A remote track fault intelligent detection system based on big data driving

Abstract

This invention discloses a remote intelligent track fault detection system driven by big data, belonging to the field of intelligent monitoring and condition maintenance technology for rail transit. It solves the problems of traditional detection methods, such as heavy reliance on manual labor, slow response, and lack of predictive and self-repair capabilities. This invention collects multi-source data through a fixed sensor network and a mobile inspection vehicle, constructs and updates a multi-scale digital twin, and generates a probabilistic fault evolution map based on a stochastic phase field model. It relies on distributed edge computing nodes for collaborative judgment, uses a hybrid geometric distance retrieval immune memory library to generate maintenance plans, and finally performs collaborative repair through mixed reality and a cluster of microrobots, achieving continuous system optimization based on federated learning and case-based reasoning. This system realizes an integrated closed loop of track condition perception, fault prediction, intelligent decision-making, and autonomous repair, improving detection predictability, maintenance accuracy, and system adaptability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and condition maintenance technology for rail transit, specifically a remote intelligent detection system for track faults based on big data. Background Technology

[0002] With the rapid development of modern rail transit, the safety, reliability, and operational efficiency of railway transportation have become increasingly important issues concerning public safety and economic operation. As the fundamental load-bearing structure for train operation, tracks are susceptible to various faults such as cracks, deformation, and internal damage due to multiple factors including load, environment, and material aging during long-term service. Traditional manual inspections and periodic maintenance methods are not only inefficient and have limited coverage, but also struggle to achieve early warning and accurate prediction of track faults. Especially in high-load, long-distance, and complex track networks, achieving real-time status monitoring and intelligent diagnosis across all time periods and sections has become a critical need that the industry urgently needs to overcome. The limitations of existing fault solutions have shown that single-technology improvements cannot systematically address the entire chain of intelligent track fault detection requirements. Only through the organic integration and collaborative innovation of multiple technologies can current technological bottlenecks be overcome. This invention is based on this understanding and constructs an integrated solution with strong logical connections between its various technical modules.

[0003] The following problems exist in the existing technology: Traditional methods rely on manual inspections or single-type sensors, making it difficult to achieve all-weather, all-road-section, and multi-dimensional (vibration, images, internal defects, etc.) collaborative perception and data fusion; traditional centralized data processing models have high response delays when faced with massive amounts of monitoring data, making it difficult to meet real-time requirements. Traditional models are mostly static or deterministic, which cannot be dynamically updated with real-time data. They often assume that the material is uniform and ignore the "butterfly effect" of the randomness of the material's microstructure on the macroscopic fracture behavior. This makes it impossible to quantify the impact of random factors on damage evolution, resulting in low reliability of prediction results and insufficient early warning accuracy. Existing technologies rely on central servers for fault diagnosis, which can easily become single points of failure. Maintenance decisions often depend on human experience or repetitive calculations, lacking a "memory-response" mechanism similar to the biological immune system. This makes it impossible to reuse historically successful maintenance solutions, resulting in significant waste of computing resources, slow response speed, and poor adaptability. Traditional fault repair techniques mainly rely on manual on-site operations, which are inefficient, risky, and difficult to guarantee accuracy. They also lack technical means for multi-robot collaborative operation and intuitive visual guidance. Traditional systems lack the ability to continuously evolve. Once deployed, models and knowledge bases are difficult to update. They cannot continuously optimize by utilizing new data and repair feedback generated during operation and maintenance while protecting data privacy and breaking down data silos, resulting in model aging and knowledge base lag. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a remote intelligent track fault detection system based on big data to solve the above-mentioned technical problem.

[0005] The first aspect of this invention provides a remote intelligent track fault detection system based on big data, comprising a remote data processing center and the following modules: Track status perception and data acquisition module: Through a distributed intelligent sensor network fixedly deployed on the track, and an unmanned inspection vehicle that can move along the track and is equipped with a multi-sensor fusion detection unit, it collaboratively collects multi-source heterogeneous data of the track. Fault Evolution Inference and Digital Twin Module: Constructs and updates multi-scale digital twins based on multi-source heterogeneous orbital data, and generates probabilistic fault evolution maps through an inference engine integrating stochastic phase field theory; Intelligent collaborative judgment and decision generation module: Constructs and maintains an immune memory library, performs collaborative fault judgment through a distributed edge computing node network based on the probabilistic fault evolution map, and queries the immune memory library using a metric based on hybrid geometric distance to generate fault repair decision schemes; Human-computer interaction and autonomous repair execution module: Transforms the fault repair decision-making scheme into realistic visual guidance and robot control commands, driving the mixed reality equipment and the cluster of micro repair robots to perform collaborative repair operations; The system's autonomous learning and optimization evolution module collects repair operation results and orbital status verification data, updates the multi-scale digital twin through federated learning, and optimizes the immune memory bank.

[0006] Preferably, the method of collecting heterogeneous data from multiple sources on the track through a distributed intelligent sensor network fixedly deployed on the track includes the following steps: The distributed intelligent sensing network consists of multiple intelligent sensing nodes fixedly installed on the track surface; each intelligent sensing node integrates a vibration sensor, a temperature sensor and a strain sensor, which are used to continuously collect vibration, temperature and micro-strain data of the track body, and upload the data to a remote data processing center after preprocessing. The intelligent sensing node determines whether a train is approaching based on the energy change and rate of change of the track vibration signal in the characteristic frequency band when the train passes. If the train is approaching, the current intelligent sensing node is automatically awakened. The awakened intelligent sensing node then wakes up the downstream intelligent sensing nodes in sequence along the track direction through surface acoustic wave communication.

[0007] Preferably, the process involves collecting heterogeneous multi-source data about the track using an unmanned inspection vehicle equipped with a multi-sensor fusion detection unit that can move along the track, including the following steps: The multi-sensor fusion detection unit carried by the unmanned inspection vehicle includes an image sensor, a laser sensor, an ultrasonic sensor, and an infrared sensor. The unmanned inspection vehicle performs inspection tasks according to a preset cycle, and autonomously cruises along the track and simultaneously collects track surface images, geometric deformation data of track profile and gauge, ultrasonic detection data reflecting internal defects, and temperature field distribution data of track surface according to the scheduling instructions of the remote data processing center.

[0008] The unmanned inspection vehicle has a built-in edge computing unit and uses a multimodal data fusion algorithm to perform real-time preprocessing and feature extraction on the collected image, geometric, ultrasonic and thermal data to generate a track health index vector. The track health index vector includes a crack risk index, a geometric distortion index, an internal defect index and a thermal anomaly index. The unmanned inspection vehicle uploads the track health index vector to a remote data processing center through a vehicle-to-ground communication link.

[0009] Preferably, constructing and updating a multi-scale digital twin based on multi-source heterogeneous orbital data includes the following steps: Construct a multi-scale digital twin model framework that includes macroscopic, mesoscopic, and microscopic scales; A macroscopic dynamic model of the track structure is constructed, which describes the overall mechanical behavior of the track structure based on beam theory. A microscopic damage model for the evolution of track damage is constructed. The microscopic damage model is described based on phase field theory and introduces a phase field variable to characterize the structural damage state. By integrating macroscopic dynamics models, mesoscopic models, and microscopic damage models, a multi-scale digital twin was constructed and deployed in a remote data processing center. Among them, the mesoscopic model serves as a bridge between the macroscopic and microscopic scales, used to describe the influence of the microstructure of the track material on the macroscopic mechanical properties and the initial initiation of damage. Based on the vibration, temperature, and micro-strain data collected by the distributed intelligent sensor network uploaded by the intelligent sensor nodes, and the track health index vector collected and generated by the unmanned inspection vehicle, the multi-scale digital twin is updated using a data assimilation method to correct its key internal states that reflect the mechanical behavior and damage state of the track structure.

[0010] Preferably, generating a probabilistic fault evolution map using a deduction engine integrating stochastic phase field theory includes the following steps: Based on the updated multi-scale digital twin, the inference engine deployed in the remote data processing center extracts its phase field variables as the initial state for inference; Fault evolution simulation was performed using a stochastic phase-field model via the Monte Carlo method. The damage evolution equation of the stochastic phase-field model includes a deterministic damage evolution term and a random noise term used to quantify the uncertainty of the microstructure of the orbital material. Where d is the phase field variable, The regularization length parameter is used to control the bandwidth of damage localization. M is a parameter related to the fracture energy of a material, and is the mobility coefficient. Let D be the strain energy of the tensile portion, and D be the noise intensity. For random noise, x is the spatial coordinate along the orbital direction, and t is time; Based on the stochastic phase field model, the inference engine executes the Monte Carlo simulation to calculate the damage values ​​at each discrete spatial location of the multi-scale digital twin at different future times, and calculates the probability that the damage values ​​exceed a preset threshold to generate a probabilistic fault evolution map.

[0011] Preferably, an immune memory bank is constructed and maintained. Based on the probabilistic fault evolution map, fault collaborative judgment is performed through a distributed edge computing node network. The immune memory bank is queried using a metric based on hybrid geometric distance to generate a fault repair decision scheme, including the following steps: Based on the probabilistic fault evolution graph, fault collaborative judgment is performed through a distributed edge computing node network deployed in a remote data processing center; the fault collaborative judgment adopts a practical Byzantine fault-tolerant consensus mechanism. An immune memory bank is constructed and maintained to store historical track fault cases and corresponding maintenance plans formed from historical track fault judgments and maintenance records; track fault patterns are matched based on hybrid geometric distance, and the hybrid geometric distance... The calculation formula is: Where P and Q are the probability distributions of the current orbital failure mode and the historical orbital failure mode, respectively; Let KL divergence be the KL divergence. For Wasserstein distance, For Bures distance, and These are the weighting coefficients; Based on the consensus results of the collaborative fault judgment, the immune memory library is queried to perform track fault mode matching, and a track fault maintenance decision scheme is generated by solving a multi-objective optimization problem.

[0012] Preferably, the fault repair decision-making scheme is transformed into realistic visual guidance and robot control commands to drive the mixed reality device and the micro repair robot cluster to perform collaborative repair operations, including the following steps: Based on the fault repair decision scheme and the probabilistic fault evolution map, a realistic visualization guide and control instructions for a cluster of micro-repair robots are generated; the realistic visualization guide is obtained based on a multi-scale digital twin and the fault repair decision scheme. The cluster of miniature repair robots executes repair tasks according to control commands and employs a cooperative strategy. This cooperative control strategy is achieved by solving the following optimization problem: The constraints are satisfied:

[0013] in, The total number of robots in the micro-repair robot swarm. and These are the state vector and control input of the k-th robot, respectively. , The harmonic coefficient, Let the desired trajectory of the k-th robot be... For integration time variable, The value function obtained through collaborative learning, Let K be the state vector of the k-th robot in collaborative learning. For the current moment, To predict the length of the time domain, for The derivative of Let G be the set of neighboring robots of the k-th robot, and let G represent the constraints of the robot cluster. For the index of neighboring robots, Let W be the set of constraints for the control input of the k-th robot, and let W and S be positive definite weighting matrices. The human-computer interaction and autonomous repair execution module is deployed at the track site, driving the mixed reality device to present the real-world visual guidance and driving the cluster of micro-repair robots to perform collaborative repair operations.

