Intelligent operation and maintenance system for high-speed rail train

Abstract

A system for intelligent operation and maintenance of a high-speed rail train relates to the technical field of operation and maintenance of the high-speed rail train and comprises an Internet of Things sensing layer, a network transmission layer, an intelligent decision-making layer and an unmanned execution layer, and the Internet of Things sensing layer is used for detecting to-be-detected parts of the train and a rail and generating multi-dimensional data with space-time labels; the network transmission layer is used for real-time data transmission and remote control among the Internet of Things sensing layer, the intelligent decision-making layer and the unmanned execution layer; the intelligent decision-making layer comprises a multi-source data fusion module and an intelligent operation and maintenance knowledge graph module and is used for performing fault diagnosis, life prediction and automatic generation of a maintenance work order on the to-be-tested component; and the unmanned execution layer receives the standard chemical order and the control instruction generated by the intelligent decision-making layer, and executes maintenance operation according to the standard chemical order and the control instruction. The intelligent operation and maintenance system for the full-link closed-loop high-speed rail train is realized by arranging the Internet of Things sensing layer, the network transmission layer, the intelligent decision-making layer, the unmanned execution layer and the safety protection layer.

Description

Technical Field

[0001] This invention relates to the field of high-speed train operation and maintenance technology, and in particular to a system for intelligent operation and maintenance of high-speed trains. Background Technology

[0002] China's high-speed railway network has formed an "eight horizontal and eight vertical" pattern. With the continuous expansion of operating mileage and equipment scale, the traditional maintenance mode, which is mainly based on manual labor, faces multiple challenges. The existing high-speed rail operation and maintenance system has problems such as limited detection capabilities, data silos, extensive resource scheduling, insufficient communication reliability, weak two-way interaction capabilities, and network security risks.

[0003] 1. Limited detection capabilities: Existing sensors have insufficient coverage, with a detection coverage rate of only 85% for key components. The missed detection rate for manual visual inspection is as high as 0.3%, and it is unable to capture millimeter-level deformation.

[0004] 2. Data silo problem: Data from vehicle-mounted TCMS, trackside monitoring equipment, and maintenance work order system are not effectively integrated, resulting in low accuracy of fault correlation diagnosis and high rate of repeated maintenance.

[0005] 3. Inefficient resource allocation: Maintenance plans are made based on manual experience, resulting in high empty running rates, long response times for adjusting train schedules during peak periods, difficulty in adapting to dynamic passenger flow demands, high maintenance costs, and an annual increase in maintenance frequency as equipment service life increases.

[0006] 4. Insufficient communication reliability: Traditional vehicle-to-ground communication relies on wired connections or single-mode wireless transmission. Signal interruptions are prone to occur in tunnels and areas with dense train traffic, resulting in high data transmission delays and affecting real-time diagnostic efficiency.

[0007] 5. Weak two-way interaction capability: The ground system can only receive basic status data from the vehicle-mounted TCMS and cannot issue dynamic maintenance commands or remotely configure sensor parameters in real time, which restricts the closed loop of operation and maintenance automation.

[0008] 6. Network security risks: The lack of encrypted transmission and access control for vehicle-to-ground data poses a risk of malicious attacks causing equipment malfunctions. Existing protection measures only meet the requirements of Level 2 Information Security Protection. Summary of the Invention

[0009] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a system for intelligent operation and maintenance of high-speed trains, which realizes a closed-loop intelligent operation and maintenance system for high-speed trains by setting up an Internet of Things sensing layer, a network transmission layer, an intelligent decision-making layer, an unmanned execution layer, and a security protection layer.

[0010] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0011] A system for intelligent operation and maintenance of high-speed trains includes an IoT sensing layer, a network transmission layer, an intelligent decision-making layer, and an unmanned execution layer. The IoT sensing layer includes sensor devices and a fusion positioning system for detecting the parts of the train and track to be tested. The fusion positioning system provides spatiotemporal tags for the sensor device data, generates multidimensional data with spatiotemporal tags, and sends it to the information processing device.

[0012] The network transmission layer is used for real-time data transmission and remote control between the IoT sensing layer, intelligent decision-making layer, and unmanned execution layer; the network transmission layer includes wired network and wireless network, used to construct a hybrid wired ring network and wireless mesh network;

[0013] The intelligent decision-making layer includes a multi-source data fusion module and an intelligent operation and maintenance knowledge graph module. These modules are used for fault diagnosis, lifespan prediction, and automated generation of maintenance work orders for the components under test. The multi-source data fusion module employs an LSTM time series analysis model and a Transformer spatiotemporal attention model. The LSTM model captures the time series features of multi-dimensional data, while the Transformer model calculates the association weights of vibration-temperature-wear data and outputs the fault type. The intelligent operation and maintenance knowledge graph module includes a Neo4j graph database, which stores the "component-fault-rule-tool" relationship. The Drools rule engine calls the graph in real time, generates standardized work orders based on the fault type, and sends them to the unmanned execution layer.

[0014] The unmanned execution layer receives standardized work orders and control instructions generated by the intelligent decision-making layer. It then executes maintenance operations through the execution equipment based on the standardized work orders and control instructions, and feeds back the operation data to the intelligent decision-making layer through the network transmission layer to update the digital twin model.

[0015] Furthermore, the multidimensional data obtained through sensor devices and fusion positioning systems is filtered and preprocessed by the on-board host and / or trackside edge nodes, and then uploaded to the central control host through the network transmission layer.

[0016] Furthermore, the sensor equipment includes a vibration acceleration sensor for detecting the vibration acceleration of the wheelset bearing housing in the running gear; an infrared thermal imaging array for detecting the temperature rise of the wheelset bearing at the bottom of the train; an ultrasonic phased array probe for detecting cracks in the wheelset tread; a terahertz imager for detecting defects in the car body welds; and a 32-line lidar for detecting the wheelset profile, tread wear, and track geometry.

[0017] Furthermore, the fusion positioning system includes a vehicle-mounted BeiDou-UWB terminal integrating a BeiDou module and a UWB tag, which transmits location information to the vehicle-mounted host via an SPI bus; the trackside UWB base station is connected to the trackside switch via Ethernet, and the positioning data is synchronized to the central control host via a fiber optic ring network.

