A high-speed rail catenary completion acceptance method and system based on 4C detection

Abstract

The application relates to the technical field of high-speed rail operation and maintenance, and discloses a high-speed rail overhead line system completion acceptance method and system based on 4C detection, which comprises the following steps: collecting overhead line system image data, acquiring pose data and spatial positioning data of an inspection vehicle, and performing synchronous calibration processing to establish a mapping relationship and obtain calibrated overhead line system images; inputting the calibrated overhead line system images into a multi-scale target detection model to output detection results, performing pose correction on the detection results, combining the mapping relationship, and generating three-dimensional positioning coordinates; constructing a time sequence feature vector, inputting the time sequence feature vector into a prediction model, outputting a defect evolution trend and a risk probability of a component, and grading the risk probability; according to the risk grade and the spatial positioning data, combining a knowledge graph driven rule engine, generating an intelligent work order mapping to field equipment, collecting maintenance results, and feeding back. The application realizes the improvement of overhead line system component identification and spatial positioning accuracy, and improves the safety and reliability of high-speed rail overhead line system completion acceptance.

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 the final acceptance of high-speed rail catenary based on 4C detection. Background Technology

[0002] The overhead contact system is a key component of the high-speed railway power supply system, and its installation quality and operational status directly affect train operation safety and power supply reliability. The final acceptance inspection of the overhead contact system is a crucial step in ensuring the safe operation of newly built or renovated high-speed railway lines, playing a decisive role in ensuring the system meets technical standards and identifying potential safety hazards.

[0003] Existing technologies employ traditional image processing algorithms, single-sensor data acquisition, and static parameter determination to achieve basic identification of overhead contact system components, measurement of some geometric parameters, and simple anomaly detection. However, they lack multi-source data fusion capabilities, lack time-series intelligent analysis, cannot dynamically adjust evaluation criteria, and lack knowledge-driven closed-loop optimization. This results in a large number of potential defects not being detected in a timely manner during the acceptance process, one-sided system status assessment, inability to predict defect evolution trends, and low efficiency in maintenance resource allocation. Ultimately, this makes it difficult to fully guarantee the safety and reliability of the high-speed rail overhead contact system, increasing operation and maintenance costs and safety risks. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for the final acceptance of high-speed railway catenary based on 4C detection. This method solves the problems of existing technologies lacking multi-source data fusion capabilities, lacking time-series intelligent analysis, being unable to dynamically adjust evaluation criteria, and lacking knowledge-driven closed-loop optimization. These problems result in a large number of potential defects not being detected in a timely manner during the acceptance process, one-sided system status assessment, inability to predict defect evolution trends, and low efficiency in the allocation of maintenance resources. Ultimately, these problems make it difficult to fully guarantee the safety and reliability of the high-speed railway catenary system, increasing operation and maintenance costs and safety risks.

[0005] The technical solution of this invention is implemented as follows: In a first aspect, this invention provides a method for the final acceptance of high-speed railway catenary based on 4C testing, comprising the following steps:

[0006] The system acquires contact network image data using a high-speed camera array, obtains the pose data and spatial positioning data of the inspection vehicle, performs synchronous calibration processing, establishes the mapping relationship between the image coordinate system and the track coordinate system, and obtains the calibrated contact network image.

[0007] The calibration catenary image is input into a deep learning-based multi-scale target detection model, which outputs the detection results. The detection results are then corrected for pose. Based on the mapping relationship, the three-dimensional positioning coordinates of the components in the track coordinate system are generated. The detection results are then denoised using a confidence-weighted filtering algorithm to remove low-confidence false detection samples.

[0008] The detection results are spatiotemporally aligned with historical detection data to construct a temporal feature vector indexed by the component number. The temporal feature vector is then input into a prediction model combining a temporal convolutional network and a long short-term memory network to output the defect evolution trend and risk probability of the component. The risk probability is then graded based on a dynamic Bayesian thresholding algorithm.

[0009] Based on risk level and spatial positioning data, combined with a knowledge graph-driven rule engine, intelligent work orders are generated, which include maintenance paths, priority rankings, and spare parts lists. The intelligent work orders are mapped to field equipment through augmented reality navigation devices, and maintenance results are collected and fed back to the digital twin database.

[0010] Based on the above technical solutions, preferably, the step of acquiring contact network image data based on a high-speed camera array, obtaining the pose data and spatial positioning data of the inspection vehicle, performing synchronous calibration processing, establishing a mapping relationship from the image coordinate system to the track coordinate system, and obtaining a calibrated contact network image includes:

[0011] The inspection vehicle travels at a constant speed along the railway line. A circular array containing multiple high-speed cameras is configured on the onboard 4C inspection device to record the six degrees of freedom motion data of the inspection vehicle body. Centimeter-level spatial positioning data of the inspection vehicle is recorded through the global navigation satellite system.

[0012] The collected six-degree-of-freedom motion data and spatial positioning data are aligned based on timestamps, and an intrinsic and extrinsic parameter matrix between multi-view high-speed cameras is constructed. Based on the real-time attitude of the vehicle body, the transformation relationship from the high-speed camera coordinate system to the track coordinate system is calculated to eliminate the influence of vehicle body vibration and shaking on image position accuracy, thus obtaining the calibration catenary image.

[0013] Based on the above technical solutions, preferably, the step of inputting the calibrated contact network image into a deep learning-based multi-scale target detection model, outputting detection results, correcting the pose of the detection results, generating the three-dimensional positioning coordinates of the components in the track coordinate system by combining the mapping relationship, and denoising the detection results using a confidence-weighted filtering algorithm to remove low-confidence false detection samples includes:

[0014] A multi-scale target detection model is used to identify and calibrate contact network images. The multi-scale target detection model includes a lightweight backbone network, a multi-scale attention feature pyramid, and a decoupled Transformer decoder. The lightweight backbone network is used to extract multi-level features of the image. The multi-scale attention feature pyramid is used to fuse features at different scales and enhance the representation ability of small targets. The decoupled Transformer decoder is used to decode the category, bounding box, and confidence of the parts in parallel.

