3c vehicle-mounted catenary operation state intelligent analysis system

Abstract

The application discloses a 3C vehicle-mounted contact network operation state intelligent analysis system and belongs to the technical field of cross of track traffic intelligent operation and maintenance and industrial artificial intelligence. The system associates multi-source heterogeneous observation data to specific equipment units through a device centralization data binding module. A multi-modal feature extraction and state management unit processes data and maintains equipment multi-dimensional state vectors by using a Kalman filter updating unit. A physical-data hybrid decision maker fuses data-driven rules and simplified physical model output diagnostic results. A topology analyzer performs global checking based on mechanical transmission rules. The system also includes a long-term health state prediction and feedback module, which predicts early failure risks by using a hidden Markov model and dynamically feedback adjusts Kalman filter process noise, so that long-term prediction and short-term estimation are coordinated. The application solves the technical problems of multi-source data fragmentation, lack of physical basis for diagnosis and inability to predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of intelligent operation and maintenance of rail transit and industrial artificial intelligence, specifically to an intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines. Background Technology

[0002] The overhead contact system of electrified railways is a key piece of equipment for train traction power supply, and its geometric parameters and the stability of its electrical connections directly affect operational safety and efficiency. The onboard overhead contact system operation status detection device (referred to as the 3C device), by integrating a visible light camera, an infrared thermal imager, and geometric parameter measurement sensors, can perform dynamic inspections of the overhead contact system during daily train operation, and has become the mainstream inspection method in the industry.

[0003] Existing technologies typically employ a "divide and conquer" architecture to process multi-source heterogeneous data collected by 3C devices: independent computer vision algorithms process images to identify appearance defects, while a rule-based system based on fixed thresholds determines whether geometric parameters exceed limits; finally, the two sets of results are simply superimposed and output. While this approach achieves a degree of automation, it still suffers from the following shortcomings:

[0004] Existing systems treat images, temperature, and geometric parameters as isolated observation signals, failing to correlate them with specific overhead contact line equipment units (such as individual droppers or clamps). For example, loose components identified in visible light images cannot be spatiotemporally correlated and cross-validated with infrared temperature rise trends and abrupt changes in pull-out values ​​at the same location, resulting in diagnostic conclusions lacking solid multidimensional evidence.

[0005] Purely visual recognition models rely solely on pixel features, failing to incorporate the physical connections and spatial topological constraints between overhead contact line components. This makes them highly susceptible to misinterpreting changes in illumination or normal shading as structural anomalies. Furthermore, parameter alarms based on fixed thresholds cannot distinguish between natural fluctuations in conductor height caused by ambient temperature and genuine structural degradation, nor can they capture early signs of faults resulting from slow, coordinated changes in multiple parameters.

[0006] When the AI ​​model outputs a defect warning, the system cannot call the corresponding mechanical or thermodynamic principle model of the device for inverse deduction. The diagnostic process becomes a "black box", and its conclusions are difficult for on-site maintenance personnel to understand and trust, which hinders the in-depth application of AI in actual decision-making.

[0007] Existing systems lack a platform for continuously tracking the evolution of equipment status. Historical trend analysis and real-time diagnosis are separated, making it impossible to dynamically and adaptively adjust diagnostic thresholds or predict chronic equipment degradation in the early stages. Essentially, they still fall under the "post-event alarm" or "periodic maintenance" model, which is insufficient to support the advanced operation and maintenance concept of predictive maintenance.

[0008] Therefore, there is an urgent need for a new technological paradigm that can break down data barriers and deeply integrate multimodal observation, device physical mechanisms and topological relationships within a unified framework, so as to achieve a leap from "data perception" to "state cognition". This is the core technical problem that this invention aims to solve. Summary of the Invention

[0009] The purpose of this invention is to provide an intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines, which solves the technical problems of fragmented multi-source heterogeneous data, disconnect between perception and cognition, lack of physical interpretability of diagnostic results, and inability to achieve unified cognition and prediction of the equipment's status throughout its entire life cycle in existing vehicle-mounted overhead contact line detection systems.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A 3C vehicle-mounted overhead contact line operation status intelligent analysis system, deployed on a ground server, is used to perform centralized unified status modeling and hybrid-driven diagnosis of multi-source heterogeneous observation data from vehicle-mounted 3C devices. The multi-source heterogeneous observation data includes visible light images, infrared thermal imaging images, geometric parameter measurements, and corresponding spatiotemporal stamp information. The system includes the following modules:

[0012] The centralized data binding module is used to receive multi-source asynchronous data packets containing visible light images, infrared thermal imaging images, geometric parameter measurements, and time stamp information. Based on the predefined catenary equipment units and their three-dimensional bounding boxes in the catenary digital twin model database, the module maps the images to the three-dimensional scene coordinate system through the perspective projection transformation module. Then, the intersection-over-union (IoU) calculation unit calculates the intersection-over-union ratio of the image projection voxels with the three-dimensional bounding boxes of each equipment unit. When the IoU is greater than 0.6, the data packet is bound to the corresponding equipment unit and the original observation segment record is generated.

[0013] The multimodal feature extraction and state management unit performs parallel processing on the original observation segments of each equipment unit: it calls the contact network component recognition neural network to process the visible light image and outputs the first structured data containing component type, visual defect type, pixel-level position mask, and model output confidence; it calls the infrared image temperature feature extraction unit and the geometric parameter offset calculation unit to process the infrared thermal imaging image and geometric parameter measurement values ​​and outputs the second structured data containing the regional average temperature, maximum temperature, temperature gradient standard deviation, guide height offset ΔH, pull-out value offset ΔS, and temperature anomaly degree Tanomaly; and based on the first structured data and the second structured data, it updates the multidimensional state vector V stored in the equipment state vector registry through the Kalman filter update unit. The multidimensional state vector V contains the equipment unique identifier, the latest timestamp, the current feature estimate and its uncertainty measure, the historical moving average and trend slope, and the associated equipment identifier list.

[0014] A physical-data hybrid decision-maker is used to input the updated multidimensional state vector V into a data-driven decision branch and a physical model decision branch: the data-driven decision branch loads a defect judgment threshold rule set according to the equipment type and outputs a preliminary defect type and confidence level; the physical model decision branch calls a simplified physical model for the corresponding equipment type from a simplified physical model library, calculates the physical rationality score; and generates equipment-level diagnostic results by combining the outputs of the two branches through an arbitration decision sub-step.