[0018] Furthermore, the wired network of the network transmission layer adopts a 10 Gigabit fiber optic ring network as its backbone, and the on-board host is connected to the trackside switch through a through fiber optic cable at the rear of the vehicle; the wireless network deploys high-gain directional antennas through trackside wireless access points (APs), the maintenance depot uses omnidirectional antennas, and the on-board wireless unit supports inter-area handover.

[0019] Furthermore, the intelligent operation and maintenance knowledge graph module uses a path optimization algorithm to allocate execution devices for the unmanned execution layer.

[0020] Furthermore, the multi-source data fusion module of the intelligent decision-making layer forms a federated learning group across multiple train depots. Local data from each depot is used to train the model without sharing the original data, and the global model parameters are aggregated and updated monthly.

[0021] Furthermore, standardized work orders include maintenance procedures, tool lists, and spare parts information.

[0022] Furthermore, it also includes a safety protection layer, which comprises an intelligent interlocking and anti-collision system, network security equipment, and a communication redundancy module, used to achieve safety protection and network security assurance in the maintenance area; the intelligent interlocking and anti-collision system deploys LiDAR to construct a dynamic electronic fence, the network security equipment deploys an industrial firewall at the control center exit, supports national cryptographic encryption tunnels, and performs two-way authentication of vehicle-to-ground data transmission, the communication redundancy module reports signal strength from the vehicle-mounted wireless unit to the safety protection host, triggering a dual-network switching command; the vehicle-mounted host has a built-in SSD that automatically caches data when communication is interrupted and retransmits it after the signal is restored.

[0023] Furthermore, the execution equipment includes intelligent robotic arms, AGV robots, tracked inspection robots, AR glasses guidance, and intelligent torque wrenches.

[0024] The beneficial effects of this invention are:

[0025] 1. The sensors of the intelligent operation and maintenance system for high-speed trains of the present invention cover all key components, achieving millimeter-level detection (such as 0.2mm crack recognition), and the missed detection rate of the system of the present invention is lower than that of traditional manual inspection.

[0026] 2. The present invention has high efficiency in data fusion: multi-dimensional data (vibration, temperature, image) are fused through LSTM+Transformer model, which greatly improves the accuracy of fault diagnosis, solves the "data island" problem of traditional systems, and reduces the rate of repeated maintenance.

[0027] 3. This invention realizes a closed-loop automation of operation and maintenance: it achieves unmanned operation of the entire process of "perception-decision-execution", improves the efficiency of component replacement, reduces the empty running rate, and reduces the average annual increase rate of maintenance costs as the service life of the equipment increases.

[0028] 4. Improved communication reliability of the invention: The end-to-end latency of the "wired ring network + wireless mesh" hybrid network is ≤100ms, and the signal availability rate is 99.99%, solving the problem of signal interruption in tunnels and dense areas.

[0029] 5. Enhanced security protection of this invention: national cryptographic encryption transmission, Level 3 information security protection, and laser radar anti-collision response of <0.2 seconds, which is more secure than traditional Level 2 information security protection.

[0030] 6. This invention features intelligent resource scheduling: It automatically generates work orders based on knowledge graphs and optimization algorithms, optimizes AGV paths, adapts to dynamic passenger flow demands, and shortens the response time for peak-hour train schedule adjustments. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the system structure for intelligent operation and maintenance of high-speed trains according to the present invention;

[0032] Figure 2 Block diagram of the IoT sensing layer subsystem;

[0033] Figure 3 This is a block diagram of the network transport layer subsystem.

[0034] Figure 4 Block diagram of the intelligent decision-making layer subsystem;

[0035] Figure 5 Block diagram of the unmanned execution layer subsystem;

[0036] Figure 6 Block diagram of the security protection layer subsystem;

[0037] Figure 7 A detailed block diagram of the wheelset detection scenario in the IoT sensing layer;

[0038] Figure 8 This is a flowchart of the closed-loop interaction process of the system.

[0039] Figure 9 This is a flowchart of the data processing for the intelligent decision-making layer. Detailed Implementation

[0040] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] This invention provides a system for intelligent operation and maintenance of high-speed trains, such as... Figure 1As shown, it includes an IoT sensing layer, a network transmission layer, an intelligent decision-making layer, and an unmanned execution layer. The IoT sensing layer includes sensor devices and a fusion positioning system for detecting the parts of the train and track to be tested. The fusion positioning system provides spatiotemporal labels for the sensor device data, generates multidimensional data with spatiotemporal labels, and sends it to the information processing device.

[0042] Specifically, the IoT sensing layer uses a global sensor cluster and positioning system to achieve millimeter-level precision detection of the status of key train components, generating multi-dimensional data with spatiotemporal tags, which is then transmitted to the intelligent decision-making layer through the network transmission layer.

[0043] like Figure 2 As shown, the sensor equipment includes a vibration acceleration sensor for detecting the vibration acceleration of the wheelset bearing housing in the running gear; an infrared thermal imaging array for detecting the temperature rise of the wheelset bearing at the bottom of the train; an ultrasonic phased array probe for detecting cracks in the wheelset tread; a terahertz imager for detecting defects in the car body welds; and a 32-line lidar for detecting the wheelset profile, tread wear, and track geometry.

[0044] More specifically, the vibration acceleration sensor can be the Kistler 8702B50 from Switzerland, installed in the wheelset bearing housing of the running gear, fixed by a magnetic clamp, and connected to the onboard edge computing host (hereinafter referred to as "onboard host") via a 485 bus. It monitors the vibration acceleration of the wheelset bearing housing in real time, collecting vibration signals from 10-5000Hz. Abnormal bearing housing vibration usually indicates bearing wear, loosening, or fatigue damage, which may lead to wheelset imbalance, abnormal noise, or even derailment. By collecting vibration signals from 10-5000Hz, early bearing fault characteristics (such as high-frequency impact signals) can be captured. Combined with intelligent decision-making layer algorithm analysis, early warnings of bearing failure risks can be issued, ensuring the safe operation of the train running gear.