[0015] The pose compensation matrix at the time of image acquisition of the catenary is calculated using inertial measurement unit data. The attitude of the detected two-dimensional coordinates of the catenary components is corrected. Based on the type of catenary components and the distribution of feature points, combined with the camera imaging model, the corrected two-dimensional coordinates are projected onto the track coordinate system to calculate the three-dimensional spatial coordinates of the catenary components. The confidence-weighted filtering algorithm is used for multiple detection results of the same catenary component. Based on the detection confidence, geometric consistency and temporal stability, the optimal fusion result is obtained and false detection samples with high confidence are eliminated.

[0016] Based on the above technical solutions, preferably, the multi-scale target detection model includes:

[0017] The lightweight backbone network adopts an improved lightweight version of the convolutional neural network architecture, introducing deformable convolutions to replace standard convolutional layers, thereby enhancing the network's ability to perceive deformable and small targets while maintaining the same computational cost. The multi-scale attention feature pyramid includes a channel attention module and a spatial adaptive aggregation module. The channel attention module is used to learn the importance weights of each channel, and the spatial adaptive aggregation module uses a learnable kernel function to dynamically fuse multi-scale features. The decoupled Transformer decoder includes three parallel branches, which are used for class prediction, position regression, and shape estimation, respectively.

[0018] Based on the above technical solutions, preferably, the pose compensation matrix and the confidence-weighted filtering algorithm include:

[0019] A temporal graph attention network is used to model the acceleration and angular velocity data collected by the inertial measurement unit (IMU) to construct a temporal state graph. The vehicle vibration mode of the inspection vehicle is extracted through graph convolution and temporal self-attention mechanism to obtain a continuous attitude estimation sequence. Based on the temporal relationship between the IMU sampling frequency and the camera frame rate, cubic spline interpolation is used to calculate the pose compensation matrix corresponding to each frame of the image to eliminate the influence of vehicle body shaking on the detection results. For the same contact network component detected, a temporal fusion tree of multi-frame detection results is constructed. Each node of the temporal fusion tree contains the component's category, position, size, and confidence information. The optimal state estimate is calculated using an adaptive Bayesian recursive estimation method. Based on the geometric consistency verification mechanism, the spatial relationship constraints between components and structural prior knowledge are used to eliminate false detection results that violate physical rules.

[0020] Based on the above technical solutions, preferably, the step of aligning the detection results with historical detection data in time and space, constructing a temporal feature vector indexed by the component number, inputting the temporal feature vector into a prediction model combining a temporal convolutional network and a long short-term memory network, outputting the defect evolution trend and risk probability of the component, and classifying the risk probability based on a dynamic Bayesian thresholding algorithm includes:

[0021] Spatiotemporal alignment of component inspection data collected from multiple inspections is performed. A spatial index is established based on global navigation satellite system positioning data and component numbers. Data from multiple time periods is time-aligned using a dynamic time warping algorithm. A time-series feature database is constructed based on the unique identifier of each component. Each record includes component type, spatial location, inspection time, appearance features, geometric parameters, environmental conditions, and historical maintenance records. The state change sequence of each component at different time points is extracted using the sliding time window method to form a multi-dimensional time-series feature vector.

[0022] A multi-dimensional temporal feature vector is input into a prediction model to predict defect trends. The prediction model includes a multi-layer temporal convolutional network, a long short-term memory network, and an attention mechanism.

[0023] The multi-layer temporal convolutional network is used to extract temporal pattern features, the long short-term memory network is used to capture long-term dependencies, the attention mechanism is used to highlight key state changes, and the prediction model is used to output the defect evolution trend and predicted risk probability of each component. A risk quantification index is established, and a dynamic Bayesian threshold algorithm is used to adaptively classify the predicted risk probability in combination with changes in environmental conditions, load levels, and maintenance cycles, generating four-level risk labels, including: R1, R2, R3, and R4, where R1 is the normal level risk label, R2 is the attention level risk label, R3 is the warning level risk label, and R4 is the danger level risk label.

[0024] Based on the above technical solutions, preferably, the step of generating intelligent work orders containing maintenance paths, priority rankings, and spare parts lists based on risk level and spatial positioning data, combined with a knowledge graph-driven rule engine, and mapping the intelligent work orders to field equipment via an augmented reality navigation device, and collecting maintenance results and feeding them back to the digital twin database, includes:

[0025] Construct a knowledge graph for the overhead contact system, which includes spatial relationships, functional dependencies, fault propagation modes, and maintenance specifications among components;

[0026] Based on the overhead contact network knowledge graph and risk level data, a combined optimization algorithm is used to calculate the optimal maintenance path and timing arrangement.

[0027] For overhead contact line components with R3 and R4 risk levels, generate smart work orders that include detailed maintenance steps, required tools, spare parts lists, and quality acceptance standards.

[0028] By using augmented reality navigation devices, work order information is spatially matched with on-site equipment, and maintenance guidance information is projected into the field of view of maintenance personnel in real time, providing visual operation guidance;

[0029] The maintenance process data and maintenance results are collected based on mobile terminals. The maintenance process data includes operation time and operation records, and the maintenance results include replacement component information and quality test results.

[0030] Maintenance feedback data is uploaded to the digital twin database in real time to update the catenary status and adjust model parameters;

[0031] An adaptive incremental learning mechanism is constructed to dynamically update the model parameters of the multi-scale target detection model and the prediction model for novel defect samples and false detection samples.

[0032] Secondly, the present invention also provides a high-speed railway catenary completion acceptance system based on 4C detection, the system comprising:

[0033] The data acquisition module is used to acquire contact network image data based on a high-speed camera array, obtain the pose data and spatial positioning data of the inspection vehicle, perform synchronous calibration processing, establish the mapping relationship between the image coordinate system and the track coordinate system, and obtain the calibrated contact network image.

[0034] The multi-scale target detection module is used to input the calibrated catenary image into the deep learning-based multi-scale target detection model, output the detection result, perform pose correction on the detection result, generate the three-dimensional positioning coordinates of the component in the track coordinate system by combining the mapping relationship, and perform noise reduction processing on the detection result through the confidence weighted filtering algorithm to remove low confidence false detection samples.

[0035] The risk grading module is used to align the detection results with historical detection data in time and space, construct a temporal feature vector indexed by the part number, input the temporal feature vector into a prediction model combining a temporal convolutional network and a long short-term memory network, output the defect evolution trend and risk probability of the parts, and grade the risk probability based on the dynamic Bayesian threshold algorithm.