[0015] The topology analyzer and diagnostic report generator are used to obtain the state vectors of topologically associated devices based on the list of associated device identifiers, verify the compatibility between the diagnostic results and the state changes of associated devices through mechanical transfer rule units, and call the system-level physical simulation model for inversion verification when there is inconsistency, and finally output a diagnostic report with mechanism explanation.

[0016] Furthermore, in the forward propagation of the neural network for identifying the contact wire components, the graph attention module uses the feature map output by the encoder as the node set and the physical connections defined in the digital twin model as edges to calculate the attention weights of neighboring nodes on the current node through the graph attention network. Its expression is:

[0017] ;

[0018] in, Let i be the feature vector of node i. For learnable weight matrix, For attention parameter vectors, For the set of adjacent nodes, This indicates vector concatenation, and the adjacency matrix is ​​frozen during training and does not participate in gradient updates.

[0019] Furthermore, in the Kalman filter update unit, the state transition matrix... Dynamically set according to equipment type: For dropper type equipment... This is a diagonal matrix with diagonal elements of 0.98; for wire clamp devices, The identity matrix; the fundamental process noise covariance matrix. diagonal elements From the formula: Determined, among which:

[0020] This is the basic factor for process noise. This is the short-term trend gain coefficient;

[0021] For the first Historical observation variance of the feature;

[0022] For the first The normalized historical trend slope of the feature is calculated using the following formula: ,in The slope of the historical trend obtained by linearly fitting historical observation data. The standard deviation of the slope of this historical trend. It is a small constant to prevent division by zero.

[0023] When the system enables the long-term health status prediction and feedback module, the final process noise covariance matrix used is... The diagonal elements are as follows: Adjustments will be made: ,in This is the long-term risk feedback gain coefficient. This is an early failure risk index.

[0024] Furthermore, in the physical model decision branch, the simplified physical model for the dropper device unit is a tension estimation model, which estimates the equivalent sag change based on the current pull-out value offset ΔS, and then calculates the tension. The specific steps are as follows:

[0025] Step S41: Calculate the equivalent sag change based on the pull-out offset ΔS using empirical conversion relationships. The formula is: ,in The conversion factor is determined based on the line parameters and dropper type; Step S42: Calculate the current estimated sag. ,in The rated sag for the dropper is derived from the digital twin model parameter library;

[0026] Step S43: Calculate the current estimated tension based on the simplified catenary model. :

[0027] ;

[0028] in, The weight per unit length of the contact wire. For span, The coefficient of thermal expansion of the contact wire material is _____. For ambient temperature, For reference temperature;

[0029] Step S44: Calculate the physical rationality score : ;

[0030] in, The rated tension for the suspension string.

[0031] Furthermore, the logic of the arbitration decision sub-step is as follows:

[0032] If the data-driven decision branch judgment is flawed and the physical rationality score is below 0.7, a high-confidence diagnostic result is generated.

[0033] If the data-driven decision branch has a flaw and the physical plausibility score is in the range of 0.7–0.85, a medium-confidence result to be reviewed is generated, and the priority check of the topology analyzer is triggered.

[0034] If the topology consistency check result is consistent with the data-driven conclusion, the confidence level is increased to "medium-high confidence".

[0035] If the topology verification results are contradictory, maintain "medium confidence pending review" and mark it as a high-priority item for manual review;

[0036] If the data-driven decision branch judgment has a flaw but the physical reasonableness score is 0.85, then a medium confidence result to be reviewed is generated;

[0037] If the data-driven decision branch does not identify a defect but the physical rationality score is below 0.5, an early warning diagnosis result will be generated.

[0038] If the data-driven decision branch does not identify a defect and the physical rationality score is ≥0.5, the equipment is judged to be in normal condition.

[0039] Furthermore, when verifying the "locator clamp loosening" diagnosis, the topology analyzer checks the historical trend slope of the pull-out value of the associated contact wire device unit. Has it been satisfied within the last 5 cycles? mm / cycle, and the direction of change is consistent with the expected loosening of the contact wire caused by the loosening of the clamp. If they are consistent, the diagnostic confidence is increased by 0.1. If they are opposite, the system-level physical simulation model is triggered.

[0040] Furthermore, the system service bus adopts the Apache Kafka message queue, and all communication between modules uses the device ID as the message topic to ensure that data flow is routed with the device as the center.

[0041] Furthermore, it also includes a vehicle-to-ground cooperative mechanism: the vehicle-mounted edge computing device deploys a lightweight contact network component identification neural network, based on the MobileNetV3 backbone network, to calculate anomaly scores in real time. :

[0042] ;

[0043] in, The preset weighting coefficients, These are the normalization constants for temperature anomaly, guide height offset, and pull-out value offset, respectively; only when The original observation segment will only be compressed and uploaded to the ground server when the image quality index is greater than the preset threshold or lower than the preset threshold.

[0044] Furthermore, after completing the diagnosis, the ground server will send the updated diagnostic results and, if necessary, the differential neural network parameters to the vehicle-mounted edge computing device to achieve online incremental evolution of the model.

[0045] Furthermore, it also includes a diagnostic result visualization and review interface, which dynamically presents four types of information: the historical status curve display area displays the historical curves of key feature values ​​over the past 30 periods; the multi-source raw data linkage view synchronously displays visible light images, infrared thermal imaging images, and geometric parameter tables; the hybrid decision intermediate logic display area displays rule matching, physical model input and output, and arbitration logic in a tree diagram; and the topology consistency verification summary area lists the status of related devices and the conclusions of contradiction analysis.

[0046] Furthermore, the diagnostic result visualization review interface supports interactive operation. All manual corrections or parameter adjustments are recorded in the audit log and automatically trigger the reverse annotation process: if an AI misjudgment is corrected, the sample is added to the hard sample set for model retraining; if rules or physical model parameters are adjusted, a configuration change request is generated and updated to the rule set library or physical model library after approval.

[0047] Furthermore, the contact network digital twin model database is built based on PostgreSQL and PostGIS, storing the three-dimensional spatial coordinates, topological connection relationships and design parameters of all contact network equipment units, with each equipment unit having a unique UUID identifier.