[0014] Preferably, the process of collecting repair operation results and orbital status verification data, updating the multi-scale digital twin and optimizing the immune memory bank through federated learning includes the following steps: The system collects the repair operation results of the human-computer interaction and autonomous repair execution module, as well as the track status verification data of the track status perception and data acquisition module. The track status verification data includes vibration, temperature and micro-strain data collected by the distributed intelligent sensor network, and track health index vector collected and generated by the unmanned inspection vehicle. Based on orbital state verification data, a federated learning framework is used to update multi-scale digital twins deployed in remote data processing centers; Based on the repair operation results, new track fault cases are generated, and the immune memory bank is optimized using an incremental learning mechanism based on case reasoning. When the difference between the new track fault case and the historical track fault cases in the immune memory bank exceeds a preset novelty threshold, the new track fault case is added to the immune memory bank.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves high-frequency, multi-dimensional, low-power all-weather monitoring and data fusion of track status through a collaborative architecture of fixed sensor network and mobile inspection vehicle, combined with intelligent wake-up and edge computing. Based on multi-scale digital twins and random phase field theory, this invention can assimilate real-time data and update it dynamically. It generates probabilistic fault evolution maps through Monte Carlo simulation, which significantly improves the foresight and scientific nature of track fault prediction. This invention employs a distributed edge computing node consensus mechanism to improve the robustness and fault tolerance of fault judgment; by constructing an immune memory library and using hybrid geometric distance for intelligent matching, it can quickly generate optimized track fault maintenance decision schemes. This invention provides intuitive visual guidance through mixed reality technology and drives a cluster of micro-repair robots to perform collaborative optimization operations, significantly improving the automation level, accuracy and efficiency of track fault repair operations; This invention utilizes a federated learning framework to update multi-scale digital twin models and employs an incremental learning mechanism to optimize the immune memory bank, enabling the system to continuously accumulate knowledge and improve itself as it is used, thus achieving continuous evolution in its level of intelligence. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Please see Figure 1This invention is a remote intelligent detection system for track faults based on big data, comprising a remote data processing center and the following modules: Track status perception and data acquisition module: Through a distributed intelligent sensor network fixedly deployed on the track, and an unmanned inspection vehicle that can move along the track and is equipped with a multi-sensor fusion detection unit, it collaboratively collects multi-source heterogeneous data of the track. Fault Evolution Inference and Digital Twin Module: Constructs and updates multi-scale digital twins based on multi-source heterogeneous orbital data, and generates probabilistic fault evolution maps through an inference engine integrating stochastic phase field theory; Intelligent collaborative judgment and decision generation module: Constructs and maintains an immune memory library, performs collaborative fault judgment through a distributed edge computing node network based on the probabilistic fault evolution map, and queries the immune memory library using a metric based on hybrid geometric distance to generate fault repair decision schemes; Human-computer interaction and autonomous repair execution module: Transforms the fault repair decision-making scheme into realistic visual guidance and robot control commands, driving the mixed reality equipment and the cluster of micro repair robots to perform collaborative repair operations; The system's autonomous learning and optimization evolution module collects repair operation results and orbital status verification data, updates the multi-scale digital twin through federated learning, and optimizes the immune memory bank.

[0019] Specifically, the system first collects multi-source heterogeneous data from the track through a distributed intelligent sensor network fixed on the track and a mobile unmanned inspection vehicle, including vibration, temperature, strain, images, and ultrasound data, and performs real-time preprocessing and feature extraction. Second, based on the collected data, a multi-scale digital twin is constructed and continuously updated. This twin integrates macroscopic dynamics and microscopic damage evolution models, realistically reflecting the mechanical behavior and damage state of the track structure. On this basis, a deduction engine integrating stochastic phase field theory is used to simulate damage evolution using the Monte Carlo method, generating a probabilistic fault evolution map. Next, the system uses a distributed edge computing node network to collaboratively judge faults and, combined with historical fault cases stored in an immune memory bank, performs similarity matching through hybrid geometric distance to generate targeted fault repair decision schemes. Then, the repair decision schemes are transformed into visual guidance and robot control commands, presented to operators through mixed reality devices, and drive a cluster of micro-repair robots to execute repair tasks according to a collaborative control strategy. Finally, based on the repair results and subsequent orbital status verification data, the system updates the digital twin using a federated learning framework and optimizes the immune memory bank through a case-based reasoning mechanism, thereby achieving continuous autonomous learning and performance evolution of the system.

[0020] In one embodiment of the present invention, collecting multi-source heterogeneous data from the track through a distributed intelligent sensor network fixedly deployed on the track includes the following steps: The distributed intelligent sensing network consists of multiple intelligent sensing nodes fixedly installed on the track surface; each intelligent sensing node integrates a vibration sensor, a temperature sensor and a strain sensor, which are used to continuously collect vibration, temperature and micro-strain data of the track body, and upload the data to a remote data processing center after preprocessing. The intelligent sensing node determines whether a train is approaching based on the energy change and rate of change of the track vibration signal in the characteristic frequency band when the train passes. If the train is approaching, the current intelligent sensing node is automatically awakened. The awakened intelligent sensing node then wakes up the downstream intelligent sensing nodes in sequence along the track direction through surface acoustic wave communication.

[0021] Specifically, the distributed intelligent sensing network consists of multiple intelligent sensing nodes. Each intelligent sensing node uses a housing with a protection rating of at least IP67 and is directly fixed to the non-working surface (i.e., the surface not in direct contact with the wheels) of the rail web or base using high-strength, weather-resistant epoxy resin adhesive. The installation spacing L is determined based on a comprehensive consideration of monitoring accuracy requirements, track importance, and cost, with a typical range of 20 to 50 meters. Each intelligent sensing node constitutes an independent embedded monitoring unit, whose hardware core includes a sensing unit, a processing and communication unit, and a power supply unit. The sensing unit integrates a triaxial MEMS (Micro-Electro-Mechanical Systems) accelerometer (range ±50g, bandwidth 0-500Hz), a patch-type digital temperature sensor (measurement range −40°C to 125°C), and a miniature resistance strain gauge (grid length 5mm, accuracy...). The processing and communication unit employs a low-power microcontroller responsible for data preprocessing, local algorithm execution, and communication control. Communication methods include long-range wireless modules (such as LoRa or 5G NR-Light) for communicating with remote data processing centers, and surface acoustic wave (SAW) communication units for collaboration between intelligent sensing nodes.

[0022] The microcontroller uses a sampling frequency (This frequency is sufficient to cover the main vibration frequency band of the track and satisfies the Nyquist sampling theorem.) Synchronously trigger the triaxial accelerometer, temperature sensor, and strain gauge to acquire data and obtain the raw signal, including vibration acceleration. (Define the z-axis perpendicular to the track plane and pointing upwards, the x-axis along the track direction, and the y-axis laterally), track temperature raw strain gauge readings Preprocessing of the original signal; processing of the vertical acceleration components. A fourth-order Butterworth bandpass filter (passband frequency 0.5Hz–200Hz) was applied; orbital temperature was utilized. The temperature-strain compensation coefficient function was obtained in advance in the laboratory. Real-time compensation is performed on the original strain readings to obtain the true mechanical strain signal. Microscopic strain data refers to mechanical strain after temperature compensation. .

[0023] The intelligent sensing nodes are in a low-power sleep state by default and are only awakened by vibration signals. The pre-processed vertical acceleration components... Time-frequency analysis was performed using wavelet packet decomposition algorithm. A three-level wavelet packet decomposition was applied to select and reconstruct the signal components corresponding to the characteristic frequency band of 5Hz-15Hz. The selection criteria for the characteristic frequency band 5Hz-15Hz are as follows: A typical section (e.g., one straight section and one curved section) was selected on-site. Vertical vibration signals from multiple passing trains were collected using temporary vibration sensors. Power spectral density analysis of the signals revealed that the main vibration energy during train passage was concentrated in the 5Hz-15Hz range (caused by wheel-rail contact, track irregularities, etc.). Therefore, this frequency band was chosen as the characteristic frequency band for identifying train approach. The characteristic frequency band 5Hz-15Hz was calculated within a sliding time window (window length...). Energy within) And the total energy of the acceleration signal within the previous time window. : , Where t is the current time; Let be the integration variable, representing the integral from Each time point traversed during the cumulative calculation within the time interval up to time t. Setting a dynamic threshold. (Typical value is 0.35) and (Typical value is 10J / s); In the initial stage, the system performs internal logic verification on events based on whether downstream nodes are awakened sequentially and the spatiotemporal continuity characteristics of vibration signals, automatically labels them as train passing or non-train interference samples, and statistically analyzes the characteristic frequency band energy ratios for the two types of events: train passing and non-train interference. ) and the absolute value of its rate of change ( The distribution of the detection rate is used to determine the optimal threshold that balances high detection rate and low false alarm rate. and Specifically, by collecting no fewer than 500 sets of train passing samples and 500 sets of non-interference samples (such as pedestrians, small machinery operations, etc.), the energy ratio and rate of change of the two types of samples are calculated respectively, ROC curves are plotted, and the threshold that maximizes the Youden's index is selected as the threshold. and The initial value can be adaptively adjusted based on actual operating data. When both conditions are met... and When a train is detected approaching, the intelligent sensor node is immediately activated.

[0024] A chain-like wake-up mechanism based on surface acoustic waves (SAW) is used to wake up the intelligent sensor nodes; the first intelligent sensor node to detect the approach of the train (denoted as node 1, trigger time) Upon being awakened, the SAW immediately transmits a specific frequency wake-up pulse signal along the track medium to the downstream smart sensor node in the direction of train travel via its SAW communication unit. The propagation speed of the SAW along the track... The value is fixed (approximately 3000 m / s for typical rail materials), and the signal processing and communication preparation delay for each smart sensor node is... (Empirical parameters determined through experimental calibration and field testing, typical value 5ms) can be calibrated. The calibration method is as follows: Under laboratory conditions, take a section of rail of known length, emit a surface acoustic wave pulse at one end, receive and record the propagation time at the other end, and take the average value after multiple measurements; during field testing, the pulse round-trip time can be measured between the installed nodes for calibration. The time difference between receiving the wake-up pulse and completing preparation to enter the data acquisition state at the measurement node is used to obtain the value, and the average of 10 measurements is taken as a fixed value. This is the time when the nth downstream smart sensor node is woken up. for: Where L represents the spacing between intelligent sensing nodes. The awakened intelligent sensing nodes immediately enter working mode, in order to... The sampling frequency continuously collects acceleration, temperature, and micro-strain data, and performs local preprocessing. During and for a period after the train's passage (e.g., 30 seconds), the data is temporarily cached. This dual-threshold judgment based on vibration energy characteristics, combined with the surface acoustic wave (SAW) communication chain-like wake-up mechanism, collaboratively solves the power consumption problem of the distributed sensor network under continuous monitoring. Wake-up is triggered by changes in characteristic frequency band energy when the train approaches, avoiding prolonged high-power operation of the sensors. Simultaneously, SAW communication utilizes the track medium as a transmission carrier, eliminating the need for additional wireless communication modules and significantly reducing communication power consumption. Calculations show that this mechanism can reduce the overall power consumption of the sensor network by approximately 65%-78%, extending the battery replacement cycle from the conventional 3-6 months to over 18 months, achieving a balance between all-weather monitoring and low-power operation.