[0036] The acceptance feedback module is used to generate intelligent work orders that include maintenance paths, priority rankings, and spare parts lists based on risk levels and spatial positioning data, combined with a knowledge graph-driven rule engine. The intelligent work orders are mapped to on-site equipment through an augmented reality navigation device, and the maintenance results are collected and fed back to the digital twin database.

[0037] Thirdly, the present invention also provides an electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus;

[0038] 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 high-speed rail catenary completion acceptance method based on 4C detection.

[0039] Fourthly, the present invention also provides a computer-readable storage medium storing computer instructions that enable a computer to perform steps such as a method for the final acceptance of a high-speed railway catenary based on 4C detection.

[0040] The high-speed railway catenary completion acceptance method and system based on 4C detection of the present invention has the following advantages over the prior art:

[0041] (1) By integrating multi-sensor synchronous calibration, high-speed camera array data acquisition and deep learning multi-scale target detection model, and combining time-series feature vectors to construct a defect prediction mechanism of time-series convolutional network and long short-term memory network, and adopting dynamic Bayes threshold algorithm, knowledge graph-driven rule engine, augmented reality navigation and digital twin feedback mechanism, the accuracy of contact network component identification and spatial positioning, dynamic performance evaluation and defect evolution trend prediction, as well as adaptive adjustment of acceptance standards and optimized allocation of maintenance resources have been achieved, thus improving the safety, reliability and work efficiency of high-speed rail contact network completion acceptance.

[0042] (2) By adopting a multi-scale target detection model with a lightweight backbone network, a multi-scale attention feature pyramid and a decoupled Transformer decoder, combined with the inertial measurement unit data processed by the temporal graph attention network and the pose compensation matrix calculated by cubic spline interpolation, and the confidence weighted filtering algorithm constructed by the adaptive Bayesian recursive estimation and geometric consistency verification mechanism, the detection accuracy and spatial positioning accuracy of the contact wire components are improved, and the false detection rate caused by environmental interference and vehicle body vibration is significantly reduced.

[0043] (3) By introducing a lightweight convolutional neural network architecture and a channel attention and spatial adaptive aggregation module in deformable convolution and multi-scale attention feature pyramid network, as well as three parallel branches in the decoupled Transformer decoder: category prediction, position regression and shape estimation, combined with multi-task learning with focus loss function and adaptive weight balancing strategy, the speed of target detection and the robustness to occluded and deformed targets are improved under the condition of limited computing resources, so that even under the condition of imbalanced data, a high recognition rate can be maintained for rare but critical component defects. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is a flowchart of a high-speed railway catenary completion acceptance method based on 4C detection according to the present invention;

[0046] Figure 2 This is a structural diagram of a high-speed rail overhead contact line completion acceptance system based on 4C detection, according to the present invention. Detailed Implementation

[0047] 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.

[0048] Please see Figure 1 This invention provides a method for the final acceptance of high-speed railway catenary based on 4C testing, comprising the following steps:

[0049] The system acquires contact network image data using a high-speed camera array, obtains the pose data and spatial positioning data of the inspection vehicle, performs synchronous calibration processing, establishes the mapping relationship between the image coordinate system and the track coordinate system, and obtains the calibrated contact network image.

[0050] The calibration catenary image is input into a deep learning-based multi-scale target detection model, which outputs the detection results. The detection results are then corrected for pose. Based on the mapping relationship, the three-dimensional positioning coordinates of the components in the track coordinate system are generated. The detection results are then denoised using a confidence-weighted filtering algorithm to remove low-confidence false detection samples.

[0051] The detection results are spatiotemporally aligned with historical detection data to construct a temporal feature vector indexed by the component number. The temporal feature vector is then input into a prediction model combining a temporal convolutional network and a long short-term memory network to output the defect evolution trend and risk probability of the component. The risk probability is then graded based on a dynamic Bayesian thresholding algorithm.

[0052] Based on risk level and spatial positioning data, combined with a knowledge graph-driven rule engine, intelligent work orders are generated, which include maintenance paths, priority rankings, and spare parts lists. The intelligent work orders are mapped to field equipment through augmented reality navigation devices, and maintenance results are collected and fed back to the digital twin database.

[0053] Specifically, this embodiment integrates multi-sensor synchronous calibration, high-speed camera array data acquisition, and deep learning multi-scale target detection model. It combines temporal feature vectors to construct a defect prediction mechanism using temporal convolutional networks and long short-term memory networks. Furthermore, it employs dynamic Bayesian thresholding algorithms, knowledge graph-driven rule engines, augmented reality navigation, and digital twin feedback mechanisms. This achieves improved accuracy in identifying and spatially positioning contact network components, dynamic performance evaluation and defect evolution trend prediction, adaptive adjustment of acceptance standards, and optimized allocation of maintenance resources. Ultimately, it enhances the safety, reliability, and efficiency of high-speed railway contact network completion acceptance.

[0054] The process of acquiring contact network image data based on a high-speed camera array, obtaining the pose data and spatial positioning data of the inspection vehicle, and performing synchronous calibration processing to establish a mapping relationship between the image coordinate system and the track coordinate system, thereby obtaining a calibrated contact network image, includes:

[0055] The inspection vehicle travels at a constant speed along the railway line. A circular array of eight high-speed cameras is configured on the onboard 4C inspection device. The field of view of each high-speed camera is set to 60°, the overlap of the field of view of adjacent high-speed cameras is set to 15% to 20%, and the shooting frequency of the high-speed cameras is set to above 8000fps. The vehicle's six degrees of freedom motion data is recorded at a sampling frequency of 200Hz based on the inertial measurement unit, and centimeter-level spatial positioning data is recorded at a sampling frequency of 10Hz through the global navigation satellite system.

[0056] In one specific embodiment, the high-speed camera is specifically configured as follows:

[0057] The high-speed camera adopts a global shutter mode, and the exposure time can be adaptively adjusted within the range of 8μs to 50μs. Each camera is equipped with a synchronously triggered structured light projection unit, and the structured light pattern is a pseudo-random dot matrix with a dot density of no less than 1000 dots / m². Under different lighting conditions, an exposure adaptive compensation algorithm is used to dynamically adjust the flash power and camera exposure parameters, so that the average image brightness is maintained within the range of 128±15 grayscale values, ensuring stable image quality in tunnels, areas with alternating light and dark, and strong lighting environments. When more than one camera captures the same component, feature point matching based on the overlapping area of ​​the camera's field of view is used to achieve cross-camera image stitching and 3D reconstruction.