[0048] Furthermore, the multidimensional state vector V is stored in a Redis cluster with “device:{ID}” as the key and fields including last_update_time, feature_list, history_stats and link_ids, supporting millisecond-level read and write and sliding window historical statistics updates.

[0049] Furthermore, in the process of updating and managing the multidimensional state vector V, a long-term health status prediction and feedback module is integrated. This module is based on a Hidden Markov Model (HMM) and works in series with the Kalman filter update unit to model the long-term degradation mode of the equipment unit and predict early failures, and dynamically feeds back to the short-term state estimation process. It performs the following steps:

[0050] S140: Steps for defining the hidden space and observation space of the health state; Define a set containing N discrete hidden states for each device unit. Where N=5, is used to characterize the long-term health level of the equipment, that is:

[0051] ;

[0052] Define observation vector Let be a two-dimensional vector consisting of the key state estimates strongly correlated with long-term degradation, output by the Kalman filter update unit at time t:

[0053] ;

[0054] in, Pull out the value offset from the state vector The current posterior estimate, Temperature anomaly The current posterior estimate, This represents the transpose of a vector.

[0055] S141: HMM parameter initialization and personalized baseline establishment steps;

[0056] The parameters of the Hidden Markov Model are denoted as The initialization method is as follows:

[0057] State transition matrix Initialized as a quasi-diagonal matrix, satisfying The remaining elements are 0, representing the prior hypothesis that the health state evolves slowly; where Indicates from state Transferred to The probability of observation; the probability matrix of observation. Parameterization based on observation vectors Given the assumption that the hidden states follow a bivariate Gaussian distribution, i.e. Mean vector With covariance matrix The initial values ​​are determined using a semi-supervised method combining historical data and expert rules: First, historical observation data of at least 100 equipment units of the same model and line section are retrieved from the contact network digital twin model database, and these data are labeled as "healthy," "faulty," or "unknown" according to the operation and maintenance records; for data samples labeled "healthy," their set is used for initialization. (Health) status For data samples labeled "fault" (usually data from several periods prior to the fault), their set is used for initialization. (Fault) status For intermediate states ,That Through and Linear or nonlinear interpolation is performed between them, and the distribution characteristics of the "unknown" category data are taken into account; the initial state distribution vector is determined. Set as Assume the device starts in a healthy state.

[0058] S142: Online recursive learning and HMM parameter update steps; at the end of each diagnostic period t, based on the observation sequence up to time t. Model parameters at the previous time step The parameters are updated using a recursive form of the online expectation-maximization algorithm. The observed mean vector for the i-th hidden state Its update formula is:

[0059] ;

[0060] in, Given all observations and old parameters, at time... In state The posterior probability is calculated using the forward-backward algorithm; For a decreasing learning rate, It is the attenuation constant; To prevent small constants from being divided by zero; covariance matrix With transition matrix A similar recursive formula based on a weighted average of posterior probabilities is used for updating.

[0061] S143: Health Status Decoding and Early Failure Probability Prediction Steps; At time t, based on the current model parameters and observation sequence The Viterbi algorithm is used to decode the most likely hidden state sequence. To obtain the current health status Simultaneously, calculating the future The probability that the equipment is in a deteriorated or faulty state after one diagnostic cycle is used as an early failure risk index. :

[0062] ;

[0063] in, It is the normalized forward probability vector at time t, satisfying ; It is the currently estimated state transition matrix; This represents the j-th component of the vector; The preset prediction step size is 10.

[0064] S144: Noise adaptation step in the process of feeding the prediction information back to the Kalman filter;

[0065] Early failure risk index As an independent adjustment factor, it is added to the adjustment term in the original Kalman filter update process to jointly determine the final process noise covariance. Let the state vector correspond to... and The component indices are respectively and Their corresponding historical variances are respectively and The normalized historical trend slopes are respectively and The adjusted process noise covariance matrix is ​​then... The diagonal elements are corrected according to the following fusion formula:

[0066] ;

[0067] in:

[0068] The index representing the state component (with values...) or );

[0069] This is the basic factor for process noise;

[0070] The historical observation variance of the corresponding feature;

[0071] The normalized historical trend slope of the corresponding feature is defined as described above. same;

[0072] It is the short-term trend gain coefficient;

[0073] It is the long-term risk feedback gain coefficient;

[0074] The early failure risk index is calculated in step S143.

[0075] Adjusted The original calculated value will be used in the next Kalman filter prediction step (time update). This fusion formula reduces the risk of long-term failures. It can independently and directly amplify process noise, so that when the risk of failure increases, even if the short-term trend is not obvious, the uncertainty of the system's state estimation increases, the Kalman gain increases, and thus responds more sensitively to new observations, enabling rapid perception of early slow degradation.

[0076] Furthermore, in the online recursive learning and HMM parameter update step S142, a memory decay window mechanism is set: when the device unit continuously... Each cycle is decoded into a healthy state ( When the learning rate is... Fixed to the minimum value To stabilize model parameters; when the decoding state enters In cases of worse or no better conditions, the learning rate is immediately reset to the initial learning rate. And start over by pressing Perform dynamic reduction, where This is the number of cycles after the state deteriorates and the counting restarts.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] This invention solves the core contradiction of fragmented multi-source data and disconnect between perception and cognition in existing technologies by constructing a centralized unified state representation and fusion cognition framework for equipment, and realizes the leap from "dispersed perception" to "unified cognition" in catenary state analysis.

[0079] This invention, through centralized data association based on digital twins and high-precision spatiotemporal binding, maps asynchronous observations from different sensors and at different times to a single device entity, forming a complete chain of state evidence. Combined with a graph attention visual recognition network that incorporates physical topological relationships, image analysis no longer relies on isolated pixels but conforms to the spatial constraints of the overhead contact system, significantly reducing false detections caused by occlusion and changes in lighting.

[0080] This invention employs a unique physics-data hybrid decision-driven mechanism, combining the rapid response of data-driven rules with mechanism verification based on simplified physical models (such as the catenary equation). The arbitration decision not only outputs "what the defect is," but also provides a physical plausibility score for "why it is credible." The results can be directly applied to engineering decisions, improving the system's acceptability and practicality.