[0025] After the train passes, the awakened intelligent sensor node cluster, through a self-organizing Time Division Multiple Access (TDMA) protocol, sequentially uploads the cached, timestamped, pre-processed data packets to the remote data processing center via a long-distance wireless communication link. The data packet content includes at least the intelligent sensor node ID (a unique identifier for each node, usually preset at the factory or written during installation and deployment using on-site configuration tools), geographic location coordinates (obtained and stored on-site using the GNSS receiver carried by the intelligent sensor node during installation, or directly imported into the system database from the construction drawings for association), and time-series data (…). ), and relevant characteristic values ​​for train approach determination (such as Once the upload is complete, all smart sensor nodes automatically return to a low-power sleep state, waiting for the next trigger.

[0026] In one embodiment of the present invention, a multi-source heterogeneous data collection method for a track is employed, which involves collecting track-based multi-sensor fusion detection unit via an unmanned inspection vehicle capable of moving along the track, including the following steps: The multi-sensor fusion detection unit carried by the unmanned inspection vehicle includes an image sensor, a laser sensor, an ultrasonic sensor, and an infrared sensor; The unmanned inspection vehicle performs inspection tasks according to a preset cycle, and autonomously cruises along the track and simultaneously collects track surface images, geometric deformation data of track profile and gauge, ultrasonic detection data reflecting internal defects, and temperature field distribution data of track surface according to the scheduling instructions of the remote data processing center. The unmanned inspection vehicle has a built-in edge computing unit and uses a multimodal data fusion algorithm to perform real-time preprocessing and feature extraction on the collected image, geometric, ultrasonic and thermal data to generate a track health index vector. The track health index vector includes a crack risk index, a geometric distortion index, an internal defect index and a thermal anomaly index. The unmanned inspection vehicle uploads the track health index vector to a remote data processing center through a vehicle-to-ground communication link.

[0027] Specifically, the unmanned inspection vehicle consists of a mobile platform, a multi-sensor fusion detection unit, a central control and edge computing unit, a communication module, and a power system. The mobile platform uses a dedicated track-based mobile chassis with four-wheel independent drive and independent steering. This chassis integrates a high-precision odometer, an IMU (Inertial Measurement Unit), a GNSS (Global Navigation Satellite System) receiver, and a LiDAR (Light Detection and Ranging) system. The multi-sensor fusion detection unit is rigidly mounted at the center of the mobile platform, maintaining the detection direction perpendicular to the track via a servo gimbal. The multi-sensor fusion detection unit specifically includes a linear CCD image sensor, a laser displacement sensor, an ultrasonic surface wave flaw detector, and an infrared thermal imager. A linear CCD image sensor (resolution ≥ 2048 pixels) is used to acquire high-resolution grayscale images of the track surface (rail head, rail web, rail bottom); a laser displacement sensor (sampling frequency 10kHz, measurement range ±50mm, accuracy ±0.01mm) uses a laser profilometer based on the triangulation principle to measure the geometric deformation of the track cross-section profile; an ultrasonic surface wave flaw detector (center frequency 5MHz, pulse repetition frequency 1kHz) uses a pitch-catch mode angled probe to excite ultrasonic surface waves on the track surface at a critical angle to detect near-surface (depth 0-10mm) defects such as cracks and peeling; an infrared thermal imager (resolution 640×480, thermal sensitivity <50mK) is used to measure the temperature field distribution on the track surface. The central control and edge computing unit uses a high-performance embedded computer (such as based on the NVIDIA Jetson AGX Orin platform) to control the operation of the entire unmanned inspection vehicle, including navigation of the mobile platform, synchronous triggering of sensor data acquisition, real-time data processing and feature extraction, and communication with a remote data processing center. The communication module uses a 5G private network or a broadband wireless local area network. The power system uses a high-energy-density lithium battery pack.

[0028] The unmanned inspection vehicle autonomously cruises along the track according to a preset cycle (e.g., weekly) or real-time dispatch instructions from a remote data processing center. Based on feedback from the high-precision odometer, the central controller synchronously triggers all sensors along the track at fixed spatial intervals (e.g., 0.1 meters) or fixed time intervals (e.g., 0.01 seconds) to collect data. The collected raw data includes high-resolution image sequences of the track surface, track profile and gauge geometric deformation data, ultrasonic A-scan waveform signals, and track surface temperature field distribution. (High-resolution image sequence of track surface) Image data acquired by a linear CCD image sensor, wherein The coordinates are the mileage coordinates along the track (obtained by fusion of odometer and GNSS data). These are the horizontal coordinates perpendicular to the track and the vertical coordinates along the track within the image plane (for a linear CCD, each trigger produces one line of image data, which is then stitched together to form a two-dimensional image). Track contour and gauge geometric deformation data. Data is collected by a laser displacement sensor, which scans the track cross-section to obtain a series of discrete point clouds. Key geometric parameters, such as rail head profile, rail web thickness, rail base width, and track gauge, can be extracted through fitting. Key geometric parameters. Ultrasound A-scan waveform signal. The ultrasonic surface wave flaw detector (SSWF) acquires data, recording the received ultrasonic echo signal (A-scan waveform) each time it is triggered. The time axis t corresponds to the propagation time of the sound wave in the track material. Track surface temperature field distribution. Acquired by an infrared thermal imager, the device captures one frame of thermal image each time it is triggered, reflecting the temperature distribution in two-dimensional space on the orbital surface. Simultaneously, it records current environmental information, including ambient temperature. Humidity (H), light intensity And the train's historical timestamps.

[0029] Edge computing units perform real-time data preprocessing and feature extraction. For image data... Denoising (median filtering), illumination unevenness correction, and image enhancement (CLAHE) are performed to extract the region of interest (ROI) on the track surface; geometric point cloud data is then processed. Outlier removal and filtering smoothing are performed, and the RANSAC algorithm is used to fit the standard track profile curve to calculate the deviation between the actual profile and the standard profile. ; for ultrasound A-scan waveform signals Bandpass filtering, DC removal, and envelope detection (Hilbert transform) are performed to extract defect echo characteristics (amplitude, time of arrival, width); thermal imaging data are analyzed. Non-uniformity correction and radiometric calibration are performed to convert the data into true temperature values. A custom multi-branch convolutional neural network (CNN) is used for end-to-end feature extraction and health index calculation; the image branch takes a ROI image patch (e.g., 256×256 pixels) on the track surface as input and uses a ResNet-18 backbone network to extract depth feature vectors related to surface cracks, indentations, and corrosion. Geometric branches use a geometric bias sequence with a fixed length of resampling. As input, a one-dimensional CNN is used to extract geometric feature vectors related to track wear, deformation, and misalignment. ; Preprocessed A-scan waveforms of ultrasound branches with fixed lengths As input, a one-dimensional CNN combined with an attention mechanism is used to extract ultrasonic feature vectors related to internal defects (such as cracks and inclusions). The thermal imaging branch takes temperature field image patches as input and uses a lightweight CNN (such as MobileNetV2) to extract thermal feature vectors related to thermal anomalies (such as frictional hot spots and temperature differences caused by internal damage). Four feature vectors ( , , , It calculates the correlation between modalities through a cross-attention mechanism and outputs a unified fusion feature vector. Fusion of feature vectors A 4-dimensional orbital health index vector is generated by mapping through a fully connected layer (output dimension 4) and a sigmoid activation function. The crack risk index CRI ∈ [0,1] reflects the overall probability of crack-like defects on and near the track surface; the higher the value, the greater the risk. The geometric distortion index GI ∈ [0,1] reflects the degree of geometric anomalies such as track profile deformation, wear, and gauge exceeding limits. The internal defect index DFI ∈ [0,1] reflects the severity of defects such as inclusions, pores, and fatigue cracks inside the track material. The thermal anomaly index TI ∈ [0,1] reflects the degree of abnormal temperature field distribution on the track surface (such as local overheating or overcooling), which may be related to internal damage or abnormal contact conditions.

[0030] The training data for the multi-branch convolutional neural network was obtained from a typical 200-kilometer section (including straight sections, curves, turnouts, bridges, and other working conditions) on a railway line under the jurisdiction of a railway bureau. Multimodal data was continuously collected for six months using an unmanned inspection vehicle. Professional flaw detectors manually annotated the collected images, ultrasonic data, and thermal images, including crack location and length, geometric deformation type and degree, and internal defect type, resulting in 120,000 valid annotated samples. These samples were divided into training, validation, and test sets in an 8:1:1 ratio for network training and performance evaluation. On the test set, the values ​​of each index in the health index vector were compared with the manually detected values. The correlation coefficient between CRI and manually annotated crack length reached 0.92, GI and gauge exceedance was 0.88, DFI and ultrasonic echo amplitude was 0.85, and TI and abnormal temperature rise in infrared thermal imaging was 0.90, indicating that each index accurately reflects the corresponding track health status.

[0031] During routine inspections, the unmanned inspection vehicle only uploads the track health index vector h and its position coordinates for each sampling point (i.e., the discrete measurement position corresponding to when the inspection vehicle triggers the sensor to collect data at fixed spatial or time intervals) or for each fixed interval (e.g., 1 meter). When any health index ( When the threshold exceeds the preset Level 1 alarm threshold (e.g., 0.7, which is set based on historical data statistics, a balance between fault identification rate and false alarm rate, and engineering experience), an abnormal mode is automatically triggered. In addition to the health index, a snapshot of the original multimodal data (compressed image, geometric contour, ultrasonic waveform, and thermal image) from the vicinity of the location (e.g., 0.5 meters before and after) is packaged and uploaded. The system can also respond to instructions from the remote data processing center to upload complete original data for a specified segment. All data is uploaded to the remote data processing center via the vehicle-to-ground communication link (using a 5G private network or broadband wireless LAN). The data packet format includes a frame header (timestamp, location, vehicle ID), a data body (health index or original data), and a frame tail (CRC checksum).

[0032] In one embodiment of the present invention, constructing and updating a multi-scale digital twin based on multi-source heterogeneous orbital data includes the following steps: Construct a multi-scale digital twin model framework that includes macroscopic, mesoscopic, and microscopic scales; A macroscopic dynamic model of the track structure is constructed, which describes the overall mechanical behavior of the track structure based on beam theory. A microscopic damage model for the evolution of track damage is constructed. The microscopic damage model is described based on phase field theory and introduces a phase field variable to characterize the structural damage state. By integrating macroscopic dynamics models, mesoscopic models, and microscopic damage models, a multi-scale digital twin was constructed and deployed in a remote data processing center. Among them, the mesoscopic model serves as a bridge between the macroscopic and microscopic scales, used to describe the influence of the microstructure of the track material on the macroscopic mechanical properties and the initial initiation of damage. Based on the vibration, temperature, and micro-strain data collected by the distributed intelligent sensor network uploaded by the intelligent sensor nodes, and the track health index vector collected and generated by the unmanned inspection vehicle, the multi-scale digital twin is updated using a data assimilation method to correct its key internal states that reflect the mechanical behavior and damage state of the track structure.