[0058] Using a time synchronization processing module, the collected six-degree-of-freedom motion data and spatial positioning data are aligned based on timestamps, and an intrinsic and extrinsic parameter matrix between multi-view high-speed cameras is constructed. Based on the real-time attitude of the vehicle body, the transformation relationship from the high-speed camera coordinate system to the track coordinate system is calculated to eliminate the influence of vehicle body vibration and shaking on image position accuracy, ensuring that the position coordinate measurement accuracy of components is better than ±5mm, and obtaining the calibrated contact wire image.

[0059] In one specific embodiment, the calculation of the transformation relationship from the camera coordinate system to the orbit coordinate system includes:

[0060] A temporal self-attention mechanism is employed to process the six-axis data of the inertial measurement unit, constructing an attitude estimation model, extracting instantaneous vibration characteristics of the vehicle body, and generating a continuous pose compensation matrix. Real-time dynamic positioning data from the Global Navigation Satellite System is matched with the overhead contact line digital map to determine the precise position of the detection system in the global coordinate system, while simultaneously marking the support pole number and positioning device number. Spatiotemporal state estimation and fusion of multi-source data are performed based on a Kalman filter to address sensor data delay, packet loss, and noise interference issues. A deep neural network-assisted bundle adjustment optimization algorithm is used to globally optimize the relative positional relationships between multi-view cameras and the image-track coordinate system transformation parameters, ensuring that the reprojection error of the three-dimensional position of components is less than 1.5 pixels.

[0061] Specifically, this embodiment achieves a four-stage process for the completion and acceptance of the high-speed railway catenary through multi-source data acquisition coordinated by a high-speed camera array and an inertial measurement unit, high-precision synchronous calibration and mapping construction, Kalman filter multi-source data fusion, and bundle adjustment optimization assisted by a deep neural network. It also utilizes a high-speed camera collaborative shooting mechanism equipped with a synchronously triggered structured light projection unit and an exposure adaptive compensation algorithm. This results in a component 3D position measurement accuracy of ±5mm, a reprojection error of less than 1.5 pixels, and stable image quality acquisition capabilities under various complex environmental conditions, such as tunnels, areas with alternating light and dark zones, and strong lighting environments. This enhances the reliability of the data foundation for the completion and acceptance of the high-speed railway catenary.

[0062] The process involves inputting the calibrated contact network image into a deep learning-based multi-scale target detection model, outputting detection results, correcting the pose of the detection results, generating the three-dimensional positioning coordinates of the components in the track coordinate system based on mapping relationships, and denoising the detection results using a confidence-weighted filtering algorithm to remove low-confidence false detection samples, including:

[0063] A multi-scale target detection model is used to identify and calibrate contact wire images. The multi-scale target detection model includes a lightweight backbone network, a multi-scale attention feature pyramid, and a decoupled Transformer decoder. The lightweight backbone network is used to extract multi-level features of the image. The multi-scale attention feature pyramid is used to fuse features of different scales and enhance the representation ability of small targets. The decoupled Transformer decoder is used to decode the category, bounding box, and confidence of the parts in parallel, realizing real-time detection of no less than 28 types of parts such as droppers, locators, and bolts.

[0064] The pose compensation matrix at the time of image acquisition of the catenary is calculated using inertial measurement unit data. The attitude of the detected two-dimensional coordinates of the catenary components is corrected. Based on the type of catenary components and the distribution of feature points, combined with the camera imaging model, the corrected two-dimensional coordinates are projected onto the track coordinate system to calculate the three-dimensional spatial coordinates of the catenary components. The confidence-weighted filtering algorithm is used for multiple detection results of the same catenary component. Based on the detection confidence, geometric consistency and temporal stability, the optimal fusion result is obtained and false detection samples with high confidence are eliminated.

[0065] In one specific embodiment, the calculation formula for the multi-scale attention feature fusion is: ;

[0066] in, For multi-scale feature fusion feature maps, For the first i Feature layer, The number of feature layers, For the first iChannel attention weights of the feature layer For the first i The learnable transformation matrix of the feature layer. This is a spatial adaptive aggregation operation. For the first i Dynamic aggregation kernel of layer feature layer, It is the first balance factor;

[0067] The formula for calculating the pose compensation matrix is:

[0068] ;

[0069] in, for t The pose compensation matrix at time step 1. for t The rotation matrix at time step, It is the identity matrix. for t The linear velocity vector at time t. for t The acceleration vector at time t. for The norm of .

[0070] In one specific embodiment, MobileNetV3-Large is used as a lightweight backbone network, and deformable convolution is introduced to enhance the ability to perceive deformable and small targets; the SE channel attention module and the spatial adaptive aggregation module are integrated in the multi-scale attention feature pyramid to realize the dynamic fusion of features at different scales.

[0071] The decoupled Transformer decoder consists of 6 stacked self-attention and cross-attention layers, with three parallel branches responsible for component category prediction, bounding box regression, and pose angle estimation, respectively.

[0072] A three-layer temporal graph attention network is used to process the acceleration and angular velocity data of the inertial measurement unit sampled at 200Hz, and a 4×4 pose compensation matrix for each frame of 8000fps image is generated by cubic spline interpolation to perform pose correction and projection on the two-dimensional detection coordinates.

[0073] A confidence-weighted filtering algorithm is applied to the five-frame detection results of the same component to eliminate false detections with a confidence level below 0.6 or a spatial geometric consistency deviation exceeding 10 mm.

[0074] This embodiment can achieve 45fps real-time inference on edge devices, and under inspection conditions of 200km / h, the detection mAP of 28 types of parts reaches 98.5%, the recall rate of small targets is improved by 30%, the 3D positioning error is better than ±5mm, and it remains stable in complex environments such as tunnels and alternating light and dark environments.

[0075] The multi-scale target detection model includes:

[0076] The lightweight backbone network employs an improved lightweight convolutional neural network architecture, introducing deformable convolutions to replace standard convolutional layers, enhancing the network's ability to perceive deformable and small targets while maintaining the same computational cost. The multi-scale attention feature pyramid includes a channel attention module and a spatial adaptive aggregation module. The channel attention module learns the importance weights of each channel, while the spatial adaptive aggregation module dynamically fuses multi-scale features using a learnable kernel function. The decoupled Transformer decoder includes three parallel branches for category prediction, position regression, and shape estimation, respectively. By introducing prior knowledge of components to guide feature decoding, the network's dependence on labeled data is reduced. During the training phase of the detection network, a multi-task learning strategy with focus loss is adopted, using adaptive weight balancing for different categories of components and samples of varying difficulty, improving the robustness of the detection network to long-tailed data distributions.