[0081] This invention employs a collaborative state estimation and prediction framework combining Kalman filtering and Hidden Markov Models (HMMs). The system not only smooths noise and estimates the current state but also models the long-term evolution patterns of equipment health, predicting early-stage chronic degradation risks. A unique adaptive mechanism that feeds long-term risk feedback to short-term process noise dynamically enhances the system's sensitivity to slow faults, achieving truly predictive maintenance early warning.

[0082] This invention utilizes topology consistency verification based on the principle of mechanical transmission, placing the anomalies of individual devices within the entire overhead contact line suspension system for verification. When local diagnostics contradict changes in the state of associated devices, a system-level physical simulation inversion is triggered. This assesses the rationality of the anomaly combination from a global mechanical balance perspective, effectively filtering isolated false alarms caused by local interference and ensuring the logical self-consistency of diagnostic conclusions at the system level.

[0083] This invention not only significantly outperforms traditional methods in technical indicators (such as accuracy and false alarm rate) through multimodal fusion, but also achieves substantial breakthroughs in interpretability, foresight, and systematicness of diagnosis by introducing physical mechanisms and state evolution models. It provides a solid and reliable technical core for the intelligent transformation of rail transit catenary operation and maintenance towards "condition-based maintenance" and "predictive maintenance". Attached Figure Description

[0084] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0085] Figure 1 This is an overall architecture diagram of the system described in this invention.

[0086] Figure 2 This is the overall flowchart of the system method of the present invention. Detailed Implementation

[0087] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0088] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0089] Example 1: See Figure 1 and Figure 2This embodiment discloses a 3C vehicle-mounted overhead contact line operation status intelligent analysis system, including:

[0090] The centralized data binding module is used to bind multi-source asynchronous data packets to the corresponding catenary equipment units and generate original observation segment records based on the catenary digital twin model database.

[0091] The multimodal feature extraction and state management unit is used to process the original observation segments, extract the first structured data and the second structured data, and update the multidimensional state vector of the device using the Kalman filter update unit;

[0092] A physical-data hybrid decision-maker is used to generate device-level diagnostic results by fusing the outputs of data-driven decision branches and physical model decision branches based on an updated multidimensional state vector.

[0093] The topology analyzer and diagnostic report generator are used to verify the compatibility of diagnostic results with changes in the status of associated devices and output diagnostic reports with mechanistic explanations.

[0094] The intelligent analysis system is deployed in the ground data center of the railway bureau or maintenance unit, and consists of a high-performance server cluster, including computing nodes, storage nodes and network nodes.

[0095] The compute nodes employ Intel Xeon Gold 6348 processors (28 cores / 56 threads), NVIDIA A100 GPUs (40GB VRAM), and 128GB DDR4 memory to perform deep learning inference, Kalman filter updates, and physical simulation calculations. The storage nodes utilize a distributed object storage architecture built on Ceph to persist the contact network digital twin model database, equipment state vector registry, and historical observation data. The network nodes are interconnected via 10 Gigabit Ethernet and configured with dedicated firewalls and API gateways to ensure secure data transmission with in-vehicle 3C devices.

[0096] The system service bus uses Apache Kafka message queue, version 3.3.1, configured with 3 Broker nodes. Topics are divided by device ID, each topic has 8 partitions, and the replication factor is 3 to ensure high throughput and high availability. All functional modules are encapsulated in Docker containers and orchestrated and scheduled through Kubernetes, supporting elastic scaling.

[0097] The implementation of the centralized data binding process begins with receiving multi-source asynchronous data packets uploaded from onboard 3C devices on high-speed trains. These packets are transmitted to a ground server via a 4G / 5G wireless network or railway private network, and after decryption through a TLS 1.3 encrypted channel, they enter a Kafka message queue. Each packet contains four core data categories:

[0098] Visible light images (resolution 2448×2048, JPEG format, frame rate 15fps), infrared thermal images (resolution 640×512, 14-bit RAW format, frame rate 9Hz), geometric parameter measurements (guide height, pull-out value, contact wire height, sampling frequency 100Hz, accuracy ±2mm), and high-precision spatiotemporal stamp information are included. The spatiotemporal stamps are generated by the vehicle-mounted GNSS / IMU integrated navigation system. The GNSS module is a u-blox F9P, supporting RTK positioning with a horizontal positioning accuracy better than ±0.3 meters; the IMU module is an XsensMTi-630, with gyroscope zero-bias stability <0.5° / h and accelerometer noise density... The two are fused through extended Kalman filtering, and the output time synchronization error is less than ±10 milliseconds in the spacetime stamp.

[0099] Upon receiving the data packet, the system immediately invokes the centralized data binding module. This module first parses the mileage information (unit: meters, accuracy ±0.5m) and UTC timestamp from the data packet, and then queries the overhead contact line digital twin model database accordingly. The database, based on PostgreSQL 14 and extended with a PostGIS spatial plugin, stores the three-dimensional spatial coordinates of all overhead contact line equipment (WGS84 coordinate system converted to a local ENU coordinate system), three-dimensional axial bounding boxes (AABB) for spatial calculations, topological connections (stored in adjacency list format), and design parameters (such as dropper length, clamp type, rated tension, etc.).

[0100] Each contact network equipment unit in the database is defined as the smallest maintainable component entity with a unique UUID (such as "DS-2024-GZ-00127"). It covers a total of 12 standard components, including droppers, positioners, clamps, electrical connectors, and segment insulators. Each component comes with standardized three-dimensional bounding box (AABB) parameters, including the center point coordinates (x, y, z) and half length (dx, dy, dz).

