Rail health state detection method based on high-speed railway air-space-ground collaborative perception

Abstract

This invention provides a track health status detection method based on high-speed railway space-air-ground collaborative sensing, comprising: based on a detection mode, retrieving surface deformation interferograms, UAV orthophotos, train vibration data, and track circuit monitoring data corresponding to the target mileage segment under the detection mode from a pre-constructed heterogeneous data spatial index database; extracting static environmental features of the target mileage segment based on the surface deformation interferograms and UAV orthophotos, and extracting dynamic response features based on the train vibration data and track circuit monitoring data; performing cross-attention operations to fuse and obtain global track state features; inputting the global track state features into a classifier to obtain the track health status detection result of the target mileage segment as healthy state, potential defect state, or defect state. This invention achieves efficient fusion and accurate evaluation of heterogeneous asynchronous data, effectively improving the accuracy and robustness of track health status detection.

Description

Technical Field

[0001] This invention relates to the field of track condition detection technology, and in particular to a track health condition detection method based on high-speed railway air-space-vehicle-ground collaborative sensing. Background Technology

[0002] Under the repeated crushing of trains and the long-term effects of the external environment, the track will gradually develop cumulative deformation, cracks and settlement, etc. Therefore, it is necessary to inspect the track to assess its health status in order to prevent track defects from developing into something that affects train operation safety.

[0003] Currently, the assessment of track health status mainly relies on a single vehicle-to-ground sensor method. This method only considers onboard vibration signals or track circuit data for assessment. However, this approach can only perform an overall assessment of the line and is difficult to accurately locate specific defect mileage sections, resulting in a high rate of missed and false alarms.

[0004] To improve positioning accuracy, multi-source fusion methods based on topological graphs have emerged in recent years. These methods construct topological relationships between track nodes using graph neural networks to associate multi-source data, aiming to enhance the detection accuracy of track defects and the accuracy of health status assessment through the synergy of graph topology and multi-source data. However, topology modeling is highly complex, and the spatial alignment accuracy of data from different sources is highly dependent on the accuracy of the topology model, resulting in high costs for graph topology-based health status detection.

[0005] Therefore, how to make full use of multi-source data to conveniently achieve accurate location and evaluation of track defects remains an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing, which solves the shortcomings of existing technologies that cannot make full use of multi-source data and have high detection method complexity. It realizes a track health status assessment method with asynchronous air-space-vehicle-ground collaborative sensing, accurate positioning of track defects, and high timeliness and accuracy.

[0007] This invention provides a method for detecting track health status based on high-speed railway air-space-vehicle-ground collaborative sensing, comprising: Based on the detection mode, the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment under the detection mode are called from the pre-built heterogeneous data spatial index database. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. Static environmental features of the target mileage section are extracted based on the surface deformation interferogram and the UAV orthophoto, and dynamic response features are extracted based on the train vibration data and track circuit monitoring data. Using the dynamic response features as the query vector and the static environment features as the key and value vectors, a cross-attention operation is performed to fuse and obtain the global orbital state features. The global track state features are input into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

[0008] According to the present invention, a track health status detection method based on high-speed railway space-air-ground collaborative sensing is provided, wherein the heterogeneous data spatial index database is constructed in the following manner: The target route is divided into multiple mileage segments, and a unique mileage identifier is assigned to each mileage segment as the primary key; Acquire surface deformation interferograms and UAV images covering the target route, process them to generate surface deformation interferograms and orthophotos respectively, segment them according to geographic coordinates, and associate each segmented image with its corresponding mileage segment through the mileage identifier; The vibration acceleration data collected by the onboard sensors of the train running on the target line and the track circuit monitoring data of the target line are obtained and associated with the corresponding mileage segment through the mileage identifier. A relational database is constructed with the mileage identifier as the primary key, so as to call the surface deformation interferogram, orthophoto, vibration acceleration data and track circuit monitoring data of the corresponding mileage segment based on the mileage identifier.

[0009] According to the present invention, a method for detecting track health status based on high-speed railway space-air-ground collaborative sensing, the step of extracting static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto specifically includes: Spatial registration was performed between the surface deformation interferograms and UAV orthophotos of the target mileage segment, and they were aligned to a unified coordinate system. The aligned surface deformation interferogram is used as the geological texture channel, and the aligned UAV orthophoto is used as the appearance texture channel. Channel fusion is performed to generate a static environment tensor. The generated static environment tensor is input into a visual extraction network to extract deep environment features as the static environment features.

[0010] According to the present invention, a method for detecting track health status based on high-speed railway space-air-ground collaborative sensing is provided. The step of extracting dynamic response features based on the train vibration data and track circuit monitoring data specifically includes: The train vibration data and the track circuit monitoring data are preprocessed respectively; The preprocessed train vibration data and track circuit monitoring data are input into a cascaded model of a one-dimensional convolutional neural network and a long short-term memory network to extract temporal features as the dynamic response features.