[0033] Specifically, a multi-scale digital twin model framework is constructed, encompassing macroscopic, mesoscopic, and microscopic scales. At the macroscopic scale (meter level), the track structure is treated as a continuous medium beam, describing its overall dynamic response (such as displacement and vibration) under train loads and environmental excitations. At the mesoscopic scale (millimeters to centimeters level), serving as a macro-micro transition bridge, the influence of material microstructures (such as grains and inclusions) on macroscopic mechanical properties and the initial initiation of damage is described. At the microscopic scale (micrometers to millimeters level), based on phase-field theory, the nucleation, evolution, merging, and eventual formation of macroscopic cracks in track material microstructures (such as grains and inclusions in rail materials) are described. These three scale models form a bidirectional interactive whole through a multi-scale coupling algorithm. The mechanical field of the macroscopic model drives the evolution of microscopic damage, while the microscopic damage state is corrected in real-time by a homogenization method to modify the track material properties of the macroscopic model. Specifically, the microscopic damage model of this system is not a traditional deterministic phase-field model, but a stochastic phase-field model. Its core lies in introducing a Gaussian white noise term to quantify the uncertainty of the material's microstructure (such as the randomness of grain boundaries and impurity distribution). This noise term directly affects the phase-field variables. The evolution of damage makes the damage evolution process inherently random, which is not present in traditional digital twins.

[0034] The macroscopic dynamic model is constructed based on the modified Timoshenko beam theory, treating the track structure as a continuous beam laid on an elastic foundation. It considers the effects of shear deformation and rotational inertia to describe the propagation and local deformation behavior of short-wavelength track waves. The model defines the spatial coordinate along the longitudinal direction of the track as x (this coordinate is the distance coordinate along the track), and time as t. This represents the vertical displacement of the neutral axis of the track beam (downward is positive), using... The angle of rotation of the cross section is represented; its dynamic behavior is described by two coupled partial differential equations: , ,in, The density of the track material (kg / m³) 3 A is the cross-sectional area of ​​the track (m). 2 ), The moment of inertia (m) of the track cross section about its horizontal neutral axis 4 E represents the elastic modulus (Pa) of the track material, and ν represents Poisson's ratio. These parameters have initial values ​​derived from track design drawings and material manuals. G represents the shear modulus of the rail material. ; This is a shear correction factor, which is related to the cross-sectional shape. For a typical rail cross-section, its value is approximately 0.85–0.95. The lateral distributed loads acting on the track mainly originate from the vertical forces of the train wheels. Their temporal history and spatial distribution can be analyzed by inversely analyzing vibration acceleration. or calculated based on the vehicle-track coupling dynamics model; The distributed bending moment acting on the track comes from creep force between wheel and rail, eccentric load and other factors. Among them, the shear correction factor can be calculated according to the cross-section geometric parameters or determined through experimental calibration. For example, select a section with a length of 5m in the track test section, install a plurality of displacement sensors, apply a known static transverse load, record the deflection and rotation angle of the beam, and perform inverse calculation by using the finite element model to minimize the error between the model prediction and the actual measurement, repeat 3 times and take the average value as the initial value of the line section. The initial values of track material and geometric parameters from design data are adopted during modeling, and these parameters will be dynamically updated through the subsequent Ensemble Kalman Filter (EnKF) data assimilation process. The model adopts the finite element method for spatial discretization, and uses the Newmark-β method for time integration. Through solving on the high-performance computing cluster in the remote data processing center, the displacement of any position x on the whole track line at any time t is obtained and the rotation angle .

[0035] The microscopic damage model is constructed based on the random phase field theory, which is used to describe the damage evolution inside the rail steel material. This model introduces a continuous phase field variable , where is the three-dimensional space coordinate; the model mainly focuses on the damage along the depth direction (y) and longitudinal direction (x) of the track, and establishes a two-dimensional (x-y plane) model; the phase field variable is defined as a scalar field, where d=0 indicates the material is intact, d=1 indicates complete fracture, and 0<d<1 indicates the damage state with microcrack aggregation. The total free energy U of the system consists of elastic strain energy and damage dissipation energy: ; wherein, is the strain tensor, which is obtained from the displacement field calculated by the macroscopic model , the local strain is derived through the geometric equation; under the two-dimensional plane assumption, the strain components can be calculated according to the deformation geometry of the beam, for example , where y is the distance from the neutral axis. is the elastic strain energy density, which is decomposed into the tensile part that drives damage evolution by adopting a method based on strain spectrum decomposition and the compression part that does not cause damage ; the specific decomposition method is: , wherein and are Lamé constants, which are calculated from the elastic modulus E and Poisson's ratio ν of the track material, that is ; represents taking a positive value, is the tensile part of the strain tensor; . For stiffness degradation function, usually take This makes the tensile stiffness of the material approach 0 when d approaches 1. The critical energy release rate (fracture toughness, J / m²) of the track material is obtained through standard fracture toughness tests on the materials used in that section of track, or from a material handbook. Specifically, according to ASTM E399 (Standard Test Method for Linear Elastic Plane Strain Fracture Toughness of Metallic Materials), compact tensile (CT) specimens are taken from the same batch of rails for fracture toughness testing, and the plane strain fracture toughness is measured. The value is then determined according to the plane strain formula. Converted . The regularization length parameter (in meters) is a numerical parameter that controls the width of the transition zone from 0 to 1 in terms of damage, and is related to the characteristic lengths within the material (such as grain size); in numerical calculations, The value of needs to ensure that the localized damage region can be analyzed by the mesh, and is usually related to the finite element mesh size. satisfy The calibration was performed by matching the theoretical fracture energy; the specific calibration method was as follows: numerical simulation of a single-sided notched tensile specimen was used to adjust... To ensure that the simulated peak load matches the theoretical solution, typically... Take 2 to 5 times the grid size, and after calibration The value is set to 2mm (corresponding to a grid size of 0.5mm). The integration region is the area of ​​track material being analyzed, and its range is determined by the deployment location of intelligent sensor nodes, the detection range of unmanned inspection vehicles, or the section specified by fault simulation. The spatial gradient of the phase field variable is obtained in numerical calculations using discrete schemes (such as finite element shape function gradients or finite difference methods).

[0036] The phase-field evolution equation is constructed based on the principles of free energy minimization and maximum dissipation, and is used to describe the dynamic behavior of the phase-field variable d as it evolves over time. Using a viscous regularized form, the phase-field evolution equation is: .in, The mobility coefficient (1 / (Pa·s)) controls the rate of damage evolution, and its value is calibrated through fatigue crack propagation experiments on rail materials. The calibration method specifically involves using a compact tensile specimen made of the same material as the rail, subjecting it to constant amplitude cyclic loading on a fatigue testing machine, and recording the crack propagation rate. By reverse-engineering the Paris formula, the corresponding phase-field model can be obtained. Value, after calibration Values . The historical strain energy driving term is defined as follows: ,in The time integral variable represents the traversal of all historical moments from the initial moment to the current moment t.

[0037] A multi-scale digital twin is obtained by integrating macroscopic and microscopic models through a unidirectional coupling strategy; the displacement field obtained by solving the macroscopic model at each time step. and strain field Driving the evolution of micro-damage models; in the context of orbital health index vectors Potential damage areas identified using methods such as high CRI or DFI are used to extract local strain fields. As boundary conditions for the microscopic damage model, the damage field is obtained by solving the phase field evolution equation. The evolution of the damage field. The system periodically (e.g., daily or after each inspection) based on statistical results of the damage field (e.g., regional average damage). ), and reduce the material parameters of the corresponding region in the macroscopic model (e.g., This allows for feedback of the damage state on the macroscopic mechanical properties. Through the above steps, a multi-scale digital twin that can reflect the dynamic interaction process of track mechanical response and damage accumulation is constructed.

[0038] To maintain synchronization between the digital twin and the real orbital state, ensemble Kalman filtering (EnKF) is used for data assimilation; a multi-scale digital twin is defined at time... state vector ,in, , For all macroscopic finite element nodes in The vertical displacement and rotation at any given moment, For all micro-damage model mesh points in The phase field variable values ​​at time t. Key model parameters to be corrected, such as local elastic modulus E and damping coefficient. Foundation stiffness etc.; Define observation vector ,in, These are the vibration acceleration observations at the locations of each intelligent sensing node. These are the temperature observation values ​​at the locations of each smart sensor node. These are the microscopic strain observations at the locations of each intelligent sensing node. Track health index vector reported by unmanned inspection vehicles .use For each set member (e.g., 100), prediction and updates are performed, analyzing the set from the previous time step. Starting from this point, the prediction set is obtained by forward integration of the model (incorporating process noise). New observations were obtained. (with observation error covariance) After that, perform a Kalman update on each member of the set: ,in For the observation operator, For observation error perturbation, The Kalman gain matrix; analyze the mean of the set. This represents the optimal estimated state of the multi-scale digital twin, incorporating all monitoring data and most closely approximating the actual orbit. (Observation error covariance) Based on the sensor's nominal accuracy settings, for example, the accelerometer observation error is set to 5% of the measured value, the temperature sensor error to ±1°C, and the strain gauge error to ±2°C. The health index error is obtained through model validation. By continuously running EnKF, the multi-scale digital twin achieves dynamic calibration, and its reflected orbital mechanical behavior and damage state remain synchronized with the physical world, providing a reliable initial state and model basis for fault simulation.

[0039] To verify the effectiveness of the multi-scale digital twin, a 200-meter test track was selected at a railway bureau, where intelligent sensor nodes (spaced 25 meters apart) were installed, and an unmanned inspection vehicle was deployed for continuous observation for one month. The initial model parameters were set to design values, and EnKF was run to assimilate the observation data from the first 10 days, resulting in an updated twin. The root mean square error between the model's predicted displacement and the measured displacement over the following 20 days was less than 0.5 mm. The evolution of the phase field variables in the damaged area showed extremely high agreement with the crack locations discovered by subsequent ultrasonic testing (e.g., over 85%), proving that the multi-scale digital twin can accurately reflect the true condition of the track.

[0040] In one embodiment of the present invention, a probabilistic fault evolution map is generated by an inference engine integrating stochastic phase field theory, including the following steps: Based on the updated multi-scale digital twin, the inference engine deployed in the remote data processing center extracts its phase field variables as the initial state for inference; Fault evolution simulation was performed using a stochastic phase-field model via the Monte Carlo method. The damage evolution equation of the stochastic phase-field model includes a deterministic damage evolution term and a random noise term used to quantify the uncertainty of the microstructure of the orbital material. Where d is the phase field variable, The regularization length parameter is used to control the bandwidth of damage localization. M is a parameter related to the fracture energy of a material, and is the mobility coefficient. Let D be the strain energy of the tensile portion, and D be the noise intensity. For random noise, x is the spatial coordinate along the orbital direction, and t is time; Based on the stochastic phase field model, the inference engine executes the Monte Carlo simulation to calculate the damage values ​​at each discrete spatial location of the multi-scale digital twin at different future times, and calculates the probability that the damage values ​​exceed a preset threshold to generate a probabilistic fault evolution map.