[0101] The core of the binding operation lies in mapping a 2D image to a 3D scene and determining its spatial attribution. The system first calls the perspective projection transformation module, implemented using OpenCV 4.8. Given the intrinsic parameter matrix K of the onboard 3C device's camera (obtained through calibration, fx=2100, fy=2100, cx=1224, cy=1024) and the extrinsic parameter matrix [R|t] (determined by pre-calibration of the IMU attitude and installation position), combined with the train pose corresponding to the current spatiotemporal stamp (calculated by GNSS / IMU), the system backprojects the pixels in the visible light and infrared images onto 3D rays in the ENU coordinate system. Subsequently, the system voxelizes the target region in the image (typically within ±300 pixels of the image center), generating a cubic mesh with a side length of 0.1 meters. Each voxel records the mean value of its corresponding image features. Next, the Intersection over Union (IoU) calculation unit iterates through the AABB of the voxel set and all device units in the digital twin model, calculating the ratio of their intersection volume to their union volume in 3D space. When the IoU value of a device unit is greater than a preset threshold of 0.6, the system determines that the original observation segment belongs to that device unit and generates an original observation segment record. Its structure is a JSON object, containing the device ID, observation time, image hash value, geometric parameter snapshot, and binding confidence (i.e., IoU value).

[0102] The multimodal feature extraction and state vector update steps are triggered immediately after device binding is completed. The system reads the multidimensional state vector V of the corresponding device unit from the Redis 7.0 cluster. This Redis instance is configured in cluster mode, containing 6 master-slave nodes, and the persistence strategy is AOF everysec to ensure strong consistency and low-latency access of state data. The state vector V is stored in a fixed schema hash structure, with the key name "device:{ID}" and fields including "last_update_time", "feature_list", "history_stats", and "link_ids".

[0103] Visible light images are fed into a contact network component recognition neural network. This neural network is implemented in PyTorch 1.13, with an input size of 1024×1024, scaled by bicubic interpolation. The encoder uses a ResNet-50 backbone network pre-trained on ImageNet, freezing the first three stages and fine-tuning only stages 4 and 5; the decoder uses an FPN structure, fusing features from layers C3 to C5, and outputting feature maps at four scales.

[0104] The key innovation lies in the integration of the Graph Attention Module (GAM):

[0105] After FPN output, the system extracts the RoIAlign features (256 dimensions) of each component candidate region, forming a set of node feature vectors { Simultaneously, the system queries the connection relationships between the device unit and its adjacent components from the digital twin model, constructing an adjacency matrix A∈{0,1}^{N×N}, where N is the total number of components detected in the image (up to an upper limit of 20). The graph attention module (GAM) is implemented using a graph attention network (GAT), whose GAT layer contains two attention heads. Each head has a weight matrix W∈R^{256×128}, an attention parameter vector a∈R^{256}, and a LeakyReLU negative slope of 0.2. During forward propagation, for each node i, the attention coefficient α_ij between it and its adjacent node j is calculated, using the formula described in the invention summary.

[0106] The aggregated features are classified by a fully connected layer, outputting component type (12 categories), visual defect type (cracks, corrosion, missing parts, looseness, 4 categories in total), pixel-level mask (sigmoid activation), and first confidence score C_defect∈[0,1]. This network is trained on a mixed dataset containing 400,000 synthetic images generated by the sample data generation method described in this invention and 20,000 real-world labeled images. The batch size is 16, and the optimizer is AdamW (lr=1e-4, weight_decay=1e-4).

[0107] During training, the adjacency matrix A is used as a fixed graph structure input. After 200 epochs of training, the results were evaluated on an independent real-world validation set (3000 images), achieving a part detection mAP@0.5 of 0.88 and a defect segmentation mIoU of 0.76.

[0108] The infrared image temperature feature extraction unit first crops the target region from the infrared image based on the AABB corresponding to the device ID (aligned to the visible light image after coordinate transformation). The temperature statistical characteristics of this region include: average temperature. (Unit: °C) Highest temperature Temperature gradient standard deviation (Calculate the standard deviation of the 8-neighborhood difference). The geometric parameter offset calculation unit then reads the current guide height. With pull value And obtain design values ​​from the digital twin model. , ,calculate , .

[0109] Simultaneously, the system obtains the current ambient temperature by querying additional ambient temperature sensor data uploaded by the onboard device or estimating it based on the background area of ​​an infrared image. Then, it reads the moving average reference temperature over the past 30 cycles under similar ambient temperatures from the historical statistics field of the state vector. and historical temperature standard deviation Calculate the temperature anomaly Tanomaly = ( - ) / The above results constitute the second structured data, which is returned in dictionary form.

[0110] Perform a state vector update. Use the output above as the observation input. The Kalman filter is invoked to update the unit. The "current eigenvalue" field in the state vector V corresponds to the state variable. The dimensions are 4 (visual defect confidence, temperature anomaly, ΔH, ΔS).

[0111] The state transition matrix F is preset according to the device type: for droppers (device type code 01), F = diag([0.98, 0.98, 0.98, 0.98]); for wire clamps (code 02), F = The observation matrix H is the identity matrix.

[0112] Process noise covariance It is a diagonal matrix, and its diagonal elements According to the formula Calculation, where , , For the first Historical observation variance of the feature For the first The normalized historical trend slope of the feature (obtained by fitting the past 30 points using linear least squares).

[0113] Observation noise covariance matrix It is a diagonal matrix, and its diagonal elements are set according to the sensor measurement accuracy and model performance:

[0114] : Confidence level corresponding to visual defects The observation noise variance (based on the classification error rate estimate of the neural network on the validation set).

[0115] : Corresponding temperature anomaly The observation noise variance (unit: dimensionless, estimated based on infrared thermometer accuracy ±1.5℃ and historical temperature fluctuations).

[0116] Corresponding guide height offset The observation noise variance (unit: m², based on the laser rangefinder accuracy of ±2mm, i.e. ±0.002m, the variance is...) );

[0117] : Corresponding pull-out value offset The observation noise variance (unit: m², based on the laser rangefinder accuracy of ±2mm, i.e. ±0.002m, the variance is approximately...) ).

[0118] Kalman gain Status update Covariance update .

[0119] The updated state vector is written back to Redis, and the historical statistics fields (sliding window queue) are updated synchronously.

[0120] The physics-data hybrid decision-making step is executed after the state vector update. The hybrid decision-maker first queries the device type based on the device ID and loads the corresponding defect judgment threshold rule set from the rule set library. The rule set library is a collection of JSON files stored in a Git repository and is version-controlled. For example, the rule set for a dropper device includes:

[0121] { “rules”: [ {“condition”: “abs(delta_S)>30 and persistent_cycles>=2”, “defect_type”: “string offset”, “confidence_base”: 0.85}, {“condition”: “T_anomaly>2.0 and C_defect>0.7”, “defect_type”: “overheating defect”, “confidence_base”:0.9} ]} The conditional expression is dynamically resolved using Python's eval() function, and persistent_cycles is provided by the continuous anomaly count field in the state vector.