[0011] According to the present invention, a track health status detection method based on high-speed railway space-air-ground collaborative sensing is provided, wherein the cross-attention operation includes: The first acquisition time of the surface deformation interferogram and the second acquisition time of the UAV orthophoto are obtained, and the earlier of the two is used as the reference acquisition time. The time difference between the reference acquisition time and the current time is calculated, and a timeliness weight is calculated according to a preset attenuation function. The timeliness weight is negatively correlated with the time difference. The timeliness weight is used to adjust the contribution of the static environmental features in performing the cross-attention operation.

[0012] According to the present invention, a method for detecting track health status based on high-speed railway air-space-vehicle-ground collaborative sensing is provided, wherein the detection mode includes a real-time monitoring mode and a target time period monitoring mode; In the real-time monitoring mode, the train vibration data of the most recent train passing through the target mileage section and the track circuit monitoring data of the most recent preset time window are called up, as well as the most recently stored surface deformation interferogram and UAV orthophoto. In the target time period monitoring mode, train vibration data and track circuit monitoring data passing through the target mileage section within the target time period are retrieved, as well as surface deformation interferograms and UAV orthophotos stored at the most recent historical time from the target time period.

[0013] According to the present invention, a track health status detection method based on high-speed railway space-air-ground collaborative sensing is provided, wherein the classifier is trained in the following manner: Acquire historical data, including historical surface deformation interferograms of the line mileage section, historical UAV orthophotos, historical train vibration data, and historical track circuit monitoring data; The track quality index was extracted from historical train vibration data, and the environmental risk level was extracted based on historical surface deformation interferograms and historical UAV orthophotos. Based on preset threshold rules, and based on the track quality index and environmental risk level of the line mileage section, a health status label is generated for the historical data of the line mileage section. The health status label includes health status, potential disease status, and disease status. On a dataset constructed from historical data with defined labels, the classifier is trained using the historical data as input and its corresponding health status labels as supervisory signals.

[0014] The track health status detection method based on high-speed railway space-air-ground collaborative sensing provided by the present invention further includes: Monitor the test results and the feedback information from manual review; When the frequency of the detection results indicating disease status exceeds a preset threshold, or when the manual review feedback information confirms a missed report, a data update instruction is automatically generated. In response to the data update command, a drone or satellite is scheduled to retake images of the target mileage segment, and the heterogeneous data spatial index database is updated based on the retaken images.

[0015] The present invention also provides a track health status detection device based on high-speed railway air-space-vehicle-ground collaborative sensing, comprising: The calling module is used to call the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment in the detection mode from a pre-built heterogeneous data spatial index database based on the detection mode. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. The extraction module is used to extract static environmental features of the target mileage section based on the surface deformation interferogram and the UAV orthophoto, and to extract dynamic response features based on the train vibration data and track circuit monitoring data. The fusion module is used to perform cross-attention operation with the dynamic response features as query vectors and the static environment features as key vectors and value vectors to fuse and obtain global orbital state features. The detection module is used to input the global track state features into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing as described above.

[0019] The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative perception provided by this invention achieves forced alignment of air-space-vehicle-ground data in the spatial dimension by constructing a heterogeneous data spatial index database with the line mileage segment as the unique primary key. Furthermore, considering the temporal scale differences between air-space data and vehicle-ground data, static environmental features and dynamic response features are extracted separately and cross-attention operations are performed to obtain global features, thereby completing temporal alignment at the semantic level. This enables the detection of track health status to fully utilize the environmental background provided by air-space imagery and the high-frequency dynamic response advantage of the more timely vehicle-ground data, achieving efficient fusion and accurate evaluation of heterogeneous asynchronous data, and effectively improving the accuracy and robustness of track health status detection. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing provided by the present invention. Figure 2 This is a schematic diagram of the track health status detection device based on high-speed railway air-space-vehicle-ground collaborative sensing provided by the present invention. Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] The following is combined Figure 1 This invention introduces a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing, such as... Figure 1 As shown, it includes: Step 101: Based on the detection mode, retrieve the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment in the detection mode from the pre-built heterogeneous data spatial index database. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. Step 102: Extract static environmental features of the target mileage section based on the surface deformation interferogram and the UAV orthophoto, and extract dynamic response features based on the train vibration data and track circuit monitoring data; Step 103: Using the dynamic response features as the query vector and the static environment features as the key vector and value vector, perform cross-attention operation to fuse and obtain global orbital state features; Step 104: Input the global track state features into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

[0024] In order to fully utilize multi-source, heterogeneous, and asynchronous space-ground, air-vehicle-ground data and improve the accuracy of track health status assessment, this invention pre-constructs a relational database with the track mileage segment as the unique primary key, so as to call the space-ground, air-vehicle-ground data corresponding to the mileage segment and realize the spatial alignment of multi-source data.