[0041] Specifically, the inference engine is a high-performance computing software module deployed in a remote data processing center, consisting of a stochastic phase-field model solver, a Monte Carlo simulation controller, and a result post-processor. Based on the updated multi-scale digital twin, it performs probabilistic predictions of future fault evolution, generating a probabilistic fault evolution map. To quantify the uncertainty of future orbital fault evolution, the inference engine introduces randomness into the digital twin phase-field model, constructing a stochastic phase-field model as its core inference engine. This model directly inherits and extends the deterministic phase-field theoretical framework in the digital twin. When initiating a fault evolution simulation, the inference engine first extracts the current moment from the multi-scale digital twin assimilated and updated using ensemble Kalman filtering (EnKF). The key states are used as the initial conditions for the simulation. Specifically, the extracted content includes the core initial field (phase field variables) that serves as the starting point for the fault evolution simulation. This directly reflects the current spatial distribution of track damage; the input conditions required for the simulation include current track temperature field data and future train load spectra predicted by the train operation history data and planned operation diagrams stored in the remote data processing center; as well as key material parameters obtained from the updated twin, such as the regularized length parameter. Parameters related to material fracture energy Mobility coefficient M, etc.

[0042] Based on the deterministic phase-field model parameters extracted from the digital twin and the current damage state, the inference engine quantifies the impact of uncertainties in the material's microstructure (such as grain boundaries and the randomness of impurity distribution) on damage evolution by introducing random noise terms, thus forming a stochastic phase-field model. The damage evolution equation of this stochastic phase-field model is described as follows: Where d=d(x,t) is the phase field variable defined in the multi-scale digital twin, representing the damage state; The regularization length parameter is used to control the bandwidth of damage localization. The parameter is related to the fracture energy of the material; M is the mobility coefficient, which controls the rate of damage evolution. The stretching component drives the damage evolution; D is the noise intensity, used to quantify the uncertainty of the model and input. Its value is obtained by statistical analysis and calibration of the residual between the historical monitoring data accumulated during the long-term operation of the system (including the track health index vector collected by the unmanned inspection vehicle and the micro-strain data collected by the distributed intelligent sensor network, etc.) and the short-term prediction results made by the updated multi-scale digital twin. To satisfy the conditions of zero mean and covariance being a Dirac function The Gaussian white noise term. The calibration method for noise intensity D is as follows: Collect the track health index vector and corresponding micro-strain data during the initial deployment phase (e.g., the first 3 months). Use a deterministic phase-field model (i.e., the equation without the noise term) to perform short-term predictions (predicting the next 24 hours). Calculate the residuals between the predicted and actual observed values, and use the variance of the residuals as an initial estimate of D. Update the D value quarterly as operational data accumulates. Based on measured data from a certain line, the typical value of D is... to Between. Initial damage field Extracted directly from the updated digital twin; for the finite orbital segments being analyzed, natural boundary conditions (Neumann conditions) are typically set at the spatial boundaries, i.e. , where n is the boundary normal, which means that the damage does not propagate at the boundary.

[0043] The simulation engine employs the Monte Carlo simulation method to simulate fault evolution. The track segment to be simulated along the track direction x is discretized into... There are grid points, located at... Discretize the total time length T of the simulation into... Each time step has a step size of [number]. , at all times Prepare the required Gaussian white noise samples for each independent Monte Carlo simulation. The superscript i indicates the i-th Monte Carlo simulation. In the i-th Monte Carlo simulation, the initial damage field is set. For each time step n=0 to -1, based on the current damage field The tensile strain energy across the entire field is calculated using a coupled macroscopic dynamic model based on the predicted external load at that time step. The discretized stochastic phase field equations are solved using numerical methods to calculate the damage field at the next time step. ,in, Indicates in Point Laplace operator The discrete approximation value; the simulation is saved at all discrete locations. and all points in time damage value The simulation engine runs in parallel and independently on a high-performance computing cluster in a remote data processing center. For example =1000) Monte Carlo simulation; each simulation uses a different random number seed to generate an independent noise sequence. , Thus obtain Different possible future damage evolution paths .

[0044] After completing all After each Monte Carlo simulation, the inference engine performs statistical analysis on the results to generate a probabilistic fault evolution map. For each discrete spatial location... And every future moment Statistically calculate the damage value at that location and time in all Monte Carlo simulations. Exceeding the preset damage threshold Number of times; damage threshold It is a predetermined value between 0 and 1, set comprehensively based on the characteristics of track materials, engineering experience, and historical system operation data, used to define the critical damage state requiring early warning; among which, the damage threshold... The specific basis for this setting is as follows: Three-point bending tests are conducted on the same batch of rail materials, and the load-displacement curves and damage values ​​in the phase field simulation are recorded simultaneously. When macroscopic cracks initiate, the corresponding phase field variable value is approximately 0.75–0.85. Combined with the industry standard's definition of severely damaged rails, the following values ​​are taken: =0.8 is used as the default warning threshold, which can be adjusted according to the line level (e.g., 0.7 for high-speed rail lines, 0.8 for ordinary lines). Calculate the probability of damage. : , where i is the sequence number of the Monte Carlo simulation; For the i-th Monte Carlo simulation, the position is... At any moment The damage value; For indicator functions, when ≥ The probability is 1 if it is true, and 0 otherwise. This represents the probability that track position x will experience damage exceeding a threshold at a future time t, given the initial simulation conditions (current state) and the predicted future train load spectrum. The post-processor of the simulation engine will calculate the... The matrix is ​​visualized to generate a probabilistic fault evolution map; this map is typically a two-dimensional color cloud map with the orbital mileage x as the horizontal axis and future time t as the vertical axis (or vice versa); each point in the map... The shade of color represents the probability of damage at that location at that moment. The size, for example, is a gradient of colors from blue (probability 0) to red (probability 1); the map clearly shows the spatial location of high-failure-risk areas (red areas) and their evolution over time t. The probabilistic map is a direct statistical output of the digital twin's evolutionary capabilities through its inherent stochastic phase field model, reflecting the deep integration of the digital twin with stochastic phase field theory.

[0045] To verify the reliability of the stochastic phase field model and the inference engine, this embodiment provides a simulation verification process: a 2-meter section of a 60kg / m standard rail on a certain railway line is selected, and the initial morphology of the real historical crack (5mm in length and 2mm in depth) collected by an unmanned inspection vehicle is used as the initial damage field. Extract the measured average traffic flow and axle load data for this section over the previous 30 days as the load spectrum input. Set the total number of Monte Carlo simulations. =1000, time step =0.01 seconds. The simulation calculation shows that, after running for an equivalent service period of 120 days, the damage probability is... The damage threshold is exceeded in the crack tip region. The probability of a crack value of 0.8 is approximately 75%. Comparing this with historical records, this section was actually detected with severe cracks meeting the track replacement standard on day 115. Data experiments demonstrate that the above-mentioned extrapolation method based on stochastic phase-field equations and Monte Carlo simulations can accurately predict the evolution trend of macroscopic track cracks, possessing both engineering feasibility and accuracy.

[0046] In one embodiment of the present invention, an immune memory bank is constructed and maintained. Based on the probabilistic fault evolution graph, fault collaborative judgment is performed through a distributed edge computing node network. The immune memory bank is queried using a metric based on hybrid geometric distance to generate a fault repair decision scheme, including the following steps: Based on the probabilistic fault evolution graph, fault collaborative judgment is performed through a distributed edge computing node network deployed in a remote data processing center; the fault collaborative judgment adopts a practical Byzantine fault-tolerant consensus mechanism. An immune memory bank is constructed and maintained to store historical track fault cases and corresponding maintenance plans formed from historical track fault judgments and maintenance records; track fault patterns are matched based on hybrid geometric distance, and the hybrid geometric distance... The calculation formula is: Where P and Q are the probability distributions of the current orbital failure mode and the historical orbital failure mode, respectively; Let KL divergence be the KL divergence. For Wasserstein distance, For Bures distance, and These are the weighting coefficients; Based on the consensus results of the collaborative fault judgment, the immune memory library is queried to perform track fault mode matching, and a track fault maintenance decision scheme is generated by solving a multi-objective optimization problem.

[0047] Specifically, the probabilistic fault evolution map generated by the simulation engine is analyzed to extract features that characterize potential fault regions. This map provides the damage probability of each discrete spatial location x (along the orbital direction) at different future times t. Set a probability threshold for identifying potential fault areas. (For example, a value of 0.5) is possible. This value is based on statistical analysis or pre-set engineering experience of the damage probability distribution in historical track fault case data (historical track fault case data comes from existing track maintenance records and publicly available fault databases imported during system deployment, as well as real case data collected by the system's own sensing module and confirmed through a closed-loop maintenance verification process after system operation). Specifically, the 25th percentile is taken as the damage probability of the actual fault area in 500 historical fault cases. This ensures coverage of the vast majority of potential fault areas while avoiding excessive false alarms. A potential fault area consisting of consecutive high-probability-of-damage points refers to a spatially continuous area that meets the following conditions: The region formed by the points identified; for each such region, extract the feature vector. ,in, The maximum probability of damage to the region at all future predicted times is derived from the maximum value in the probabilistic fault evolution graph. For this region at the current moment The average damage probability is calculated by measuring the average damage probability of all spatial points within the region. time The average of the values ​​is obtained; The spatial variance of the damage probability in this region is calculated by... The spatial distribution variance of the values ​​is obtained; The orbital mileage coordinates of the center point of this area; This is the length of the region along the track direction; The probability of damage to this area exceeds a preset engineering warning threshold for the first time. Prediction time, engineering early warning threshold This is a predetermined value between 0 and 1 (e.g., 0.6 or 0.7), the specific value of which is determined based on industry safety maintenance standards, line operation requirements, and historical decision-making experience. It is used to define the urgency level at which a maintenance decision-making process needs to be initiated; for example, for high-speed railways with speeds exceeding 300 km / h, a value of 0.6 is used. =0.6; For ordinary passenger and freight railways, take =0.7, this value can be optimized and adjusted using historical fault data of the line. Extract the current material parameters (such as local elastic modulus) of this region from the updated multi-scale digital twin. ) and damage field The statistical characteristics are analyzed, and the current ambient temperature is extracted from the real-time data stream of the orbital state sensing module. Recent train arrival time Information such as these forms a complete description of the current failure mode. Track failure modes have high repeatability and patterning characteristics similar to biological "antigens." Considering the relative stability of track materials, load conditions, and environmental factors, similar failures often exhibit similar feature vector distributions. This characteristic provides an engineering basis for the application of biomimetic immune memory mechanisms.