[0122] In physical model decision-making, the system calls a simplified physical model of the corresponding device from the physical model library. Taking the dropper as an example, the tension estimation model is executed according to the following steps: First, based on the current pull-out value offset... Calculate equivalent sag change (In this example, take) ), to obtain the current estimated chasm Then, substitute the values ​​into the simplified catenary model to calculate and estimate the tension: .

[0123] The arbitration decision sub-step integrates the outputs of the two branches:

[0124] If the rule matches successfully and Physical_Score < 0.7, output a high-confidence diagnosis (confidence = confidence_base + 0.1); if the rule matches but Physical_Score > 0.85, output a medium-confidence diagnosis (confidence = confidence_base - 0.15); if no rule matches but Physical_Score < 0.5, output a warning (confidence = 0.6).

[0125] The topology consistency verification and diagnostic result generation steps further enhance diagnostic reliability. The system reads the associated device identifier list Link_IDs (such as the contact wire, catenary wire, etc. associated with the locator) from the state vector and retrieves the latest state vectors of these devices in batches through Redis.

[0126] The mechanical transmission rule unit has a built-in expert knowledge base, for example: "Loosening of the positioner clamp → the contact wire pull-out value should show a negative acceleration change." The system checks whether the slope Slp_S of the pull-out value trend of the associated equipment satisfies Slp_S < -0.5 mm / cycle (last 5 cycles), and whether the sign is consistent with the expectation. If consistent, the diagnostic confidence is increased by 0.1; if contradictory (e.g., Slp_S > 0), the system-level physical simulation model is activated.

[0127] This physical simulation model is built on FEniCS 2019.1.0 and uses linear elastic beam elements to simulate the overhead contact line suspension structure. The input is the state vector of key equipment along the entire line (guide height, pull-out value, temperature). Boundary conditions are set as the fixed ends of the supports, and loads include gravity, wind load (based on meteorological data), and dynamic lifting force (simplified to concentrated force) when a train passes. The model solves the static equilibrium equation Ku = f and outputs the displacement and internal forces of each node. The system compares the residuals of the simulated displacements with the measured values. If the L2 norm of the residuals is less than a threshold (e.g., 5mm), the current observation combination is considered physically feasible, and the original diagnosis is retained; otherwise, it is marked as "physically infeasible," the confidence level is reduced to 0.4, and manual review is recommended. The final diagnostic report includes the defect type, confidence level, Physical_Score, topological consistency conclusion, and remedial recommendations (e.g., "replace within 72 hours").

[0128] The implementation of the vehicle-to-ground cooperative mechanism relies on onboard edge computing devices. This embodiment uses the NVIDIA Jetson AGX Orin module, running Ubuntu 20.04 LTS and JetPack 5.1.1. The lightweight contact network component recognition neural network is reconstructed based on the MobileNetV3-small backbone network, with the number of parameters compressed to 3.2M (original model 16M), input size 512×512, and inference speed reaching 45fps (TensorRT FP16 accuracy). The anomaly score calculation unit calculates the Anomaly_Score in real time according to the aforementioned formula, where the normalization constant is taken as... : The original observation segments are uploaded in JPEG2000 format (10:1 compression ratio) only if the Anomaly_Score > 0.7 or the BRISQUE image quality score < 35. After the ground server returns the diagnostic results, if model updates are involved, the differential parameters (Delta parameters) are issued. The onboard equipment then fine-tunes the final classification layer through incremental learning, achieving online evolution.

[0129] The diagnostic results visualization and review interface is implemented using a web technology stack. The front end uses Vue 3 + TypeScript, the chart library is ECharts 5.4, and the back end uses FastAPI 0.95. The interface dynamically presents four types of information: the historical status curve display area displays key feature values ​​and trend lines over the past 30 periods; the multi-source raw data linkage view supports simultaneous browsing and annotation of visible light, infrared, and geometric parameters; the hybrid decision-making intermediate logic display area uses a tree diagram to show the rule matching path, physical model input and output, and arbitration logic; and the topology consistency verification summary area lists the status of related devices and conflict analysis.

[0130] All interactive operations are recorded in the MongoDB audit log and trigger the reverse annotation process: manually corrected samples are added to the Hard Example Bank for monthly model retraining; rule or parameter adjustments generate YAML format change requests, which are then merged into the main branch after three levels of approval.

[0131] In practical implementation, the catenary equipment of this invention was selected for a section (120km long) of the Beijing-Guangzhou High-Speed ​​Railway to be deployed. The testing period was from March 1st to March 31st, 2024, during which a total of 187,432 valid data packets were collected, covering three main types of equipment: droppers, positioners, and clamps. The statistical indicators of the system diagnostic results compared with those of manual inspection are shown in the table below.

[0132] Comparative Example 1: Using the traditional threshold method: alarms are based solely on geometric parameter thresholds (|ΔH|>50mm or |ΔS|>30mm), without multimodal fusion, physical model, or topology verification.

[0133] The specific results are shown in Table 1:

[0134] Table 1:

[0135]

[0136] Experiments show that this invention significantly outperforms traditional methods on various devices, especially in reducing false alarm rates. The physical-data hybrid decision-making and topology verification mechanism effectively filters out false alarms caused by temporary disturbances (such as birds landing or momentary vibrations).

[0137] In summary, through rigorous engineering design, this invention achieves a complete closed loop from initial observation to interpretable diagnosis, providing reliable technical support for intelligent operation and maintenance of overhead contact lines.

[0138] Example 2: This example is a further optimization based on Example 1. In this example, early prediction of chronic loosening of droppers is based on the HMM-KF series model. This example also includes a long-term health status prediction and feedback module. This module is based on a hidden Markov model and works in series with the Kalman filter update unit to model the long-term degradation mode of the equipment unit and predict early failures, and dynamically feeds back to the short-term state estimation process; it performs the following steps:

[0139] The steps for defining the hidden space and observation space of health status are as follows: Define a set containing multiple discrete hidden states for each device unit to characterize the long-term health level of the device; define the observation vector as a vector composed of key state estimates that are strongly correlated with long-term degradation, which are output from the Kalman filter update unit.