[0077] The pose compensation matrix and the confidence-weighted filtering algorithm include:

[0078] A temporal graph attention network is used to model the acceleration and angular velocity data collected by the inertial measurement unit (IMU) to construct a temporal state graph. The vehicle vibration mode of the inspection vehicle is extracted through graph convolution and temporal self-attention mechanism to obtain a continuous attitude estimation sequence. Based on the temporal relationship between the IMU sampling frequency and the camera frame rate, cubic spline interpolation is used to calculate the pose compensation matrix corresponding to each frame of the image to eliminate the influence of vehicle body shaking on the detection results. For the same contact wire component detected, a temporal fusion tree of multi-frame detection results is constructed. Each node of the temporal fusion tree contains the component's category, position, size, and confidence information. The optimal state estimate is calculated using an adaptive Bayesian recursive estimation method. Based on the geometric consistency verification mechanism, the spatial relationship constraints between components and structural prior knowledge are used to eliminate false detection results that violate physical rules, thereby improving the overall detection accuracy.

[0079] Specifically, this embodiment employs a multi-scale target detection model using a lightweight backbone network, a multi-scale attention feature pyramid, and a decoupled Transformer decoder. It combines inertial measurement unit data processed by a temporal graph attention network with pose compensation matrices calculated by cubic spline interpolation, and a confidence-weighted filtering algorithm constructed using adaptive Bayesian recursive estimation and geometric consistency verification mechanisms. This improves the detection accuracy and spatial positioning precision of overhead contact line components and significantly reduces the false detection rate caused by environmental interference and vehicle vibration.

[0080] The process involves aligning the detection results with historical detection data in a spatiotemporal manner to construct a temporal feature vector indexed by the component number. This temporal feature vector is then input into a prediction model combining a temporal convolutional network and a long short-term memory network. The model outputs the defect evolution trend and risk probability of the component. The risk probability is then graded based on a dynamic Bayesian thresholding algorithm, including:

[0081] Spatiotemporal alignment of component inspection data collected from multiple inspections is performed. A spatial index is established based on global navigation satellite system positioning data and component numbers. The data from multiple time periods is time-aligned using a dynamic time warping algorithm. A temporal feature database is constructed based on the unique identifier of each component. Each record includes component type, spatial location, inspection time, appearance features, geometric parameters, environmental conditions, and historical maintenance records. The state change sequence of each component at different time points is extracted using the sliding time window method to form a multidimensional temporal feature vector.

[0082] In one specific embodiment, the specific implementation of the spatiotemporal alignment and temporal feature vector construction includes:

[0083] A spatial mapping relationship is established based on the physical location of the overhead contact system components and the line mileage. A spatial hash function is designed to quickly index adjacent inspection areas, realizing the spatial correspondence of components under different inspection cycles. An improved dynamic time warping algorithm is used to perform nonlinear matching on different time series. This algorithm introduces an adaptive window mechanism based on quantized time error to solve the problem of uneven sampling caused by different vehicle speeds. Feature extraction templates are designed for each type of component to extract structural features, surface features, and geometric features from the original inspection data. Features include, but are not limited to, component size deviation, wear degree, surface defect area, bolt loosening angle, connection status, and geometric deformation. Environmental context information is introduced, including external factors such as temperature, humidity, wind speed, rain and snow, and electrical environment, as well as operating parameters such as mechanical load and current load, to establish a multi-dimensional environment-component correlation matrix. A hierarchical feature fusion strategy is adopted to normalize features from different sources and with different measurement accuracies, constructing a multi-level time series feature representation that includes historical evolution, current state, and environmental impact, supporting subsequent trend prediction and risk assessment.

[0084] A multi-dimensional temporal feature vector is input into a prediction model to predict defect trends. The prediction model includes a multi-layer temporal convolutional network, a long short-term memory network, and an attention mechanism.

[0085] The multi-layer temporal convolutional network is used to extract temporal pattern features, the long short-term memory network is used to capture long-term dependencies, the attention mechanism is used to highlight key state changes, and the prediction model is used to output the defect evolution trend and predicted risk probability of each component. A risk quantification index is established, and a dynamic Bayesian threshold algorithm is used to adaptively classify the predicted risk probability in combination with changes in environmental conditions, load levels, and maintenance cycles, generating four-level risk labels, including: R1, R2, R3, and R4, where R1 is the normal level risk label, R2 is the attention level risk label, R3 is the warning level risk label, and R4 is the danger level risk label.

[0086] In one specific embodiment, the defect trend prediction and risk classification are specifically implemented as follows:

[0087] The prediction model adopts a multi-branch parallel architecture. The multi-layer temporal convolutional network branch contains three one-dimensional convolutional layers, with kernel sizes of 3, 5, and 7 respectively, to capture change patterns at different time scales. The long short-term memory network branch adopts a bidirectional structure with a hidden layer dimension of 256 to capture long-term dependencies. A multi-head self-attention mechanism is introduced with 8 heads and an attention dimension of 64 to enhance the ability to perceive key state transition points.

[0088] For different types of components, an expert knowledge-guided loss function is designed. This loss function combines mean squared error, negative log-likelihood and ranking loss to give higher weights to safety-critical components and improve the accuracy of risk prediction.

[0089] A Bayesian network model is trained based on historical component failure data to establish a conditional probability table of component status, environmental conditions, and failure probability, achieving precise quantification of risk probability. A dynamic Bayesian threshold algorithm adaptively adjusts the risk threshold based on the current operating environment, seasonal variations, and line importance. Risk level boundaries are determined through Bayesian posterior probability estimation, where R1 corresponds to a failure probability < 0.05, R2 to 0.05-0.25, R3 to 0.25-0.6, and R4 to > 0.6. For each risk level, a detailed risk description report is generated, including the risk source, development trend, prediction time window, and recommended measures, forming a structured risk assessment file.