[0045] The infrared thermal imaging array, such as the FLIR A655sc, is integrated into a detection bracket at the bottom of the train, 0.5m vertically from the bearing. It transmits thermal images to the onboard host via gigabit Ethernet, enabling real-time monitoring of the temperature rise of the train's wheel bearings (such as axle box bearings and traction motor bearings). Excessive bearing temperature rise is a direct manifestation of poor lubrication, abnormal load, or increased friction, potentially leading to serious malfunctions such as bearing seizure and axle overheating. By scanning thermal images in real-time with an infrared thermal imaging array 0.5m from the bearing, abnormal temperature rise areas can be quickly located. Combined with temperature thresholds, the bearing's operating status can be determined, preventing traffic safety accidents caused by bearing overheating.

[0046] The ultrasonic phased array probe, such as the Olympus OmniScan MX2, is deployed in the trackside inspection pit, directly facing the wheelset tread. It connects to the trackside edge node via a waterproof cable to detect cracks as small as 0.2mm in the wheelset tread. The wheelset tread is a critical component in direct contact with the track; if cracks are not detected promptly, they can propagate during high-speed train operation, leading to serious accidents such as wheelset breakage and derailment. The ultrasonic phased array probe emits ultrasonic waves directly onto the tread through the trackside inspection pit, using the reflected signals to identify minute cracks, ensuring the integrity of the wheelset structure and guaranteeing its strength and stability during operation.

[0047] The terahertz imager, such as the Advantest TAS7200, can be installed on a trackside gantry, 1.2m from the track center and 0.8m directly above the track. It connects to a trackside switch via fiber optic cable, scans the car body welds, identifies weld defects smaller than 0.2mm, and triggers synchronously with the lidar. The terahertz image is transmitted to the trackside edge node via a dedicated fiber optic link, spatiotemporally aligned with the lidar point cloud data. Car body welds are critical nodes connecting various car body components. Weld defects (such as porosity, incomplete penetration, and cracks) can weaken the car body structure and cause cracking under vibration and load, affecting the overall safety of the train. The terahertz imager identifies minute defects through penetrating scanning, ensuring the quality of car body welding and guaranteeing the structural reliability of the car body during high-speed operation.

[0048] The 32-line lidar, employing the SICK LMS511, is symmetrically installed in the trackside inspection area. It detects wheelset profiles, tread wear, and track geometry. The 3D point cloud data is transmitted to the trackside edge computing node via UDP protocol, and then connected to the onboard host via a fiber optic ring network. The 32-line lidar, symmetrically installed on both sides of the trackside inspection area, transmits 3D point cloud data to the onboard host via UDP protocol. Point cloud modeling allows analysis of parameters such as wheelset diameter, tread wear, and flange thickness. Simultaneously, it detects geometric defects such as track elevation and gauge variations, providing data support for wheelset inspection and track maintenance, ensuring wheel-rail matching accuracy, and reducing operating noise and wear risks.

[0049] Multidimensional data obtained from sensor devices and fusion positioning systems is preprocessed, such as filtered and dimensionality reduced, by the onboard host and / or trackside edge nodes, and then uploaded to the central control host through the network transmission layer.

[0050] The fusion positioning system includes a vehicle-mounted BeiDou-UWB terminal that integrates a BeiDou module and a UWB tag, which transmits location information to the vehicle-mounted host via an SPI bus; the trackside UWB base station is connected to the trackside switch via Ethernet, and the positioning data is synchronized to the central control host via a fiber optic ring network.

[0051] Edge computing nodes: The onboard host incorporates an adaptive Kalman filter algorithm to perform FFT transformation on the original vibration signal, extract feature values, compress them, and then upload them. The trackside edge nodes perform outlier filtering on the laser point cloud and synchronously package it with the terahertz image into a spatiotemporally labeled data stream. The trackside edge nodes use the RANSAC algorithm to filter outliers in the laser point cloud, retaining valid points with errors ≤0.1mm, achieving a compression ratio of 10:1.

[0052] Specifically, the onboard terminal integrates a BeiDou module and a UWB tag, installed in the center of the vehicle's undercarriage. It communicates with the onboard edge computing host via an SPI bus, sending the train's real-time position (latitude, longitude, and speed) to the edge computing host to provide geographic coordinate references for sensor data. The trackside base station (UWB positioning base station) is deployed every 20 meters within the maintenance depot, connected to a trackside switch via Ethernet. It primarily provides positioning references (positioning error ±5cm) for unmanned execution layer equipment such as AGV robots and six-axis robotic arms, without directly collecting train position information; the train's position is independently obtained by the onboard BeiDou-UWB terminal.

[0053] The data transmission path is as follows: Onboard sensor data: Vibration acceleration sensors, infrared thermal imaging arrays, etc., are connected to the onboard host via a 485 bus and gigabit Ethernet. The onboard host transmits the pre-processed data to the 10-gigabit fiber optic ring network via a fiber optic cable running through the rear of the vehicle, and then to the ground central control host (core equipment of the intelligent decision-making layer). Trackside sensor data: Ultrasonic phased array probes, terahertz imagers, 32-line lidar, etc., are connected to trackside edge nodes or trackside switches via waterproof cables, fiber optics, and UDP protocol. The data is then fed into the onboard host via the fiber optic ring network and transmitted to the central control host. Positioning data: The onboard Beidou-UWB terminal transmits location information to the onboard host via an SPI bus. The trackside UWB base station connects to the trackside switch via Ethernet. The positioning data is synchronized to the central control host via the fiber optic ring network, providing spatiotemporal tags for the sensor data.

[0054] The network transmission layer is used for real-time data transmission and remote control between the IoT sensing layer, intelligent decision-making layer, and unmanned execution layer. The network transmission layer includes wired and wireless networks, which are used to build a hybrid wired ring network and wireless mesh network to achieve end-to-end latency ≤100ms, signal availability 99.99%, and support concurrent access of 100,000+ devices.

