defect.

[0032] To identify skateboard bending defects, the longitudinal centerline of the skateboard extracted in the early stage is used as a reference. Perform a quadratic polynomial fitting to obtain the fitted curve. , where the coefficient The overall bending tendency of the skateboard is characterized, and then the curvature function of the fitted curve is calculated using the curvature calculation formula. It depicts the curvature of the center line at various points, while also capturing low-resolution, large-scale features. Project onto the centerline position to construct the structural response weight function. This function can reflect the degree of abnormality in the overall structure of the skateboard, combined with the curvature function. Constructing the comprehensive energy function of skateboard bending Set bending judgment threshold ,when At that time, it was determined that the skateboard had a bending defect.

[0033] To identify skateboard fracture defects, the first step is to mask the skateboard area. Perform connected component analysis to obtain the set of connected components. When the number of connected components At that time, it was determined that the skateboard had a topological structure split, which was a potential fracture defect. Subsequently, the adjacent connected domains were analyzed. Calculate the minimum Euclidean distance between its boundary pixels. Define the maximum distance between all adjacent connected components. The length of the fracture gap Set a structural continuity tolerance threshold Pixels, when At that time, feature discontinuity verification is performed, and a neighborhood of the fracture gap is constructed with the midpoint of the fracture gap as the center. Calculate the fused feature map within this neighborhood. Mean spatial gradient intensity Set an upper bound for the gradient statistics of normal skateboard features. ,when When the skateboard is found to have a fracture defect, the fracture defect is accurately identified through dual verification of topological structure and feature continuity.

[0034] Finally, a multi-factor coupled comprehensive defect risk assessment is conducted. This step, based on defect type identification and location, integrates multiple influencing factors to achieve a quantitative assessment of defect risk and outputs a tiered risk alarm level and operational decisions. First, for each identified defect, its corresponding key discrimination metric is determined. Among them, foreign matter adhesion defects are texture deviation. The bending defect of the skateboard is the maximum curvature value. The fracture defect of the sliding plate is the length of the fracture gap. Through the Sigmoid normalization function Will Mapped to A unified risk scoring range is used to obtain the basic risk value for a single defect. The sigmoid function, as a monotonically increasing normalization function, can achieve unified quantification of quantities with different dimensions.

[0035] Subsequently, based on the degree of impact of various defects on the safety of skateboard operation, a defect type weighting coefficient was set. Among them, foreign object adhesion defects Skateboard bending defects Slide plate fracture defects Fracture defects have the greatest impact on driving safety, so the weighting coefficient is set to the maximum value, while also taking into account the high-risk areas identified in the early stages. Set spatial location weight coefficient If the defective area ,but ,otherwise The higher risk of defects in high-risk areas is reflected by spatial location weighting coefficients, and then the comprehensive risk value of a single defect is calculated. .

[0036] The total risk score of the skateboard is obtained by summing the combined risk values ​​of all individual defects on the same skateboard. Simultaneously calculate the spatial distribution entropy of defects. ,in Let be the proportion of the area of ​​the i-th defect region to the total area of ​​all defect regions. A larger value indicates a more dispersed defect distribution and a higher risk of overall structural degradation of the skateboard. An entropy adjustment coefficient is then introduced. ,pass Basic alarm threshold Adaptive adjustment is performed to obtain dynamic alarm thresholds. .

[0037] Construct a hierarchical and ordered set of risk thresholds This threshold set is set based on statistical data of pantograph failures in high-speed railways, industry standards and specifications, and on-site operation and maintenance experience, and incorporates the total risk score. Compared with the threshold set, it is mapped to the corresponding risk level: when When the risk level is low, the system outputs "normal operation" status information, and only the pantograph contact plate is periodically monitored every 7 days; when... When the risk level is medium, a "key focus" or "planned maintenance" alert will be displayed, and manual review or partial maintenance is recommended within 3 days; when When the risk level is high, an "emergency alarm" or "immediate action" command is output, triggering the maintenance early warning mechanism and taking safety measures such as train speed limit and offline maintenance. This achieves the integration of defect detection and operation and maintenance decision-making, providing accurate basis for the intelligent operation and maintenance of high-speed railway pantographs.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0039] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways, characterized in that: Includes the following steps: Step S1: Obtain the original image of the pantograph, perform skateboard area localization and structure extraction on the image to obtain the skateboard area mask and the high-risk area of ​​the pantograph skateboard working contact surface; Step S2: Under the dual constraints of the skateboard area mask and the high-risk area, multi-scale feature extraction is performed on the original pantograph image, and a multi-scale feature adaptive fusion model is constructed to obtain the defect fusion feature map through the model; Step S3: Based on the physical mechanism of pantograph sliding plate defects, the defect fusion feature map is decoupled and identified to distinguish three types of defects: foreign object attachment, sliding plate bending, and sliding plate breakage. Step S4: Combining the defect type, defect spatial location, and high-risk areas of the working contact surface, perform a multi-factor coupled comprehensive assessment of defect risk based on the defect identification results, and output the risk alarm level of the pantograph sliding plate defect and the corresponding operation and maintenance decision.

2. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: The implementation method of skateboard region localization and structure extraction in step S1 is as follows: skateboard region localization is achieved through a deep learning model with a joint detection-segmentation structure. Based on the skateboard region mask, the geometric parameters of the skateboard are calculated, and a structural consistency constraint term is constructed to correct the localization result. The formula for the structural consistency constraint term is: in These are the proportional parameters of the skateboard structure. The standard aspect ratio for a predefined skateboard.

3. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 2, characterized in that: The proportional parameter of the skateboard structure is the ratio of the projected length to the projected width of the skateboard in the image coordinate system. Both the projected length and projection width are calculated based on the extreme values ​​of the horizontal and vertical coordinates of the effective pixels in the skateboard region mask; when If the result exceeds the geometric consistency tolerance threshold, the skateboard area positioning result will be discarded or corrected a second time.

4. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: The high-risk area of ​​the working contact surface in step S1 is constructed as follows: the longitudinal centerline of the skateboard is extracted based on the skateboard area mask, and width control parameters are set on both sides of the centerline. The absolute value of the difference between the ordinate of a pixel and its corresponding centerline position is not greater than 1. The set of pixels was identified as a high-risk area.

5. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: In step S2, multi-scale features are extracted by a feature pyramid network equipped with a channel-space-global attention module. The multi-scale features include at least high-resolution small-scale features representing tiny foreign objects and microcracks, medium-scale features representing the initial features of the skateboard bending, and low-resolution large-scale features representing the overall bending and large-size fractures. Scale-adaptive fusion weights based on mutual information: Where k is the scale index. The mutual information between scale features and defect category random variables is used to achieve adaptive fusion of multi-scale features through this weight.

6. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 5, characterized in that: The formula for calculating the defect fusion feature map obtained by multi-scale feature adaptive fusion in step S2 is as follows: Where s, m, and l are the small, medium, and large scale indices, respectively. The adaptive fusion weights are for each scale. Feature maps extracted at each scale.

7. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: The discrimination logic for foreign object attachment defects in step S3 is as follows: perform saliency response analysis on the defect fusion feature map to generate candidate defect regions, calculate the Euclidean norm of the texture features of the candidate defect regions and the texture features of the normal skateboard region to obtain the texture deviation degree, and when the texture deviation degree is greater than the texture judgment threshold and the area of ​​the candidate defect region is within the preset minimum identifiable area and maximum coverage area range, it is judged as a foreign object attachment defect.

8. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: The discrimination logic for skateboard bending defects in step S3 is as follows: perform a quadratic polynomial fitting on the longitudinal centerline of the skateboard, calculate the curvature function of the centerline using the curvature calculation formula to characterize the bending trend strength of the skateboard, construct the structural response weight function by combining low-resolution large-scale features, and then obtain the comprehensive energy function of skateboard bending. When the combined energy function value of bending exceeds the bending judgment threshold, it is judged as a bending defect of the skateboard. The criteria for identifying skateboard fracture defects must simultaneously satisfy the following: the number of connected components in the skateboard region mask is greater than 1, the length of the maximum discontinuity gap between adjacent connected components is greater than the allowable threshold for structural continuity, and the average gradient intensity of the fused feature map in the neighborhood of the fracture gap is greater than the upper bound of the statistical gradient of the normal skateboard feature.

9. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 1, characterized in that: In step S4, the risk quantification method for a single defect is as follows: the key discrimination quantity corresponding to each defect is mapped to the basic risk value through a monotonically increasing normalization function. The defect type weight coefficient and the spatial location weight coefficient are introduced. If the defect area is in a high-risk area, the spatial location weight coefficient is taken as a value greater than 1. The basic risk value is multiplied by the spatial location weight coefficient and the defect type weight coefficient in sequence to obtain the comprehensive risk value of a single defect. If the defective area is not located in a high-risk area, the spatial location weight coefficient is set to 1, and the basic risk value is obtained by multiplying the defect type weight coefficient.

10. The defect analysis method for the 5C power supply safety detection and monitoring system of electrified railways according to claim 9, characterized in that: The specific implementation of the multi-factor coupled defect risk comprehensive assessment in step S4 is as follows: The total risk score of the skateboard is obtained by summing the comprehensive risk values ​​of all individual defects; the defect spatial distribution entropy is introduced to adaptively adjust the basic alarm threshold to obtain the dynamic alarm threshold. The defect spatial distribution entropy is shown in the following formula: in, Let be the probability of the area proportion of the i-th defect region; The dynamic alarm threshold is shown in the following formula: in, Basic alarm threshold, This is the entropy adjustment coefficient; A hierarchical and orderly set of risk thresholds is constructed. The total risk score is compared with the dynamic alarm threshold and the hierarchical threshold, and mapped to low, medium and high risk levels. Low risk outputs the operation and maintenance decision of normal operation and only periodic monitoring. Medium risk outputs the operation and maintenance decision of key attention or planned maintenance. High risk outputs the operation and maintenance decision of emergency alarm and triggering immediate action.

Images