[0090] In one specific embodiment, the formula for calculating the spatiotemporal alignment is:

[0091] ;

[0092] in, Time series x and y The Middle j and the kThe cumulative distance between points The feature distance function, For time-related weights, To add or remove penalty coefficients, For the time difference, This is a time-scale parameter.

[0093] Specifically, this embodiment introduces a lightweight convolutional neural network architecture and a channel attention and spatial adaptive aggregation module in a deformable convolutional, multi-scale attention feature pyramid network, as well as three parallel branches in a decoupled Transformer decoder: category prediction, position regression, and shape estimation. Combined with multi-task learning with a focus loss function and an adaptive weight balancing strategy, it achieves improved target detection speed and enhanced robustness to occluded and deformed targets under limited computing resources. This enables a high recognition rate for rare but critical component defects even under imbalanced data conditions.

[0094] Based on risk level and spatial positioning data, and combined with a knowledge graph-driven rule engine, a smart work order is generated, containing maintenance paths, priority rankings, and spare parts lists. This smart work order is mapped to on-site equipment via an augmented reality navigation device, and maintenance results are collected and fed back to a digital twin database. This includes:

[0095] Construct a knowledge graph for the overhead contact system, which includes spatial relationships, functional dependencies, fault propagation modes, and maintenance specifications among components;

[0096] Based on the overhead contact network knowledge graph and risk level data, a combined optimization algorithm is used to calculate the optimal maintenance path and timing arrangement.

[0097] For overhead contact line components with R3 and R4 risk levels, generate smart work orders that include detailed maintenance steps, required tools, spare parts lists, and quality acceptance standards.

[0098] By using augmented reality navigation devices, work order information is spatially matched with on-site equipment, and maintenance guidance information is projected into the field of view of maintenance personnel in real time, providing visual operation guidance.

[0099] In one specific embodiment, the specific implementation of the knowledge graph construction and intelligent work order generation includes:

[0100] The knowledge graph adopts a multi-layered structure. The bottom layer contains component ontology, attributes, and basic relationships; the middle layer contains functional groups and system topology; and the top layer contains fault modes, propagation rules, and maintenance knowledge. It is stored and managed through a graph-structured database management system oriented towards relational networks. A graph embedding method based on recurrent neural networks maps component nodes and their attributes to a low-dimensional vector space, calculates the functional correlation and fault correlation between components, and constructs a fault propagation model for the overhead contact system. A mixed-integer programming algorithm, combined with risk level, geographical location, maintenance time window, maintenance resource constraints, and operational safety requirements, solves for the globally optimal maintenance plan. The scheduling scheme minimizes maintenance costs and operational impact; for each component to be maintained, applicable maintenance procedures are extracted from the knowledge graph, and parameterized maintenance steps are automatically generated based on environmental conditions, defect type, and equipment model; maintenance tasks for multiple components in the same area are intelligently merged to optimize workflows and reduce repetitive work; intelligent work orders support multi-level expansion and context awareness, dynamically adjusting the level of detail in guidance information based on the maintenance personnel's qualification level and on-site conditions, highlighting key operations and safety precautions; work order execution progress is tracked in real time, and warnings are issued for deviations from standard procedures to ensure maintenance quality.

[0101] The maintenance process data and maintenance results are collected based on mobile terminals. The maintenance process data includes operation time and operation records, and the maintenance results include replacement component information and quality test results.

[0102] Maintenance feedback data is uploaded to the digital twin database in real time to update the catenary status and adjust model parameters;

[0103] An adaptive incremental learning mechanism is constructed to dynamically update the model parameters of the multi-scale target detection model and the prediction model for novel defect samples and false detection samples.

[0104] In one specific embodiment, a distributed federated learning framework is adopted to aggregate the detection and maintenance experience of multiple lines while protecting data security, continuously improve the generalization ability of the algorithm model, and form an intelligent closed loop of "detection-prediction-maintenance-feedback-optimization".

[0105] The specific implementation of the maintenance feedback and incremental learning includes:

[0106] A standardized maintenance feedback template is adopted, which includes objective measurement data and subjective evaluation information, and is collected in real time through mobile terminals; key maintenance operations are recorded by video, and computer vision methods are used to extract work standardization indicators.

[0107] A maintenance quality evaluation model is constructed, which integrates maintenance time, operational accuracy, spare parts usage, and functional test results to generate a maintenance quality score; the digital twin database adopts a layered architecture, including a physical layer, a data layer, and a business layer, to realize a panoramic digital representation of the overhead contact system from components to the whole.

[0108] Based on the temporal graph data model, the system records the status changes of components throughout their entire life cycle, supporting status history queries and trend analysis. An incremental update mechanism is designed to only synchronize data that has changed, reducing communication load and storage pressure. The incremental learning mechanism is based on an uncertainty sampling strategy, prioritizing the selection of samples with low model confidence or large prediction bias for training.

[0109] An elastic weight merging algorithm is adopted to balance the parameters of the old and new models and avoid the catastrophic forgetting problem. A multi-level feedback optimization system is constructed, with short-term feedback used to adjust the detection threshold and filtering rules, medium-term feedback used to update the model parameters and feature representation, and long-term feedback used to improve the model structure and algorithm process.

[0110] Specifically, this embodiment integrates the multi-level structure of knowledge graphs, the graph embedding method of recurrent neural networks, the mixed integer programming algorithm, the parameterized maintenance procedure generation technology, the context-aware work order display mechanism, the standardized maintenance feedback template, the hierarchical digital twin database architecture, the temporal graph data model, and the incremental learning mechanism based on uncertainty sampling to achieve accurate characterization and prediction of catenary fault propagation modes.

[0111] This embodiment improves the efficiency of global optimization planning of maintenance paths by 40%, increases maintenance resource utilization by 35%, and enhances maintenance quality consistency by 50%. At the same time, it constructs a complete closed-loop system of "detection-prediction-maintenance-feedback-optimization", which continuously improves the adaptive capability and generalization performance, supports edge computing, and significantly reduces communication load and storage pressure. It effectively solves the problems of traditional acceptance methods, such as lack of scientific basis for maintenance decision-making, low efficiency of resource allocation, large fluctuations in maintenance quality, and difficulty in accumulating and passing on experience.