[0055] Specifically, such as Figure 3As shown, the wired network of the network transmission layer adopts a 10 Gigabit fiber optic ring network as its backbone, and the vehicle-mounted host is connected to the trackside switch through a fiber optic cable at the rear of the vehicle; the wireless network deploys high-gain directional antennas through trackside wireless access points (APs), and the maintenance depot uses omnidirectional antennas. The vehicle-mounted wireless unit supports cross-area handover, achieving an information transmission delay of ≤100ms between IoT sensing layer sensors and vehicle-mounted host / trackside edge nodes, vehicle-mounted host and ground central control host (intelligent decision-making layer), central control host and unmanned execution layer equipment (AGV, robotic arm, etc.), and each layer of equipment and the safety protection layer. All IoT sensing layer data is transmitted through the network transmission layer.

[0056] like Figure 4 As shown, the intelligent decision-making layer includes a multi-source data fusion module and an intelligent operation and maintenance knowledge graph module, used for fault diagnosis, life prediction, and automated generation of maintenance work orders for the components under test. The multi-source data fusion module has an LSTM time series analysis model and a Transformer spatiotemporal attention model. The LSTM time series analysis model is used to capture the time series features of multi-dimensional data, while the Transformer model is used to set the association weights of data in multi-dimensional data and output the fault type. The hardware configuration of the multi-source data fusion module is based on a ground-based central control host (referred to as the "central host") as its core, connected to an external GPU acceleration cluster and a Neo4j graph database server. The data transmission and processing process of the multi-source data fusion module is as follows: the vehicle host transmits the pre-processed data to the central control host, which then distributes it to the GPU cluster via the PCIe 4.0 bus. The multi-source data fusion module has an LSTM+Transformer model. LSTM captures time series features (such as the trend of vibration signal changes over time), and Transformer calculates the correlation weights of vibration-temperature-wear data (such as the correlation between vibration abnormality and bearing temperature rise), and outputs the fault type (such as bearing wear, tread cracks, weld defects, etc.), with an accuracy of 98%.

[0057] The multi-source data fusion module consists of an input layer, an attention layer, and an output layer. The input layer receives preprocessed multi-dimensional data and captures time-series features through an LSTM network. The attention layer calculates the correlation weights of the data in each dimension. The output layer outputs the fault type.

[0058] Data input: preprocessed multidimensional time series data such as vibration acceleration data (10-5000Hz) from the IoT sensing layer, infrared thermal imaging temperature data, ultrasonic phased array crack detection data, and lidar contour data.

[0059] In the context of intelligent operation and maintenance of high-speed railways, this invention utilizes an LSTM model to capture the dynamic changes in sensor time-series data (such as vibration and temperature), providing temporal feature support for fault diagnosis. The core formula of the LSTM layer is as follows:

[0060] h t =σ(W h ·[h t-1 ,x t ]+b h )☉tanh(W c ·[h t-1 ,x t ]+b c );

[0061] Among them, h t σ represents the hidden state at time t, which is the output of the LSTM layer after processing the input information at the current time. It contains a fusion of historical information and current input features, used to pass important temporal information to subsequent time steps or layers. σ is the sigmoid function, which maps the input to a value between 0 and 1, and is used in LSTM to control the retention or forgetting of information. W h It is for the hidden state h t-1 And input x t The concatenated linear transformation weight matrix is ​​used to transform [h] t-1 ,x t ](h t-1 and x t Perform a linear transformation on the concatenated vector, which means multiplying it by the matrix and adding the bias b. h To extract relevant features. t-1 The hidden state at time t-1 represents the temporal information saved by the LSTM layer in previous time steps; it is the historical state passed to the current time step. t The input data at time t is the external input information received by the LSTM layer at the current time. h In order to cooperate with W h The corresponding bias term, in relation to [h] t-1 ,x t Added during linear transformations to adjust the result of the linear transformation. c It is for [h] t-1 ,x t Another set of weight matrices is used to perform linear transformations, in conjunction with the linear transformations related to generating candidate hidden states. h Is with W c The corresponding bias term, in W c For [h] t-1 ,x t Add this when performing a linear transformation to adjust the result of the linear transformation.

[0062] The core objective is to generate a current hidden state h that includes long-term dependent features by fusing historical states with the current input. tThis is used to characterize the dynamic patterns of time-series data (such as the decay or enhancement trend of bearing vibration signals over time). In the formula, σ (sigmoid function) and tanh (hyperbolic tangent function) constitute a simplified gating logic—σ outputs 0-1 "weights" to control the retention ratio of information, and tanh generates candidate features. The two are then used to filter key time-series features (such as high-frequency components related to faults in vibration signals) by multiplying them element by element (⊙).

[0063] The LSTM algorithm performs the fusion calculation of historical states and current inputs, following these steps:

[0064] 1. Input concatenation

[0065] The hidden state h at time t-1 t-1 (Historical time-series characteristics, such as the bearing vibration characteristics at the previous moment) and the input data x at time t t (Current time series data, such as the vibration acceleration value at the current moment) are concatenated into a vector [h] t-1 ,x t This will enable the initial integration of historical and current information.

[0066] 2. Linear Transformation and Activation

[0067] For the concatenated vector [h t-1 ,x t Perform two linear transformations:

[0068] First time: W h ·[h t-1 ,x t ]+b h (W h Let b be the weight matrix. h (As a bias term), used to extract features related to "information filtering".

[0069] Second time: W c ·[h t-1 ,x t ]+b c (W c For another set of weight matrices, b c (For the corresponding bias term), used to extract features related to "candidate feature generation".

[0070] Processed separately using activation functions:

[0071] The first result is activated by the σ function: σ(W) h ·[h t-1 ,x t ]+b h It outputs a "gating weight" of 0-1 to control which information is retained (such as filtering out outliers in vibration signals that exceed a threshold).

[0072] The second result is activated by the tanh function: tanh(W c ·[h t-1 ,x t ]+b c The output is a "candidate hidden state" of -1-1, which contains the fusion features of the current input and historical states (such as the amplitude change trend of the vibration signal).