[0140] Hidden Markov Model Parameter Initialization and Personalized Benchmark Establishment Steps: Initialize the state transition matrix, observation probability matrix, and initial state distribution vector of the Hidden Markov Model; the parameters of the observation probability matrix are initialized based on historical normal observation data of the same type of equipment unit through cluster analysis;

[0141] Online recursive learning and Hidden Markov Model parameter update steps: At the end of each diagnostic cycle, the model parameters are updated in a recursive form using an online expectation-maximization algorithm based on the observation sequence up to the current time and the model parameters of the previous time step.

[0142] Health status decoding and early failure probability prediction steps: Based on the current model parameters and observation sequence, decode the most likely hidden state at present, and calculate the probability that the equipment will be in a deteriorated or failed state after several diagnostic cycles in the future, as an early failure risk index;

[0143] The process noise adaptation step of feeding the predicted information back to the Kalman filter is as follows: the early failure risk index is used as an independent adjustment factor and added to the adjustment term in the original Kalman filter update process to jointly determine the final process noise covariance matrix. This adjustment enables the increase in long-term failure risk to independently amplify the process noise, thereby increasing the Kalman gain in the Kalman filter and responding more sensitively to new observations.

[0144] Furthermore, in the online recursive learning and hidden Markov model parameter update steps, a memory decay window mechanism is set: when the device unit is decoded to the healthiest state for multiple consecutive cycles, the learning rate is fixed to the minimum value to stabilize the model parameters; when the decoding state changes to a sub-healthy or worse state, the learning rate is immediately reset to the initial value and the dynamic decay restarts.

[0145] This embodiment aims to verify the early warning capability of the long-term health status prediction and feedback module for chronic loosening faults of droppers in actual lines.

[0146] Test environment and data preparation:

[0147] The test was conducted on a section of a heavy-haul railway line with a steep gradient. Due to prolonged exposure to high dynamic tension, droppers in this section were prone to chronic loosening. Thirty droppers that had been in service for more than five years were selected as test subjects, numbered D001 to D030. The test period was 90 days, with the onboard 3C device performing daily inspections, collecting a total of 2700 valid data packets. All data was processed by the system's front-end centralized data binding module (5) and multimodal feature extraction module, generating a daily data set for each dropper including the pull-out offset. and temperature anomaly The state estimates of features such as [these characteristics] are used as the input data source in this embodiment.

[0148] Specific implementation of the HMM-KF cascade model:

[0149] HMM initialization:

[0150] From the line asset database, historical daily observation data (approximately 60,000 records, each containing data) were extracted from 200 normal droppers with the same environment and model as the test section over the past year. and Using the Python sklearn.cluster.KMeans library, with n_clusters=5, cluster this historical dataset. Then, based on the clustering results, cluster the centers of each class. Based on absolute value (smaller is healthier) and expert experience, the five cluster centers are mapped sequentially to hidden states. to initial observed mean to The covariance of each type of sample is used as the corresponding to The initial values. State transition matrix. Initialize as described above.

[0151] KF configuration: Kalman filter state vector is Initial process noise covariance .

[0152] Parameter settings: Feedback gain , Predicted step size Learning rate decay constant .

[0153] Model Interaction and Early Warning Process Analysis:

[0154] Taking the D015 dropper as an example, it showed significant loosening on the 88th day. The model interaction process is as follows:

[0155] Days 1-70 (Healthy Operation Period): Smoothed state estimates from daily KF outputs. HMM observation vectors. Small fluctuations, Viterbi decoding state Always for , . There are basically no adjustments.

[0156] Days 71-80 (Early Deterioration Period): Due to material fatigue, D015... It begins to exhibit a slow, unidirectional positive drift (from +3mm to +15mm). KF tracks this change and updates its internal state according to the process noise adaptive mechanism. HMM decodes the state from day 75. Begin to become (Sub-health).

[0157] On the 80th day, the calculation is as follows Based on the fusion formula in step S144, a long-term risk feedback term is independently added to the original process noise calculation.

[0158] Given: Components (index) The corresponding historical observation variance Normalized historical trend slope Process noise baseline factor Short-term trend gain coefficient Long-term risk feedback gain coefficient .

[0159] Based on the additive fusion formula of S144, the adjusted process noise is calculated directly: .

[0160] Calculations show that long-term risk As an independent adjustment term, the process noise variance is adjusted from the value when only short-term trends are considered. Increase to This amplifies the signal by approximately 42%. This allows the Kalman filter to improve its performance in the next cycle. The confidence level (Kalman gain) of the new observations increases accordingly, thus enabling more sensitive tracking of subtle deterioration trends in the state.

[0161] Days 81-87 (Accelerated Deterioration Period): The acceleration increases to +42mm. Due to the increased process noise variance, the Kalman gain of KF correspondingly increases. The response to new observations is faster, and the lag in state estimation is reduced. HMM decoding of states is rapid. Transition to (deterioration). The value rose sharply, reaching 0.78 on day 85. The system reached this value on day 85, three days before the failure occurred. At that time, an "early high-risk warning" is automatically generated, indicating that D015 is very likely to detach within 10 days. This contrasts with the traditional fixed threshold method (…). The alarm was triggered on the 86th day.

[0162] Test Results and Comparative Analysis: A full-cycle analysis was conducted on 90 days of data from 30 test droppers. The statistical results are shown in Table 2 below.

[0163]

[0164] This embodiment demonstrates that by introducing the HMM-KF cascade model, predictive maintenance is achieved: the system identifies chronic degradation trends on average more than 3 weeks in advance, shifting the operation and maintenance mode from "post-failure emergency response" to "pre-failure prevention," addressing the core pain point. A perception-cognition-decision enhancement closed loop is formed: the HMM's "cognition" of long-term health status optimizes the KF's "perception" sensitivity of instantaneous status in real time through a feedback mechanism; the two work together deeply, improving the overall system's sensitivity and robustness. The reliability of the results is enhanced: the output health status level and failure probability provide operation and maintenance personnel with decision-making basis far exceeding that of a single alarm signal, promoting the integration of artificial intelligence and expert experience.