[0112] Please see Figure 2 The present invention also provides a high-speed railway catenary completion acceptance system based on 4C detection, the system comprising:

[0113] The data acquisition module is used to acquire contact network image data based on a high-speed camera array, obtain the pose data and spatial positioning data of the inspection vehicle, perform synchronous calibration processing, establish the mapping relationship between the image coordinate system and the track coordinate system, and obtain the calibrated contact network image.

[0114] The multi-scale target detection module is used to input the calibrated catenary image into the deep learning-based multi-scale target detection model, output the detection result, perform pose correction on the detection result, generate the three-dimensional positioning coordinates of the component in the track coordinate system by combining the mapping relationship, and perform noise reduction processing on the detection result through the confidence weighted filtering algorithm to remove low confidence false detection samples.

[0115] The risk grading module is used to align the detection results with historical detection data in time and space, construct a temporal feature vector indexed by the part number, input the temporal feature vector into a prediction model combining a temporal convolutional network and a long short-term memory network, output the defect evolution trend and risk probability of the parts, and grade the risk probability based on the dynamic Bayesian threshold algorithm.

[0116] The acceptance feedback module is used to generate intelligent work orders that include maintenance paths, priority rankings, and spare parts lists based on risk levels and spatial positioning data, combined with a knowledge graph-driven rule engine. The intelligent work orders are mapped to on-site equipment through an augmented reality navigation device, and the maintenance results are collected and fed back to the digital twin database.

[0117] Specifically, this embodiment of a high-speed rail catenary completion acceptance system based on 4C detection provides multi-sensor synchronous measurement and adaptive image acquisition capabilities with an accuracy of ±5mm through a data acquisition module; achieves a component recognition rate of over 98% and millimeter-level spatial positioning accuracy through a multi-scale target detection module; realizes a defect prediction accuracy of over 85% and dynamic environment adaptive risk assessment through a risk classification module; and constructs a complete closed loop from detection to maintenance through an acceptance feedback module. The overall system acceptance efficiency is improved by more than 3 times, the workload of manual inspection is reduced by 70%, the missed detection rate is reduced by 85%, and the defect early warning lead time is increased, fundamentally solving the key problems of low efficiency and poor accuracy of traditional acceptance methods.

[0118] 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, the memory, and the 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 the final acceptance of high-speed railway catenary based on 4C detection.

[0119] This invention also discloses a computer-readable storage medium that stores computer instructions, which instruct the computer to implement all or part of the steps of the high-speed rail catenary completion acceptance method based on 4C detection 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.

[0120] 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 method for the final acceptance of high-speed railway overhead contact lines based on 4C testing, characterized in that, Includes the following steps: The system acquires contact network image data using a high-speed camera array, obtains the pose data and spatial positioning data of the inspection vehicle, performs synchronous calibration processing, establishes the mapping relationship between the image coordinate system and the track coordinate system, and obtains the calibrated contact network image. The calibration catenary image is input into a deep learning-based multi-scale target detection model, which outputs the detection results. The detection results are then corrected for pose. Based on the mapping relationship, the three-dimensional positioning coordinates of the components in the track coordinate system are generated. The detection results are then denoised using a confidence-weighted filtering algorithm to remove low-confidence false detection samples. The detection results are spatiotemporally aligned with historical detection data to construct a temporal feature vector indexed by the component number. The temporal feature vector is then input into a prediction model combining a temporal convolutional network and a long short-term memory network to output the defect evolution trend and risk probability of the component. The risk probability is then graded based on a dynamic Bayesian thresholding algorithm. Based on risk level and spatial positioning data, combined with a knowledge graph-driven rule engine, intelligent work orders are generated, which include maintenance paths, priority rankings, and spare parts lists. The intelligent work orders are mapped to field equipment through augmented reality navigation devices, and maintenance results are collected and fed back to the digital twin database. The process involves aligning the detection results with historical detection data in a spatiotemporal manner to construct a temporal feature vector indexed by the component number. This temporal feature vector is then input into a prediction model combining a temporal convolutional network and a long short-term memory network. The model outputs the defect evolution trend and risk probability of the component. The risk probability is then graded based on a dynamic Bayesian thresholding algorithm, including: Spatiotemporal alignment of component inspection data collected from multiple inspections is performed. A spatial index is established based on global navigation satellite system positioning data and component numbers. Data from multiple time periods is time-aligned using a dynamic time warping algorithm. A time-series feature database is constructed based on the unique identifier of each component. Each record includes component type, spatial location, inspection time, appearance features, geometric parameters, environmental conditions, and historical maintenance records. The state change sequence of each component at different time points is extracted using the sliding time window method to form a multi-dimensional time-series feature vector. A multi-dimensional temporal feature vector is input into a prediction model to predict defect trends. The prediction model includes a multi-layer temporal convolutional network, a long short-term memory network, and an attention mechanism. The multi-layer temporal convolutional network is used to extract temporal pattern features, the long short-term memory network is used to capture long-term dependencies, the attention mechanism is used to highlight key state changes, and the prediction model is used to output the defect evolution trend and predicted risk probability of each component. A risk quantification index is established, and a dynamic Bayesian threshold algorithm is used to adaptively classify the predicted risk probability in combination with changes in environmental conditions, load levels, and maintenance cycles, generating four-level risk labels, including: R1, R2, R3, and R4, where R1 is the normal level risk label, R2 is the attention level risk label, R3 is the warning level risk label, and R4 is the danger level risk label.

2. The method for acceptance testing of high-speed railway overhead contact lines based on 4C testing as described in claim 1, characterized in that, The process of acquiring contact network image data based on a high-speed camera array, obtaining the pose data and spatial positioning data of the inspection vehicle, and performing synchronous calibration processing to establish a mapping relationship between the image coordinate system and the track coordinate system, thereby obtaining a calibrated contact network image, includes: The inspection vehicle travels at a constant speed along the railway line. A circular array containing multiple high-speed cameras is configured on the onboard 4C inspection device to record the six degrees of freedom motion data of the inspection vehicle body. Centimeter-level spatial positioning data of the inspection vehicle is recorded through the global navigation satellite system. The collected six-degree-of-freedom motion data and spatial positioning data are aligned based on timestamps, and an intrinsic and extrinsic parameter matrix between multi-view high-speed cameras is constructed. Based on the real-time attitude of the vehicle body, the transformation relationship from the high-speed camera coordinate system to the track coordinate system is calculated to eliminate the influence of vehicle body vibration and shaking on image position accuracy, thus obtaining the calibration catenary image.