[0048] The remote data processing center deploys a distributed network consisting of N distributed edge computing nodes. Each distributed edge computing node is a high-performance computing unit deployed at key locations along the track (such as stations, substations, or dedicated communication rooms beside the track), possessing independent computing, storage, and communication capabilities. It can process data locally and participate in collaborative decision-making. These nodes are interconnected via dedicated communication links (such as industrial Ethernet or 5G private networks), forming a decentralized distributed network. Each distributed edge computing node runs the same fault diagnosis algorithm. When the system generates a new current fault mode description, it sends it as a pending fault collaborative diagnosis request to all N nodes. The value of N is determined based on the line length, node coverage, and redundancy requirements. For example, if one node is deployed every 10 kilometers, and 5 nodes are deployed on a 50-kilometer section of the line, then N=5. In this case, the tolerable number of faulty nodes f satisfies... That is, f≤1. Each distributed edge computing node Based on the received fault mode description, diagnostic calculations are performed independently by using feature vectors. The fault type (such as surface cracks, internal inclusions, etc.) is determined by comparing and analyzing the fault characteristics with the system's built-in fault feature library (derived from statistical analysis of historical track fault cases); based on the damage probability characteristic value... The system, referencing a pre-defined severity grading standard (based on the correlation between damage probability and actual damage severity in historical track fault case data), classifies severity levels (e.g., mild, moderate, severe). Based on the determined fault type and severity level, it queries a pre-defined maintenance rule base (built from historical successful track fault maintenance records) to obtain corresponding maintenance action suggestions (e.g., grinding, welding, replacement). Distributed edge computing nodes generate a confidence score. ∈[0,1], this score originates from the node's internal calculation process, such as classification probability or matching score, and is used to characterize the degree of confidence in the judgment result. The distributed edge computing node outputs a score containing fault type, severity level, recommended maintenance action type, and confidence score. Preliminary judgment results .

[0049] To ensure reliable consensus even with faulty nodes, a Practical Byzantine Fault Tolerance (PBFT) consensus mechanism is adopted, requiring the total number of distributed edge computing nodes to meet certain conditions. Where f is the maximum number of potentially faulty nodes in the system (including malicious or malfunctioning nodes). The value of f is typically set based on the system size and security requirements; for example, f ≥ 1 and f < 1. The consensus process is as follows: the master distributed edge computing node designated by the system collects the preliminary judgment results reported by each distributed edge computing node and broadcasts a pre-preparation message containing a summary of these results to the entire network; subsequently, each distributed edge computing node verifies the received pre-preparation message, and if the message is valid, it broadcasts a preparation message to other distributed edge computing nodes in the network to indicate agreement; when any distributed edge computing node receives a pre-preparation message from at least After receiving valid preparation messages from at least [number] different distributed edge computing nodes, a submission message is broadcast to the entire network; when the distributed edge computing node receives at least [number] valid preparation messages, it broadcasts a submission message to the entire network. A consensus is considered reached when a valid submission message is received. The judgment results of the distributed edge computing nodes that have reached a consensus are aggregated to generate the final collaborative judgment consensus result. : ,in, It represents the number of distributed edge computing nodes that reach a consensus, and r iterates through all the preliminary judgment results. ,when The indicator function value is 1 when it equals r, and 0 otherwise.

[0050] The system constructs and maintains an immune memory bank to store cases formed from historical track fault judgments and successful repair records. Each historical case's immune memory bank stores a historical track fault case. Includes case ID and historical failure mode description vector. Corresponding successful repair solutions Post-repair verification effect score Among them, the fault mode description vector Analogous to "antigens" in a biological immune system, a successful repair solution This is analogous to a specific "antibody" produced against the antigen; this biomimetic design enables the system to, like the immune system, quickly recall its memory and call upon proven effective repair solutions (antibodies) when encountering similar faults (antigens), thereby skipping complex global optimization calculations and significantly improving decision-making efficiency.

[0051] To accurately measure the similarity between the current fault mode (neoantigen) and historical orbital fault cases (memory antigen), the probabilistic distribution characteristics of fault features must be considered, rather than a simple numerical comparison. This is because orbital damage evolution has inherent randomness, and fault modes exhibit probabilistic distributions rather than deterministic values. Therefore, the feature vector f of the current fault mode and the description vector of historical fault modes should be considered. Convert them to probability distributions respectively and For each feature component k, it is assumed to follow a Gaussian distribution, i.e. ,in, For the current eigenvalue, Its uncertainty variance (estimated based on measurement error); , These are the characteristic statistics of historical track fault cases. The mixed geometric distance between the current fault mode and each historical case is calculated. This distance integrates three different probabilities distribution difference measures: Where P and Q are the probability distributions of the current orbital failure mode and the historical orbital failure mode, respectively; KL divergence measures the information difference between two probability distributions P and Q; The Wasserstein distance measures the minimum cost required to transform distribution P into distribution Q. The Bures distance, based on quantum fidelity, measures the geometric fidelity between non-Gaussian distributions. Defined as: , . and The weighting coefficients are determined through cross-validation optimization using historical matching data, and their typical value range is [insert range here]. ∈[0.3,0.6], ∈[0.2,0.5], and satisfy + <1; where the weighting coefficient is... and The specific optimization process is as follows: Collect 5000 historical fault cases with clear matching results, aiming to maximize the Top-1 matching accuracy, and use a grid search method to... ∈[0,1]、 ∈[0,1] and + Traverse the region where <1, with a step size of 0.05; after cross-validation, when =0.45、 A value of 0.35 yields the highest matching accuracy, therefore it is set as the system default. This hybrid distance metric comprehensively captures the differences between probability distributions across three dimensions: information entropy, spatial structure, and geometric similarity, ensuring matching accuracy. All hybrid geometric distances are calculated. Then, select the historical case with the smallest distance. If the minimum distance Less than the preset matching threshold (This threshold is set by analyzing the distance distribution of historical successful matches), then historical cases... As the best matching case, its repair solution Use it as an initial reference; otherwise, mark the current fault as a new type of fault. The analysis was conducted by examining the distance distribution of historical successful matches. Specifically, the mixed geometric distances of all successful matches were calculated, and the average value was taken. and standard deviation ,set up = +2 This is to cover most of the successful matching cases.

[0052] Based on the consensus result of collaborative judgment and the matched best historical maintenance plan (if the match is successful), a multi-objective optimization problem is constructed to generate the final maintenance plan; the decision variable is the maintenance parameter vector. (Such as grinding depth, welding parameters, material usage, etc.). The optimization objective is to minimize maintenance costs simultaneously. Minimize maintenance operation time And to maximize the expected repair results ,Right now Constraints must be met (Minimum performance requirements) (Maximum permissible working time, usually based on the track maintenance window) , This represents the feasible region for maintenance parameters. Constraints. and Determined based on the route operation requirements; After repair, the probability of damage is reduced by at least half, or the health index is restored to below 0.2; Based on the maintenance window time setting, for example, the maintenance window time for a regular line is 4 hours, then =4 hours. The proposed maintenance parameters will be... Input the updated multi-scale digital twin, simulate maintenance operations (such as modifying the damage field and adjusting material parameters), and then run a short-term fault evolution simulation (e.g., for the next 30 days). Calculate the average reduction in the probability of damage in critical areas after maintenance as... The quantitative value is obtained. The problem is solved using multi-objective optimization algorithms such as NSGA-II, and a set of Pareto optimal solutions (i.e., Pareto fronts) are obtained. The system selects the final solution from the Pareto fronts based on actual operating constraints (such as time urgency and cost constraints), and outputs the fault repair decision scheme.

[0053] When encountering a new type of fault (i.e., a fault case without a matching historical track), the system generates a completely new maintenance plan through the above multi-objective optimization calculation. After successful maintenance verification, the new fault mode and its solution are stored as a new "antigen-antibody" pair in the immune memory bank, realizing the active evolution of system knowledge. This is similar to the mechanism by which the biological immune system produces new antibodies.

[0054] To verify the effectiveness of hybrid geometric distance matching and multi-objective optimization, this embodiment selects 100 cases from the historical fault records of a certain line for testing. If 80 of these cases are used as an immune memory bank and 20 are used as new fault inputs, the system successfully matches 17 cases (accuracy of 85%), and the 3 unmatched cases are correctly marked as new faults, thus proving the reliability of the method.

[0055] In one embodiment of the present invention, the fault repair decision scheme is transformed into realistic visual guidance and robot control commands to drive the mixed reality device and the micro repair robot cluster to perform collaborative repair operations, including the following steps: Based on the fault repair decision scheme and the probabilistic fault evolution map, a realistic visualization guide and control instructions for a cluster of micro-repair robots are generated; the realistic visualization guide is obtained based on a multi-scale digital twin and the fault repair decision scheme. The cluster of miniature repair robots executes repair tasks according to control commands and employs a cooperative strategy. This cooperative control strategy is achieved by solving the following optimization problem: The constraints are satisfied:

[0056] in, The total number of robots in the micro-repair robot swarm. and These are the state vector and control input of the k-th robot, respectively. , The harmonic coefficient, Let the desired trajectory of the k-th robot be... For integration time variable, The value function obtained through collaborative learning, Let K be the state vector of the k-th robot in collaborative learning. For the current moment, To predict the length of the time domain, for The derivative of Let G be the set of neighboring robots of the k-th robot, and let G represent the constraints of the robot cluster. For the index of neighboring robots, Let W be the set of constraints for the control input of the k-th robot, and let W and S be positive definite weighting matrices. The human-computer interaction and autonomous repair execution module is deployed at the track site, driving the mixed reality device to present the real-world visual guidance and driving the cluster of micro-repair robots to perform collaborative repair operations.

[0057] Specifically, based on the fault repair decision-making scheme and the probabilistic fault evolution map, realistic visualization guidance and control commands for the micro-repair robot cluster are generated. The fault repair decision-making scheme specifically includes: fault location information, i.e., the track mileage coordinates of the center point of the fault area. (from feature vectors) The maintenance operation instructions include specific operating steps, process parameters (such as grinding depth, welding current and voltage), and operation sequence, which are mainly formulated based on the system's built-in maintenance rule base and the results of multi-objective optimization. The maintenance rule base is derived from historical successful track fault maintenance records, and pre-sets corresponding maintenance measure frameworks for different fault types and severity levels. Specific process parameters are determined by the maintenance parameter vector generated by the multi-objective optimization process. Repair path planning is based on the length L of the fault area (derived from the feature vector). ) and maintenance operation instructions, planning the spatial trajectory of robots or maintenance personnel entering and leaving the work area, moving, and performing operations; repair priority, based on critical time The urgency of the operation is determined by the time when the probability of damage is expected to exceed the engineering early warning threshold; the resource requirements list determines the materials (such as welding rods, filler material types and quantities), equipment, and personnel configuration required to complete the maintenance operation based on the process parameters, material usage, and operation scope in the maintenance operation instruction; the expected effect indicators are the target values ​​of the track health index (such as the reduction of CRI and DFI) to be achieved after maintenance; the risk control measures are the safety warning areas, protection requirements, and emergency plans during the operation; and the verification and inspection points are the quality inspection items, methods, and standards used for acceptance after the maintenance is completed.

[0058] Using the updated multi-scale digital twin as a base map, information from the fault repair decision-making scheme is overlaid, while the damage probability distribution in the probabilistic fault evolution map is also included. By integrating the heatmap into the base map, an augmented reality scene can be generated that can be displayed on mixed reality devices such as head-mounted displays or tablets, providing on-site maintenance personnel with intuitive guidance for maintenance operations.