[0073] 3. Hidden state update

[0074] By combining the "gating weights" with the "candidate hidden states" through element-wise multiplication (⊙), the hidden state h at time t is obtained. t :

[0075] h t =Gated weights ⊙Candidate hidden states

[0076] At this time h t Irrelevant information has been filtered out, while key dependencies in the time series data (such as the increasing pattern of vibration signals at multiple consecutive time points) have been preserved.

[0077] In this invention's system, LSTM is mainly used for time-series data processing from the IoT sensing layer to the intelligent decision-making layer. Its core function is to capture the long-term dependencies of sensor data, supporting early fault warning and diagnosis. Specific applications are as follows:

[0078] 1. Temporal Feature Extraction

[0079] Input data: Time-series data collected by the IoT sensing layer, such as vibration acceleration signals (10-5000Hz) of the running gear bearings and the temperature variation curve of the wheelset over time. For example, during early bearing wear, the high-frequency impact component of the vibration signal gradually increases over time. LSTM uses h... t-1 With x t The continuous integration captures this gradual trend of change.

[0080] Output characteristics: h t As a carrier of temporal characteristics, it transmits dynamic patterns such as "amplitude increase" and "frequency shift" of vibration signals, providing a temporal dimension basis for fault diagnosis.

[0081] 2. Collaborate with Transformer to achieve multi-source data fusion

[0082] In the "LSTM+Transformer" fusion model of the intelligent decision-making layer: LSTM is responsible for extracting single-dimensional temporal features (such as the fluctuation pattern of vibration signals within 1 hour); Transformer calculates the correlation weights of multi-source data (such as vibration and temperature) based on these temporal features (such as the correlation between vibration enhancement and temperature rise); the two work together to achieve a fault diagnosis accuracy of ≥98%. For example, when LSTM captures a continuous enhancement of the high-frequency components of the vibration signal (h... t When the Transformer calculates a strong correlation between the vibration and the temperature rise, the system can determine that it is a bearing wear failure.

[0083] By capturing long-term dependencies, LSTM can identify potential risks in the early stages of a fault (such as abnormal vibration signals in early bearing wear). Combined with data from other sensors, it can achieve "early detection of hidden dangers," reducing the system's false negative rate to <0.05% and supporting the efficient operation of the "perception-decision-execution" closed loop.

[0084] Transformer attention layer:

[0085]

[0086] Where Q, K, and V are query, key, and value matrices, respectively, and the association weights of data in each dimension are calculated. k The dimension of the key vector.

[0087] Output results: Fault types include common faults in key high-speed rail components such as bearing wear, tread cracks, weld defects, wheelset wear, and component loosening, with an accuracy rate of 98%.

[0088] A federated learning group is formed across multiple train depots. The model is trained using local data from each depot without sharing the original data. The global model parameters are aggregated and updated monthly to improve the model's ability to generalize fault diagnosis for different lines and train types, thus avoiding model bias caused by data silos.

[0089] The intelligent operation and maintenance knowledge graph module includes the Neo4j graph database, which stores the relationship between "component-fault-rule-tool". The Drools rule engine calls the graph in real time, generates standardized work orders based on the fault type, and sends them to the unmanned execution layer.

[0090] Specifically, the data transmission and processing of the intelligent operation and maintenance knowledge graph module is as follows: the Neo4j graph database server stores entities and relationships, and is interconnected with the central control host via a 10 Gigabit Ethernet. The Drools rule engine calls the knowledge graph in real time to generate standardized work orders including maintenance steps, tool lists, and spare parts information, and pushes them to the unmanned execution system. Based on the "Railway Locomotive and Rolling Stock Maintenance Regulations" TB / T 3567-2020, association rules between wheelset tread cracks and vibration acceleration are defined, such as "crack ≥ 0.2mm and vibration > 0.5g → Level 3 repair". The path can also be optimized using existing path optimization algorithms, prioritizing allocation to the nearest AGV robot.

[0091] The construction and reasoning of the knowledge graph in the intelligent operation and maintenance knowledge graph module includes construction tools and definitions. Entity definitions: Based on the Neo4j graph database, "components" are defined, such as wheelsets, bearings, and welds; "faults" are defined, such as cracks, wear, and temperature rise; "rules" are defined, such as the maintenance standards in the "Railway Locomotive and Rolling Stock Maintenance Regulations" TB / T 3567-2020; and "tools" are defined, such as ultrasonic flaw detectors and torque wrenches. Relationship definitions: Relationships such as "component-fault" (e.g., "wheelset tread-crack"), "fault-rule" (e.g., "crack ≥ 0.2mm - Level 3 maintenance procedure"), and "fault-tool" (e.g., "crack - ultrasonic phased array probe") are established to form the knowledge graph. Rule parsing: The "Maintenance Procedure Rules" are parsed using natural language processing technology to extract the rule "execute a specific maintenance procedure when the component state meets a certain condition," which is then transformed into entity relationships in the graph, such as "tread crack ≥ 0.2mm and vibration > 0.5g → Level 3 maintenance procedure."

[0092] Inference Engine and Path Optimization: The Drools rule engine receives fault types from the multi-source data fusion module in real time, matches the "fault-rule" relationship in the graph, and automatically generates standardized work orders, including maintenance steps such as "disassembling wheelset → flaw detection → replacing bearing", tool list such as "ultrasonic flaw detector, hydraulic puller", and spare parts information such as "bearing model 6308".

[0093] Dijkstra's algorithm can be used for path optimization.

[0094] D(v)=min(D(v),D(u)+weight(u,v)),

[0095] Where D(v) is the currently known shortest distance from the starting node (e.g., the current position of the AGV robot) to the target node v (e.g., the parts warehouse, maintenance station), D(u) represents the known shortest distance from the starting node to the intermediate node u, and weight(u,v) represents the edge weight between node u and node v. In the high-speed rail operation and maintenance scenario, this can be understood as the actual distance between the two nodes (e.g., track length, path length within the warehouse) or the passage cost (e.g., congestion coefficient, equipment avoidance time). The core logic of min(D(v),D(u)+weight(u,v)) is a "relaxation operation," which compares the currently known shortest distance to v, D(v), with the sum of "the shortest distance from the starting node to u + the weight from u to v," and updates D(v) with the smaller of the two values ​​until the shortest path from the starting node to all target nodes is found. The Dijkstra algorithm is used to calculate the shortest path from the AGV robot's current position to the parts warehouse and maintenance station, prioritizing the allocation to the nearest AGV to reduce empty running distance, reducing the empty running rate to 6%.