3. The method for acceptance testing of high-speed railway overhead contact lines based on 4C testing as described in claim 2, characterized in that, The process involves inputting the calibrated contact network image into a deep learning-based multi-scale target detection model, outputting detection results, correcting the pose of the detection results, generating the three-dimensional positioning coordinates of the components in the track coordinate system based on mapping relationships, and denoising the detection results using a confidence-weighted filtering algorithm to remove low-confidence false detection samples, including: A multi-scale target detection model is used to identify and calibrate contact network images. The multi-scale target detection model includes a lightweight backbone network, a multi-scale attention feature pyramid, and a decoupled Transformer decoder. The lightweight backbone network is used to extract multi-level features of the image. The multi-scale attention feature pyramid is used to fuse features at different scales and enhance the representation ability of small targets. The decoupled Transformer decoder is used to decode the category, bounding box, and confidence of the parts in parallel. The pose compensation matrix at the time of image acquisition of the catenary is calculated using inertial measurement unit data. The attitude of the detected two-dimensional coordinates of the catenary components is corrected. Based on the type of catenary components and the distribution of feature points, combined with the camera imaging model, the corrected two-dimensional coordinates are projected onto the track coordinate system to calculate the three-dimensional spatial coordinates of the catenary components. The confidence-weighted filtering algorithm is used for multiple detection results of the same catenary component. Based on the detection confidence, geometric consistency and temporal stability, the optimal fusion result is obtained and false detection samples with high confidence are eliminated.

4. The method for acceptance testing of high-speed railway overhead contact lines based on 4C testing as described in claim 3, characterized in that, The multi-scale target detection model includes: The lightweight backbone network adopts an improved lightweight version of the convolutional neural network architecture, introducing deformable convolutions to replace standard convolutional layers, thereby enhancing the network's ability to perceive deformable and small targets while maintaining the same computational cost. The multi-scale attention feature pyramid includes a channel attention module and a spatial adaptive aggregation module. The channel attention module is used to learn the importance weights of each channel, and the spatial adaptive aggregation module uses a learnable kernel function to dynamically fuse multi-scale features. The decoupled Transformer decoder includes three parallel branches, which are used for class prediction, position regression, and shape estimation, respectively.

5. The method for acceptance testing of high-speed railway overhead contact lines based on 4C testing as described in claim 3, characterized in that, The pose compensation matrix and the confidence-weighted filtering algorithm include: A temporal graph attention network is used to model the acceleration and angular velocity data collected by the inertial measurement unit (IMU) to construct a temporal state graph. The vehicle vibration mode of the inspection vehicle is extracted through graph convolution and temporal self-attention mechanism to obtain a continuous attitude estimation sequence. Based on the temporal relationship between the IMU sampling frequency and the camera frame rate, cubic spline interpolation is used to calculate the pose compensation matrix corresponding to each frame of the image to eliminate the influence of vehicle body shaking on the detection results. For the same contact network component detected, a temporal fusion tree of multi-frame detection results is constructed. Each node of the temporal fusion tree contains the component's category, position, size, and confidence information. The optimal state estimate is calculated using an adaptive Bayesian recursive estimation method. Based on the geometric consistency verification mechanism, the spatial relationship constraints between components and structural prior knowledge are used to eliminate false detection results that violate physical rules.

6. The method for acceptance testing of high-speed railway overhead contact lines based on 4C testing as described in claim 1, characterized in that, Based on risk level and spatial positioning data, and combined with a knowledge graph-driven rule engine, a smart work order is generated, containing maintenance paths, priority rankings, and spare parts lists. This smart work order is mapped to on-site equipment via an augmented reality navigation device, and maintenance results are collected and fed back to a digital twin database. This includes: Construct a knowledge graph for the overhead contact system, which includes spatial relationships, functional dependencies, fault propagation modes, and maintenance specifications among components; Based on the overhead contact network knowledge graph and risk level data, a combined optimization algorithm is used to calculate the optimal maintenance path and timing arrangement. For overhead contact line components with R3 and R4 risk levels, generate smart work orders that include detailed maintenance steps, required tools, spare parts lists, and quality acceptance standards. By using augmented reality navigation devices, work order information is spatially matched with on-site equipment, and maintenance guidance information is projected into the field of view of maintenance personnel in real time, providing visual operation guidance; The maintenance process data and maintenance results are collected based on mobile terminals. The maintenance process data includes operation time and operation records, and the maintenance results include replacement component information and quality test results. Maintenance feedback data is uploaded to the digital twin database in real time to update the catenary status and adjust model parameters; An adaptive incremental learning mechanism is constructed to dynamically update the model parameters of the multi-scale target detection model and the prediction model for novel defect samples and false detection samples.

7. A high-speed railway catenary completion acceptance system based on 4C detection, used to execute the high-speed railway catenary completion acceptance method based on 4C detection as described in any one of claims 1-6, characterized in that, The system includes: The data acquisition module is used to acquire contact network image data based on a high-speed camera array, obtain the pose data and spatial positioning data of the inspection vehicle, perform synchronous calibration processing, establish the mapping relationship between the image coordinate system and the track coordinate system, and obtain the calibrated contact network image. The multi-scale target detection module is used to input the calibrated catenary image into the deep learning-based multi-scale target detection model, output the detection result, perform pose correction on the detection result, generate the three-dimensional positioning coordinates of the component in the track coordinate system by combining the mapping relationship, and perform noise reduction processing on the detection result through the confidence weighted filtering algorithm to remove low confidence false detection samples. The risk grading module is used to align the detection results with historical detection data in time and space, construct a temporal feature vector indexed by the part number, input the temporal feature vector into a prediction model combining a temporal convolutional network and a long short-term memory network, output the defect evolution trend and risk probability of the parts, and grade the risk probability based on the dynamic Bayesian threshold algorithm. The acceptance feedback module is used to generate intelligent work orders that include maintenance paths, priority rankings, and spare parts lists based on risk levels and spatial positioning data, combined with a knowledge graph-driven rule engine. The intelligent work orders are mapped to on-site equipment through an augmented reality navigation device, and the maintenance results are collected and fed back to the digital twin database.

8. An electronic device, characterized in that, include: At least one processor, at least one memory, a communication interface, and a bus; 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 method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 6.

Images