[0059] The analysis of the specific repair actions (such as spraying, filling, etc.) and the scope of the fault area to be addressed (defined by L and ...) in the fault repair decision-making plan is crucial. Based on key parameters such as the determination of the number of repair materials and the amount of various repair materials used, a path planning and task allocation algorithm is used to generate control instructions for each robot in the cluster. These instructions include the initial target position planned for the robot, the motion trajectory covering the repair area, and the sequence of tasks arranged in chronological order. The collaborative strategy of the micro-repair robot cluster is implemented using distributed model predictive control. The cluster consists of M robots (the specific value is determined by comprehensively calculating the length L of the area to be repaired, the complexity of the required repair work, the standard operating capability of a single robot, and the total allowable time for the task to ensure efficient task completion; its typical range is 1 to 10 robots). The k-th ( The state vector of the robot (including position, velocity, working device state, etc.) is denoted as... The control input vector (including motor commands, job execution signals, etc.) is denoted as The cluster's cooperative control objective is modeled as an optimization problem: The constraints include: .in, Let S be the desired trajectory of the k-th robot, whose data comes from the micro-repair robot cluster control commands generated for this robot. W and S are positive definite weighted matrices, usually taken as diagonal matrices; W is used to penalize trajectory tracking error, i.e., the actual state of the robot. With the expected state trajectory The increased weighting coefficient of the deviation indicates a higher requirement for trajectory tracking accuracy; S is used to penalize the control input. The increased weighting coefficient of W and S indicates a stricter restriction on energy consumption control. The specific values ​​of the weighting coefficients of W and S are determined during system deployment through experimental calibration, simulation optimization, or based on engineering experience. The specific method for determining the weighting coefficients of W and S is as follows: set up a typical repair scenario (such as crack filling with L=2m) in the simulation environment, and use the root mean square of tracking error and control energy consumption as performance indicators. Use the grid search method to adjust the diagonal elements of W and S, and select the value that optimizes the comprehensive performance index. After calibration, for the standard operation scenario, W is set to diag(10,10,1) (position weight 10, speed weight 1), and S is set to diag(0.1,0.1) (control quantity weight 0.1). The harmonic coefficient is used to balance the weights among the three sub-objectives: trajectory tracking error, control energy consumption, and long-term repair benefits. Its value is calibrated through offline simulation and field testing, with a typical range of values. ; The value function, trained through collaborative reinforcement learning (using algorithms such as multi-agent deep deterministic policy gradient for offline training), is used to evaluate the robot in extended states. (Including its own state) The long-term repair benefits that can be obtained by performing actions under the premise of environmental information and neighbor status are given by the following inputs: The output is a scalar value assessment; The current moment; For integration time; The predicted time domain length (in seconds) is set based on real-time control requirements and the computing power of distributed edge computing nodes; a typical value is 5. 30 seconds; For state differentiation; The mathematical model (usually a set of nonlinear ordinary differential equations) describing the dynamics of the k-th robot individual is obtained through the robot's factory dynamics calibration. The coupling term model describes the interactions between robots (such as collision avoidance, communication, and task coordination), and its form is defined according to specific cooperation rules (such as artificial potential field method and consensus protocol). The set of neighbors of the k-th robot is dynamically determined by the real-time communication distance and perception range between robots. The constraints are cluster-level, including total operating area limits, total repair material limits, minimum safe distance between robots, etc. The specific inequalities are derived from the resource list and safety specifications in the fault repair decision-making scheme. The feasible region for the k-th robot control input is determined by the robot motor performance, the physical limits of the working device, etc., and is usually expressed as the upper and lower limits of the control quantity.

[0060] This optimization problem is solved in real time using a distributed model predictive control algorithm on distributed edge computing nodes (which can be deployed on unmanned inspection vehicles or fixed nodes) at the track site. The solution process is as follows: within each control cycle, each robot k obtains its current state. and through communication from the neighbor set Receive its prediction information (i.e., its state prediction sequence in the future prediction time domain); based on the prediction information and its own desired trajectory Each robot constructs and solves a local finite-time optimal control problem (i.e., a subset of the objective function and constraints in its own dimension). All robots solve their respective local optimization problems in parallel (usually transformed into quadratic programming or nonlinear programming problems, solved using numerical optimization methods such as interior-point methods or sequential quadratic programming), obtaining a set of predicted control input sequences and state sequences for their own future time domain. Each robot exchanges the parts of its predicted sequence that are coupled with those of its neighboring robots; each robot updates the coupling constraints in its local problem based on the new information received and solves it again; this process is iterated until a preset convergence condition is met (such as reaching the maximum number of iterations or the change in the predicted sequence being less than the tolerance), which is considered as achieving distributed consensus; each robot uses the first control variable in its finally solved optimal control input sequence. ,Will The commands are converted into specific execution instructions for the drive motor and working device (such as spray valve and filling pump), and then sent to the corresponding robot for execution via a wireless communication network.

[0061] Mixed reality devices receive real-world visual guidance and project it onto the maintenance personnel's display device in augmented reality form to assist them in performing maintenance tasks; each robot executes repair tasks (such as crack filling and coating spraying) according to the received control instructions; during the operation, the robot uses sensors to provide real-time feedback on the operation status to the edge nodes.

[0062] In one embodiment of the present invention, the process of collecting repair operation results and orbital status verification data, updating the multi-scale digital twin and optimizing the immune memory bank through federated learning includes the following steps: The system collects the repair operation results of the human-computer interaction and autonomous repair execution module, as well as the track status verification data of the track status perception and data acquisition module. The track status verification data includes vibration, temperature and micro-strain data collected by the distributed intelligent sensor network, and track health index vector collected and generated by the unmanned inspection vehicle. Based on orbital state verification data, a federated learning framework is used to update multi-scale digital twins deployed in remote data processing centers; Based on the repair operation results, new track fault cases are generated, and the immune memory bank is optimized using an incremental learning mechanism based on case reasoning. When the difference between the new track fault case and the historical track fault cases in the immune memory bank exceeds a preset novelty threshold, the new track fault case is added to the immune memory bank.

[0063] Specifically, it collects repair operation result data from the human-computer interaction and autonomous repair execution modules, including: the actual operation trajectory of each micro-repair robot. Actual control input The data includes: material consumption, start and end times of the operation, images of the repaired site, solution logs of the distributed edge computing nodes during the operation for the optimization problem described in the human-computer interaction and autonomous repair execution module (including objective function values ​​and constraint satisfaction status), and a preliminary repair quality assessment (e.g., "good" or "requires re-inspection") submitted by on-site operators via mixed reality devices at the end of the operation. Simultaneously, track condition verification data from the track condition perception and data acquisition module is collected to objectively evaluate the repair effect, including: vibration acceleration collected by the distributed intelligent sensor network after the repair operation is completed and a stabilization period (e.g., after several train passes). ,temperature Microscopic strain data And the track health index vector, which is collected and generated again by the unmanned inspection vehicle under the next preset inspection cycle or special review instruction. However, rail transit systems inherently suffer from data fragmentation and privacy barriers. Track data from different lines and sections often form "data silos" due to management ownership, trade secrets, or security considerations, making it impossible to directly aggregate them to the central server. At the same time, the massive amounts of raw monitoring data (especially high-definition images and continuous vibration waveforms) put enormous pressure on communication bandwidth, making full uploading neither economical nor practical. To address this, this system adopts a federated learning framework to collaboratively update multi-scale digital twin models without sharing raw data.

[0064] Based on the collected orbital state verification data, a federated learning framework is used to update the multi-scale digital twin deployed at a remote data processing center (which serves as the central coordination node, also known as the server, in this federated learning framework). The remote data processing center, acting as the server of the federated learning framework, updates the global model parameters of the current multi-scale digital twin. The data is distributed to each client. Each client is a computing unit deployed on distributed edge computing nodes in each unmanned inspection vehicle and key areas. Each client stores a local copy of the multi-scale digital twin model (hereinafter referred to as the local model copy) for training based on local data. Each client only uses its locally stored track state verification data for training; the original data does not need to leave the local machine, thus protecting data privacy and reducing communication overhead. Each client c uses its locally stored track state verification data (mainly including the unmanned inspection vehicle health index vectors before and after repair). , The training is performed locally using data from intelligent sensor nodes within the corresponding time period, with the training objective being to minimize the local loss function. An example of this loss function in the form of mean squared error: ;in, The parameter is Multi-scale digital twin model; parameters This is a set that encompasses all variables in the model that can be adjusted through data learning, primarily including material property parameters in the macroscopic dynamic model (such as elastic modulus E, density, etc.). Phase field parameters in the micro-damage model (such as critical energy release rate) Regularization length ), and the state variables updated during data assimilation; and These are the spatial coordinates (mileage along the track) and timestamp of the i-th data sample, respectively; It corresponds to the actual observed value, for example, it could be A specific index, or data from smart sensing (such as microscopic strain) The mechanical state quantities derived from ) This represents the number of local data samples on client c.

[0065] The client uses optimization algorithms such as stochastic gradient descent to calculate the gradient of the model parameters based on local data, and then updates the local model parameters based on this gradient. Specifically, taking stochastic gradient descent (SGD) as an example, parameter updates follow rules. ,in For local learning rate, Let be the gradient of the loss function with respect to the parameters; where the local learning rate is... The value was determined through cross-validation. During the initial system deployment, historical data from one client was selected for testing. The convergence rate and final loss at 0.001, 0.01, and 0.1 are compared, and the value that minimizes the loss without oscillation is chosen; typically, a value of 0.1 is used. =0.01. Each client will encrypt the result. The data is uploaded to the server (i.e., the remote data processing center). The server uses a secure aggregation algorithm (such as Secure Aggregation) to decrypt and aggregate updates from all clients, and then updates the global model parameters using a federated averaging algorithm. ,in, C is the global learning rate; C is the total number of clients participating in this round of updates. The federal average algorithm is explained in detail below: , This represents the total number of data samples from all participating clients. The global learning rate is included. Typically, it's set to 1.0, indicating that the average update after aggregation is applied directly; if a smoother convergence is desired, it can be reduced appropriately, such as... =0.8; After each server update, the loss on the validation set is calculated. When the loss decreases less than a threshold for 5 consecutive rounds (e.g., ...), the loss is calculated. Training stops when the updated global parameters are displayed. The parameters of the original multi-scale digital twin model are replaced, so that the multi-scale digital twin can more accurately reflect the real mechanical behavior and long-term evolution trend of the track structure after repair.

[0066] The federated learning mechanism enables the system to achieve global knowledge fusion while protecting data privacy and reducing communication burden. Based on the results of the repair operation, an incremental learning mechanism based on case reasoning is used to optimize the immune memory. For each complete "detection-diagnosis-repair-verification" closed loop, the system generates a new track fault case. ; The feature vectors are extracted from this fault event, including but not limited to: the orbital health index vector at the time of the fault. Key features from probabilistic fault evolution maps (such as maximum damage probability) The damaged area along the track (length L), the fault type and severity level labels determined by collaborative judgment, and the final fault repair decision plan. ;also The verification results after the associated storage repair, i.e., the evaluation of the repair effect reflected in the orbital status verification data. Its quantitative indicators are derived from orbital state verification data, such as Each indicator relative to The extent of improvement.