[0096] In the system of this invention, the Dijkstra path optimization algorithm supports closed-loop collaboration between the "intelligent decision-making layer and the unmanned execution layer" by accurately calculating the shortest path of the AGV robot: when the intelligent decision-making layer generates a work order, it calls the algorithm to calculate the path and sends it to the AGV; the AGV travels according to the planned path, avoiding the path redundancy problem of traditional manual scheduling, and helping to improve the efficiency of parts replacement by 40%.

[0097] Taking the AGV robot parts scheduling in the unmanned execution layer as an example, the AGV robot needs to go from its current docking point (starting node) to the parts warehouse (node ​​A) to pick up spare parts, and then transport them to the maintenance station (node ​​B). The shortest path needs to be planned to reduce the empty running rate (the system requires the empty running rate to be reduced to 6%). The Dijkstra path optimization algorithm performs the following steps:

[0098] 1. Node and weight definition:

[0099] Starting node: Current position of AGV (e.g., entrance to the maintenance warehouse); Intermediate node u: branch road, obstacle avoidance point, etc. within the warehouse; Target node v: Parts warehouse (A), maintenance station (B); weight(u,v): Physical distance between nodes (e.g., 50 meters from u to A, 30 meters from A to B).

[0100] 2. Path calculation:

[0101] Initially, the distance D (starting node) from the starting node to itself is set to 0, and the distance to other nodes is set to infinity.

[0102] Iterate through intermediate nodes u and update the shortest distance to the target node v using the formula:

[0103] For example, if the current D(A) = 100 meters (the initial distance from the starting node to A), and the distance from the starting node to u, D(u) = 40 meters, and the weight of u to A is 30 meters, then D(A) = min(100, 40 + 30) = 70 meters, that is, the shortest distance to update A is 70 meters.

[0104] Repeat the above steps until the shortest path from the starting node to the parts warehouse (A) and the maintenance station (B) is calculated. Using this algorithm, the AGV robot can automatically plan the optimal path, prioritizing routes with short distances and high throughput, reducing unnecessary empty runs (such as avoiding congested sections of other AGVs), ultimately reducing the empty run rate to 6%, improving the efficiency of maintenance resource scheduling, and adapting to the dynamic operation and maintenance needs of high-speed rail.

[0105] like Figure 5 As shown, the unmanned execution layer receives standardized work orders and control commands generated by the intelligent decision-making layer through the network transmission layer. It then uses intelligent robotic arms, such as a six-axis robotic arm, to grasp spare parts, AGV robots to schedule parts, and remote maintenance terminals, such as AR glasses for guidance and intelligent torque wrenches for data uploading, to automate the entire maintenance process. Based on the standardized work orders and control commands, maintenance operations are executed, and the operation data, such as torque values ​​and parts replacement records, are fed back to the intelligent decision-making layer through the network transmission layer to update the digital twin model, forming a closed loop.

[0106] Specifically, unmanned execution layer equipment can include AGV robots, six-axis robotic arms, tracked inspection robots, AR glasses, and intelligent torque wrenches. The AGV robot's command response process is as follows: receiving a work order → UWB positioning (±10cm) and driving to the parts warehouse → vision camera identifying the spare parts' QR code → robotic arm grasping the parts → transporting them to the maintenance station according to an optimized path (such as the Dijkstra algorithm). The six-axis robotic arm's end effector gripper integrates a vision camera, and its command response process is as follows: receiving the maintenance steps in the work order → connecting to the edge controller via EtherCAT bus → performing disassembly and assembly actions with a positioning accuracy of ±0.5mm → the end effector vision camera confirming that the parts are installed in place. The tracked inspection robot's chassis is equipped with an ultrasonic flaw detector and a laser thickness gauge, communicating with the on-board host via a Wi-Fi 6 module. The detection data is uploaded via an encrypted tunnel, and its command response process is as follows: automatically cruising along the route specified in the work order → inertial navigation positioning → ultrasonic flaw detector and laser thickness gauge scanning components → triggering local magnified detection when a suspected fault is detected → encrypted data upload to the central host. The AR glasses connect to the vehicle's onboard unit via Bluetooth, displaying a 3D assembly animation. The command response process is as follows: scanning the component's QR code → the onboard unit pushes the 3D assembly animation → overlaying torque parameters and installation sequence guidance → real-time display of the robotic arm's operating perspective, assisting in human / robot collaborative work. The intelligent torque wrench's command response process is as follows: tightening the bolt according to the torque value required by the work order → BLE5.0 uploading real-time torque data to the work order system → generating an electronic record to ensure assembly accuracy meets standards.

[0107] It also includes a security protection layer, such as Figure 6 As shown, the safety protection layer, through lidar anti-collision, national cryptographic encryption transmission, and an industrial firewall, achieves a collision risk warning response time of less than 0.2 seconds in the maintenance area, and the network security protection level reaches Level 3 of the Information Security Protection Standard. The safety protection layer includes an intelligent interlocking and anti-collision system, network security equipment, and a communication redundancy module to achieve safety protection and network security assurance in the maintenance area. The intelligent interlocking and anti-collision system features millimeter-wave radar installed at the maintenance depot entrance, and lidar deployed around the depot area to construct a dynamic electronic fence. The network security equipment utilizes an industrial firewall deployed at the control center exit, supporting national cryptographic encryption tunnels and performing two-way authentication for vehicle-to-ground data transmission. The communication redundancy module reports signal strength from the vehicle-mounted wireless unit to the safety protection host, triggering a dual-network switching command. The vehicle-mounted host has a built-in SSD that automatically caches data when communication is interrupted and retransmits it after the signal is restored.