[0165] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

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

Claims

1. 3C Vehicle-Mounted Overhead Contact Line Operation Status Intelligent Analysis System, characterized in that, Includes the following modules: The centralized data binding module is used to receive multi-source asynchronous data packets containing visible light images, infrared thermal imaging images, geometric parameter measurements, and time stamp information. Based on the predefined contact network equipment units and their three-dimensional bounding boxes in the contact network digital twin model database, the module maps the images to the three-dimensional scene coordinate system through perspective projection transformation. Then, the intersection-union ratio (IUU) calculation unit calculates the IUU between the image projection voxels and the three-dimensional bounding boxes of each equipment unit. When the IUU is greater than a set threshold, the data packet is bound to the corresponding equipment unit and an original observation segment record is generated. The multimodal feature extraction and state management unit is used to perform parallel processing on the original observation segments of each equipment unit: it calls the contact network component recognition neural network to process the visible light image and outputs the first structured data containing component type, visual defect type, pixel-level position mask and model output confidence. The infrared image temperature feature extraction unit and the geometric parameter offset calculation unit are invoked to process the infrared thermal imaging image and the geometric parameter measurement value, and output the second structured data, which includes the regional average temperature, the highest temperature, the standard deviation of the temperature gradient, the guide height offset, the pull-out value offset, and the temperature anomaly degree. Based on the first structured data and the second structured data, the multidimensional state vector stored in the device state vector registry is updated by the Kalman filter update unit. The multidimensional state vector includes the device unique identifier, the latest timestamp, the current feature estimate and its uncertainty measure, the historical moving average and trend slope, and the list of associated device identifiers. A physics-data hybrid decision-maker is used to input the updated multidimensional state vector into a data-driven decision branch and a physics model decision branch: the data-driven decision branch loads a defect judgment threshold rule set according to the equipment type and outputs a preliminary defect type and confidence level; the physics model decision branch calls a simplified physics model for the corresponding equipment type from a simplified physics model library and calculates the physical rationality score. The device-level diagnostic results are generated by integrating the outputs of the two branches through the arbitration decision sub-step. The topology analyzer and diagnostic report generator are used to obtain the state vectors of topologically associated devices based on the list of associated device identifiers, verify the compatibility between the diagnostic results and the state changes of associated devices through mechanical transfer rule units, and call the system-level physical simulation model for inversion verification when there is inconsistency, and finally output a diagnostic report with mechanism explanation.

2. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, The contact network component identification neural network includes a graph attention module. During forward propagation, this module uses the feature map output by the encoder as the node set and the physical connection relationships defined in the contact network digital twin model as edges to construct a graph structure. The graph attention network calculates the attention weights of adjacent nodes to the current node. The node feature vector, the learnable weight matrix, and the attention parameter vector all participate in the calculation of the attention weights. The attention weights are obtained by normalizing the nonlinear transformation result of the concatenated vectors using a normalized exponential function. Furthermore, the adjacency matrix used to construct the graph structure is frozen during training and does not participate in gradient updates.

3. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, In the Kalman filter update unit, the state transition matrix is ​​dynamically set according to the equipment type: for dropper type equipment, the state transition matrix is ​​a diagonal matrix with diagonal elements less than 1; for wire clamp type equipment, the state transition matrix is ​​an identity matrix. The process noise covariance matrix is ​​a diagonal matrix, and the values ​​of its diagonal elements are obtained by multiplying the process noise base coefficient, the historical observation variance of the corresponding feature, and an adjustment term that is proportional to the absolute value of the normalized historical trend slope. The normalized historical trend slope is the maximum value of the historical trend slope obtained by linearly fitting the historical observation data divided by its standard deviation and a small constant.

4. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, In the physical model decision branch, the simplified physical model for the dropper device unit is a tension estimation model, which estimates the equivalent sag change based on the current pull-out value offset, and then calculates the tension. Specifically, the following steps are performed: The equivalent sag change is calculated based on the pull-out value offset using an empirical conversion relationship, which includes conversion coefficients determined based on line parameters and dropper type. Add the design rated sag to the equivalent sag change to obtain the current estimated sag; Based on a simplified catenary model, the current estimated tension is calculated by combining the unit length weight of the contact wire, the span, the material's coefficient of thermal expansion, the ambient temperature, and the reference temperature. A physical rationality score is calculated by comparing the relative deviation between the current estimated tension and the design rated tension.

5. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, The logic of the arbitration decision sub-step is as follows: if the data-driven decision branch determines that there is a defect and the physical rationality score is lower than the first threshold, a high-confidence diagnostic result is generated; if the data-driven decision branch determines that there is a defect but the physical rationality score is higher than the second threshold of the first threshold, a diagnostic result pending review is generated; if the data-driven decision branch does not determine that there is a defect but the physical rationality score is lower than the third threshold, a warning diagnostic result is generated.

6. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, When verifying the diagnosis of locator clamp loosening, the topology analyzer checks whether the historical trend slope of the pull-out value of the associated contact wire device unit meets the condition of being less than a set negative threshold in recent cycles, and whether the direction of change is consistent with the expected contact wire slack caused by clamp loosening. If they are consistent, the diagnostic confidence is increased; otherwise, the system-level physical simulation model is triggered.

7. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, The system service bus uses a message queue, and all communication between modules is routed using the device's unique identifier as the message subject.

8. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, It also includes a vehicle-to-ground collaborative mechanism: the vehicle-mounted edge computing device deploys a lightweight contact network component recognition neural network to calculate the comprehensive anomaly score in real time; the comprehensive anomaly score is a weighted sum of visual defect confidence, normalized temperature anomaly, and normalized geometric parameter offset; the original observation segment is compressed and uploaded to the ground server only when the comprehensive anomaly score is greater than a preset threshold or the image quality index is lower than a preset threshold.

9. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 8, characterized in that, After completing the diagnosis, the ground server sends the updated diagnostic results and differential neural network parameters to the vehicle-mounted edge computing device to realize the online incremental evolution of the model.

10. The intelligent analysis system for the operating status of 3C vehicle-mounted overhead contact lines according to claim 1, characterized in that, It also includes a diagnostic result visualization and review interface, which dynamically presents historical status curves, multi-source raw data linkage views, hybrid decision-making intermediate logic display, and topology consistency verification summary.

Images