[0067] New Case of Computation With all historical cases in the immune memory bank Mixed geometric distance between The formula for calculating this distance is: Among them, probability distribution and It is obtained by kernel density estimation from their respective eigenvectors; and These are preset weighting coefficients, consistent with the system's initial settings. Let them be... The historical case with the smallest distance is The corresponding distance is The system will With a preset novelty threshold Comparison; the novelty threshold Based on the statistical characteristics (such as mean) of the set of mixed geometric distances between all historical cases in the immune memory bank. with standard deviation Configure it as follows: A typical configuration formula is... ,in An adjustable coefficient set according to the system's novelty sensitivity requirements; The value is determined by the system's trade-off between false positives (misclassifying novel cases as similar) and false negatives (misclassifying similar cases as novel), and is plotted through cross-validation of historical cases. The accuracy curve below is selected to maximize the F1 score. Value, usually taken =2.0 (e.g., corresponding to approximately a 95% confidence interval); this threshold The system can be periodically updated based on newly added case data after it is running. Then it is considered a new case Compared with existing cases If the cases are highly similar, the system will not add new cases, but will instead use historical cases. Corresponding repair solutions Perform a weighted update and incorporate the successful solution from this study. Experience; for example, can Update to a weighted average of the two or iteratively update using a small learning rate. If Then it is believed This is a novel case; the system will ( , This is added as a new antigen-antibody pair to the immune memory bank, thereby expanding the system's ability to cope with novel faults. To prevent the immune memory bank from expanding indefinitely, the system's maximum capacity can be set according to actual needs. Maximum capacity Configure according to storage resources and the diverse needs of different cases, for example, take =10000. When the maximum capacity is exceeded, some old cases are removed based on utility metrics such as call frequency and timestamps, using strategies similar to Least Recently Used (LRU) or utility eviction. For example, an Least Recently Used (LRU) strategy or a utility eviction strategy (such as overall score = call frequency × time decay factor) can be used to remove old cases, specifically setting a utility score. ,in For call frequency, For the current time, Last access time The decay coefficient is set to 0.01 / day, and the lowest 10% of cases are removed each time. A learning and optimization process is automatically triggered after each complete remediation job is executed and its effectiveness verified. A fixed time period (e.g., weekly or monthly) is set to aggregate data from all remediation events within that period, and batch-process federated learning updates and immune memory bank optimizations are performed.

[0068] The federated learning framework and the incremental learning mechanism together constitute the system's continuous evolution engine. The former solves the problem of global model updates under data dispersion and privacy constraints, enabling the digital twin to securely integrate diverse track states and maintenance experience across the entire network without aggregating the original data, thereby continuously improving its generalization ability and prediction accuracy. The latter ensures that the maintenance decision knowledge base can be continuously enriched with the accumulation of practice. The combination of the two enables the system to truly realize an intelligent closed loop of learning in use and improving in evolution.

[0069] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A big data driven based remote track failure intelligent detection system, characterized in that, Includes a remote data processing center and the following modules: Track status perception and data acquisition module: Through a distributed intelligent sensor network fixedly deployed on the track, and an unmanned inspection vehicle that can move along the track and is equipped with a multi-sensor fusion detection unit, it collaboratively collects multi-source heterogeneous data of the track. Fault Evolution Inference and Digital Twin Module: Constructs and updates multi-scale digital twins based on multi-source heterogeneous orbital data, and generates probabilistic fault evolution maps through an inference engine integrating stochastic phase field theory; Intelligent collaborative judgment and decision generation module: Constructs and maintains an immune memory library, performs collaborative fault judgment through a distributed edge computing node network based on the probabilistic fault evolution map, and queries the immune memory library using a metric based on hybrid geometric distance to generate fault repair decision schemes; including: Based on the probabilistic fault evolution graph, fault collaborative judgment is performed through a distributed edge computing node network deployed in a remote data processing center; the fault collaborative judgment adopts a practical Byzantine fault-tolerant consensus mechanism. An immune memory base is constructed and maintained to store historical track fault cases and corresponding maintenance solutions formed by historical track fault judgment and maintenance records; a track fault mode is matched based on a mixed geometric distance The calculation formula of the mixed geometric distance is as follows: where P and Q are probability distributions from the current track failure mode and the historical track failure mode, respectively; for the KL divergence, for the Wasserstein distance, for the Bures distance, and is a weight coefficient; Based on the consensus results of the collaborative fault judgment, the immune memory bank is queried to perform track fault mode matching, and a track fault maintenance decision scheme is generated by solving a multi-objective optimization problem. Human-computer interaction and autonomous repair execution module: Transforms the fault repair decision-making scheme into realistic visual guidance and robot control commands, driving the mixed reality equipment and the cluster of micro repair robots to perform collaborative repair operations; The system's autonomous learning and optimization evolution module collects repair operation results and orbital status verification data, updates the multi-scale digital twin through federated learning, and optimizes the immune memory bank.

2. The remote track failure intelligent detection system based on big data driving according to claim 1, wherein, Collecting heterogeneous data from multiple sources on the track through a distributed intelligent sensor network that is fixedly deployed on the track includes the following steps: The distributed intelligent sensing network consists of multiple intelligent sensing nodes fixedly installed on the track surface; each intelligent sensing node integrates a vibration sensor, a temperature sensor and a strain sensor, which are used to continuously collect vibration, temperature and micro-strain data of the track body, and upload the data to a remote data processing center after preprocessing. The intelligent sensing node determines whether a train is approaching based on the energy change and rate of change of the track vibration signal in the characteristic frequency band when the train passes. If the train is approaching, the current intelligent sensing node is automatically awakened. The awakened intelligent sensing node then wakes up the downstream intelligent sensing nodes in sequence along the track direction through surface acoustic wave communication.

3. The remote track failure intelligent detection system based on big data driving according to claim 1, characterized in that, Collecting multi-source heterogeneous data about the track using an unmanned inspection vehicle equipped with a multi-sensor fusion detection unit that can move along the track, including the following steps: The multi-sensor fusion detection unit carried by the unmanned inspection vehicle includes an image sensor, a laser sensor, an ultrasonic sensor, and an infrared sensor. The unmanned inspection vehicle performs inspection tasks according to a preset cycle, and autonomously cruises along the track and simultaneously collects track surface images, geometric deformation data of track profile and gauge, ultrasonic detection data reflecting internal defects, and temperature field distribution data of track surface according to the scheduling instructions of the remote data processing center. The unmanned inspection vehicle has a built-in edge computing unit and uses a multimodal data fusion algorithm to perform real-time preprocessing and feature extraction on the collected image, geometric, ultrasonic and thermal data to generate a track health index vector. The track health index vector includes a crack risk index, a geometric distortion index, an internal defect index and a thermal anomaly index. The unmanned inspection vehicle uploads the track health index vector to a remote data processing center through a vehicle-to-ground communication link.

4. The remote intelligent track fault detection system based on big data as described in claim 1, characterized in that, Constructing and updating multi-scale digital twins based on multi-source heterogeneous orbital data includes the following steps: Construct a multi-scale digital twin model framework that includes macroscopic, mesoscopic, and microscopic scales; A macroscopic dynamic model of the track structure is constructed, which describes the overall mechanical behavior of the track structure based on beam theory. A microscopic damage model for the evolution of track damage is constructed. This microscopic damage model is described based on phase-field theory, and a phase-field variable characterizing the structural damage state is introduced. By integrating macroscopic dynamics models, mesoscopic models, and microscopic damage models, a multi-scale digital twin was constructed and deployed in a remote data processing center. Among them, the mesoscopic model serves as a bridge between the macroscopic and microscopic scales, used to describe the influence of the microstructure of the track material on the macroscopic mechanical properties and the initial initiation of damage. Based on the vibration, temperature, and micro-strain data collected by the distributed intelligent sensor network uploaded by the intelligent sensor nodes, and the track health index vector collected and generated by the unmanned inspection vehicle, the multi-scale digital twin is updated using a data assimilation method to correct its key internal states that reflect the mechanical behavior and damage state of the track structure.

5. The remote intelligent track fault detection system based on big data as described in claim 1, characterized in that, The probabilistic fault evolution map is generated using a derivation engine that integrates stochastic phase field theory, including the following steps: Based on the updated multi-scale digital twin, the inference engine deployed in the remote data processing center extracts its phase field variables as the initial state for inference; Fault evolution simulation was performed using a stochastic phase-field model via the Monte Carlo method. The damage evolution equation of the stochastic phase-field model includes a deterministic damage evolution term and a random noise term used to quantify the uncertainty of the microstructure of the orbital material. Where d is the phase field variable, The regularization length parameter is used to control the bandwidth of damage localization. M is a parameter related to the fracture energy of a material, and is the mobility coefficient. Let D be the strain energy of the tensile portion, and D be the noise intensity. For random noise, x is the spatial coordinate along the orbital direction, and t is time; Based on the stochastic phase field model, the inference engine executes the Monte Carlo simulation to calculate the damage values ​​at each discrete spatial location of the multi-scale digital twin at different future times, and calculates the probability that the damage values ​​exceed a preset threshold to generate a probabilistic fault evolution map.

6. The remote intelligent track fault detection system based on big data as described in claim 1, characterized in that, The fault repair decision-making scheme is transformed into realistic visual guidance and robot control commands, driving the mixed reality device and the cluster of micro repair robots to perform collaborative repair operations, including the following steps: Based on the fault repair decision scheme and the probabilistic fault evolution map, a realistic visualization guide and control instructions for a cluster of micro-repair robots are generated; the realistic visualization guide is obtained based on a multi-scale digital twin and the fault repair decision scheme. The cluster of miniature repair robots executes repair tasks according to control commands and employs a cooperative strategy. This cooperative control strategy is achieved by solving the following optimization problem: The constraints are satisfied: in, The total number of robots in the micro-repair robot swarm. and These are the state vector and control input of the k-th robot, respectively. , The harmonic coefficient, Let the desired trajectory of the k-th robot be... For integration time variable, The value function obtained through collaborative learning, Let K be the state vector of the k-th robot in collaborative learning. For the current moment, To predict the length of the time domain, for The derivative of Let G be the set of neighboring robots of the k-th robot, and let G represent the constraints of the robot cluster. For the index of neighboring robots, Let W be the set of constraints for the control input of the k-th robot, and let W and S be positive definite weighting matrices. The human-computer interaction and autonomous repair execution module is deployed at the track site, driving the mixed reality device to present the real-world visual guidance and driving the cluster of micro-repair robots to perform collaborative repair operations.

7. The remote intelligent track fault detection system based on big data as described in claim 1, characterized in that, The process of collecting repair operation results and orbital status verification data, updating the multi-scale digital twin and optimizing the immune memory bank through federated learning includes the following steps: The system collects the repair operation results of the human-computer interaction and autonomous repair execution module, as well as the track status verification data of the track status perception and data acquisition module. The track status verification data includes vibration, temperature and micro-strain data collected by the distributed intelligent sensor network, and track health index vector collected and generated by the unmanned inspection vehicle. Based on orbital state verification data, a federated learning framework is used to update multi-scale digital twins deployed in remote data processing centers; Based on the repair operation results, new track fault cases are generated, and the immune memory bank is optimized using an incremental learning mechanism based on case reasoning. When the difference between the new track fault case and the historical track fault cases in the immune memory bank exceeds a preset novelty threshold, the new track fault case is added to the immune memory bank.

Images