[0108] The security protection layer covers network security and operational safety. Network security includes encrypting and protecting data transmitted between layers (such as sensor data from the IoT sensing layer to the intelligent decision-making layer, and work order instructions from the intelligent decision-making layer to the unmanned execution layer) through industrial firewalls (supporting national standard SM4 encryption) and two-way authentication, achieving a network security level of Level 3. Operational safety includes ensuring the operational safety of unmanned execution layer equipment (AGVs, robotic arms) through lidar anti-collision systems (collision warning response time in maintenance areas <0.2 seconds) and electronic fences, and ensuring the reliability of data transmission between layers through communication redundancy modules (SSD caching, breakpoint resume).

[0109] Example 1

[0110] Examples of application scenarios for wheelset detection using the system of this invention, such as... Figure 7 As shown, the hardware deployment and connection include trackside detection equipment: a 32-line lidar (1.5m from the track center) via UDP protocol → trackside edge node → fiber optic ring network → onboard host; onboard sensors: vibration acceleration sensor (running unit) via 485 bus → onboard host, infrared thermal imaging array via gigabit Ethernet → onboard host; positioning system: Beidou-UWB fusion positioning, SPI bus transmits positioning data to onboard host.

[0111] Specifically, the trackside inspection equipment includes: a 32-line lidar (German SICK LMS511-10100): installed on both sides of the trackside inspection area, 1.5m from the track center and 1.2m high, installed in pairs to cover the entire circumference of the wheelset. It transmits 3D point cloud data to the trackside edge computing node via UDP protocol, and then connects to the onboard host via a fiber optic ring network. A terahertz imager (Japanese Advantest TAS7200): installed 0.8m directly above the track, triggered synchronously with the lidar, transmitting terahertz images to the trackside edge node via a dedicated fiber optic link, spatiotemporally aligned with the lidar point cloud data. Onboard sensors include: a vibration acceleration sensor (Swiss Kistler 8702B50): mounted on the surface of the wheelset bearing housing, fixed with magnetic clamps, and connected to the onboard host via a 485 bus, uploading vibration signals in real time. Infrared thermal imaging array (FLIR A655sc): Installed on the inspection bracket at the bottom of the train, vertically aligned with the bearing, 0.5m away from the component being measured, transmitting thermal images to the onboard host via gigabit Ethernet. Positioning and synchronization system: BeiDou-UWB fusion positioning: The onboard terminal integrates a BeiDou module and a UWB tag, installed in the center of the train's undercarriage, communicating with the onboard host via an SPI bus; one trackside base station is deployed every 20m within the maintenance depot, connected to a trackside switch via Ethernet, providing spatiotemporal tags for the inspection data. Edge computing nodes: The onboard host incorporates an adaptive Kalman filter algorithm to perform FFT transformation on the raw vibration signal, extract feature values, compress them, and upload them; the trackside edge nodes perform outlier filtering on the laser point cloud, synchronously packaging it with the terahertz image into a spatiotemporally tagged data stream. The trackside edge nodes use the RANSAC algorithm to filter outliers in the laser point cloud, retaining valid points with an error ≤0.1mm, achieving a compression ratio of 10:1.

[0112] The data acquisition process includes a triggering mechanism: when a train enters the detection area at a speed of ≤5km / h, the trackside inductive loop triggers the detection equipment to start synchronously, with a single detection cycle of 200ms. Spatiotemporal annotation: the UWB positioning system outputs the train's position coordinates in real time, which are aligned with the sensor data via timestamps to form a "position-time-state" three-dimensional dataset. Preprocessing: the onboard host performs frequency domain analysis on the vibration signal, and the trackside edge nodes perform noise reduction on the laser point cloud. Both types of data are transmitted to the central host via the vehicle-to-ground communication system.

[0113] The data processing flow includes the train entering the detection area at a speed of ≤5km / h, the ground inductive loop triggering the detection, UWB adding spatiotemporal tags to the data, the onboard host performing FFT transformation on the vibration signal, the trackside edge node processing the laser point cloud, and the data being uploaded to the ground central host.

[0114] The intelligent decision-making layer integrates multi-source data. The hardware consists of a central host equipped with an NVIDIA A100 GPU cluster, processing wheel pair contour data, vibration characteristics, and temperature data in parallel. The algorithm uses an LSTM+Transformer model to calculate the correlation weights between vibration, temperature, and wear data. Figure 9 As shown, the input layer receives preprocessed multidimensional data and captures time-series features through an LSTM network; the attention layer calculates the association weights of data in each dimension; and the output layer outputs the fault type. Federated learning optimization: A federated group is formed across multiple train depots, and the global model is updated monthly to improve generalization ability. Knowledge graph reasoning: The Neo4j database stores the "part-fault-rule" relationship, which is called in real time by the Drools engine. For example, when a wheelset crack ≥0.2mm and vibration acceleration >0.5g is detected, a standard work order containing repair steps is automatically generated. Construction tools: Based on the Neo4j graph database, entities and relationships such as "part-fault-rule-tool" are defined, and entity relationships are extracted from the "Repair Procedure Rules". Reasoning engine: The Drools rule engine calls the graph in real time. When a wheelset tread crack ≥0.2mm and vibration acceleration >0.5g is detected, the "Repair Procedure Rules" are automatically associated to generate a standardized work order, including repair steps, tool list, spare parts information, etc., and the path is optimized through algorithms, prioritizing allocation to the nearest AGV robot.

[0115] Unmanned automated operation layer (wheelset maintenance closed loop)

[0116] The intelligent warehousing and logistics system includes AGV robots: These robots achieve ±10cm positioning via UWB tags and trackside base stations. After receiving work orders, they drive to the parts warehouse, precisely docking at shelves using QR code markers. A vision camera identifies the QR code on the bearing spare parts, a robotic arm grasps the parts, and a conveyor belt transports them to the maintenance station. Human-robot collaborative operation: An AR-assisted terminal scans the wheelset's QR code, and the onboard host pushes a 3D disassembly and assembly animation to the glasses, overlaying torque parameters and installation sequence guidance; an intelligent torque wrench uploads torque data to the work order system in real time, creating an electronic record.

[0117] Wheel-rail inspection robot: The tracked chassis is equipped with an ultrasonic flaw detector and a laser thickness gauge, and the lithium battery provides an 8-hour battery life. The inspection process involves automatic cruising along the track, using an inertial navigation system for real-time positioning, and performing a full-coverage scan of the undercarriage components; when a suspected crack is detected, a local magnified scan is triggered, and the inspection data is uploaded to the ground system through a encrypted tunnel.

[0118] System closed-loop verification, such as Figure 8 As shown, this embodiment represents a closed-loop wheelset maintenance system:

[0119] 1. The intelligent decision-making layer generates a work order → the AGV robot receives the instruction → it uses UWB positioning to drive to the parts warehouse.

[0120] 2. A six-axis robotic arm grasps bearing spare parts, AR glasses guide disassembly and assembly, and an intelligent torque wrench uploads torque data.

[0121] 3. After the maintenance is completed, the lidar re-measures the contour → the vehicle host compares with the standard model → the digital twin is updated, forming a closed loop of "perception-decision-execution-feedback".

[0122] Data closed loop: After the wheelset maintenance is completed, the lidar remeasures the contour data, the on-board host compares it with the standard model, generates a maintenance quality report, and updates the digital twin model.

[0123] Communication closed loop: The vehicle-to-ground two-way communication system supports remote configuration of sensor parameters, and the ground system monitors the AGV's operation progress in real time, dynamically calibrating the robotic arm path through a digital twin model.

[0124] Safety closed loop: During maintenance, the lidar of the safety protection layer monitors the distance between the two trains in real time. When the distance between the two trains is less than 5m, the safety protection host cuts off the power supply through hard wire with a response time of less than 0.2 seconds to ensure operational safety.

[0125] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any modifications, alterations, substitutions, and variations made by those skilled in the art to the above embodiments are within the scope of the present invention.

Claims

1. A system for intelligent operation and maintenance of high-speed trains, characterized in that: It includes an IoT sensing layer, a network transmission layer, an intelligent decision-making layer, and an unmanned execution layer. The IoT sensing layer includes sensor devices and a fusion positioning system for detecting the parts of the train and track to be tested. The fusion positioning system provides spatiotemporal labels for the sensor device data, generates multidimensional data with spatiotemporal labels, and sends it to the information processing device. The network transmission layer is used for real-time data transmission and remote control between the IoT sensing layer, intelligent decision-making layer, and unmanned execution layer; the network transmission layer includes wired network and wireless network, used to construct a hybrid wired ring network and wireless mesh network; The intelligent decision-making layer includes a multi-source data fusion module and an intelligent operation and maintenance knowledge graph module. These modules are used for fault diagnosis, lifespan prediction, and automated generation of maintenance work orders for the components under test. The multi-source data fusion module employs an LSTM time series analysis model and a Transformer spatiotemporal attention model. The LSTM model captures the time series features of multi-dimensional data, while the Transformer model calculates the association weights of vibration-temperature-wear data and outputs the fault type. The intelligent operation and maintenance knowledge graph module includes a Neo4j graph database, which stores the "component-fault-rule-tool" relationship. The Drools rule engine calls the graph in real time, generates standardized work orders based on the fault type, and sends them to the unmanned execution layer. The unmanned execution layer receives standardized work orders and control instructions generated by the intelligent decision-making layer. It then executes maintenance operations through the execution equipment based on the standardized work orders and control instructions, and feeds back the operation data to the intelligent decision-making layer through the network transmission layer to update the digital twin model.

2. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: Multidimensional data obtained through sensor devices and fusion positioning systems is filtered and preprocessed by the on-board host and / or trackside edge nodes, and then uploaded to the central control host through the network transmission layer.

3. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: The sensor equipment includes a vibration acceleration sensor for detecting the vibration acceleration of the wheelset bearing housing in the running gear; an infrared thermal imaging array for detecting the temperature rise of the wheelset bearings at the bottom of the train; an ultrasonic phased array probe for detecting cracks in the wheelset tread; a terahertz imager for detecting defects in the car body welds; and a 32-line lidar for detecting wheelset profiles, tread wear, and track geometry.

4. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: The fusion positioning system includes a vehicle-mounted BeiDou-UWB terminal that integrates a BeiDou module and a UWB tag, which transmits location information to the vehicle-mounted host via an SPI bus; the trackside UWB base station is connected to the trackside switch via Ethernet, and the positioning data is synchronized to the central control host via a fiber optic ring network.

5. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: The wired network of the network transmission layer adopts a 10 Gigabit fiber optic ring network as its backbone, and the on-board host is connected to the trackside switch through a fiber optic cable at the rear of the vehicle; the wireless network deploys high-gain directional antennas through trackside wireless access points (APs), the maintenance depot uses omnidirectional antennas, and the on-board wireless unit supports inter-area handover.

6. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: The intelligent operation and maintenance knowledge graph module uses a path optimization algorithm to allocate execution devices in the unmanned execution layer.

7. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: The multi-source data fusion module of the intelligent decision-making layer forms a federated learning group across multiple train depots. The local data of each depot is used to train the model without sharing the original data, and the global model parameters are aggregated and updated monthly.

8. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: Standardized work orders include maintenance procedures, tool lists, and spare parts information.

9. The intelligent operation and maintenance system for high-speed trains according to claim 1, characterized in that: It also includes a safety protection layer, which comprises an intelligent interlocking and anti-collision system, network security equipment, and a communication redundancy module, used to achieve safety protection and network security assurance in the maintenance area. The intelligent interlocking and anti-collision system deploys LiDAR to build a dynamic electronic fence. The network security equipment deploys an industrial firewall at the control center exit, supports national cryptographic encryption tunnels, and performs two-way authentication for vehicle-to-ground data transmission. The communication redundancy module reports signal strength from the vehicle-mounted wireless unit to the safety protection host, triggering a dual-network switching command. The vehicle-mounted host has a built-in SSD that automatically caches data when communication is interrupted and retransmits it after the signal is restored.

10. A system for intelligent operation and maintenance of high-speed trains according to claim 1, characterized in that: The execution equipment includes intelligent robotic arms, AGV robots, tracked inspection robots, AR glasses guidance, and intelligent torque wrenches.

Images