[0025] The data from space, air, vehicle, and ground includes satellite synthetic aperture radar (SAR) imagery, UAV optical image sequences, train vibration data monitored by onboard vibration sensors, and track circuit monitoring data. Specifically, continuous multi-phase SAR satellite imagery reflects the settlement changes of the soil around the track, UAV optical image sequences reflect the surface morphology around the track, and train vibration data and track circuit monitoring data capture the impact of subtle changes, such as cracks, on the track.

[0026] Preferably, in this embodiment, the surface deformation interferogram is obtained by processing two consecutive satellite synthetic aperture radar images, and the UAV optical image sequence is processed into UAV orthophotos. Then, the images are associated with their corresponding route mileage segments for storage, so as to reduce the amount of data that the heterogeneous data spatial index database needs to store.

[0027] By constructing the heterogeneous data spatial index database, various applications for track health status detection can be performed. For example, during routine track inspections, recently stored space-ground-vehicle-ground data can be used to replace manual inspections and assess the track's health status. Another example is the prediction and analysis of disease evolution trends for specific track mileage sections based on long-term accumulated multi-period data. Furthermore, after extreme weather or geological disasters, historical images and vibration data of affected mileage sections can be quickly retrieved to compare and evaluate the impact of the disaster on track health status.

[0028] Optionally, multiple different detection modes are predefined based on different detection applications. When assessing the track health status, the corresponding space-air-ground data is retrieved according to the data retrieval rules pre-configured for each detection mode. For example, if the detection mode is a routine inspection of a certain mileage segment, the vibration data of the 10 trains that recently passed through that mileage segment is retrieved for evaluation; if the detection mode is a historical status assessment, the vibration data of multiple trains that passed through this mileage segment in the historical period, as well as the space-air data stored in the historical period, are retrieved for evaluation.

[0029] Furthermore, since satellite data based on satellite synthetic aperture radar and UAV orthophoto data collected by UAVs have long update cycles, usually on a weekly or monthly basis, while train vibration data and track circuit monitoring data are collected in real time as the train runs, with sampling frequencies reaching the millisecond level, there is a huge difference in time scale between space-air data and vehicle-to-ground data.

[0030] Based on this, if we forcibly merge the four types of data (space, air, vehicle, and ground) at the timestamp level, the amount of space data within the same mileage segment is far less than that of vehicle-ground data, making effective alignment difficult. Therefore, in this embodiment, space data and vehicle-ground data are processed separately.

[0031] Specifically, features are extracted from aerospace data, namely surface deformation interferograms and UAV orthophotos of the target mileage section, to obtain static environment features; features are extracted from vehicle-to-ground data, namely train vibration data and track circuit monitoring data of the target mileage section, to obtain dynamic response features. Then, through a cross-attention mechanism, using dynamic response features as queries and static environment features as keys and values, semantic alignment is achieved to obtain global track state features.

[0032] This approach eliminates the need to forcibly integrate space-air data and vehicle-to-ground data into the same time grid, thereby fully utilizing the environmental background information provided by space-air data and the high-frequency dynamic response provided by vehicle-to-ground data to achieve efficient fusion and accurate evaluation of heterogeneous asynchronous data.

[0033] Finally, the obtained global track state features are input into a pre-trained classifier to obtain the health assessment results of the target mileage segment under the selected detection mode, which are: healthy state, potential disease state, or disease state.

[0034] Among them, "healthy status" means that the track geometry is good and the environment is safe; "potentially defective status" means that the track geometry has slight deviations but the environment is extremely dangerous; and "defective status" means that the track geometry exceeds the limit and requires immediate repair.

[0035] This invention constructs a heterogeneous data spatial index database with the line mileage segment as the unique primary key, thereby achieving forced alignment of space-air-vehicle-ground data in the spatial dimension. Furthermore, addressing the temporal scale differences between space-air and vehicle-ground data, it extracts static environmental features and dynamic response features respectively, performing cross-attention operations to obtain global features. This achieves temporal alignment at the semantic level, enabling the detection of track health status to fully utilize the environmental background provided by space-air imagery and leverage the high-frequency dynamic response advantage of the more timely vehicle-ground data. This achieves efficient fusion and accurate assessment of heterogeneous asynchronous data, effectively improving the accuracy and robustness of track health status detection.

[0036] In the above process, a cross-attention module is designed to achieve feature fusion. First, the dot product similarity between the query vector and the key matrix is ​​calculated, and a scaling factor is applied to stabilize the gradient. The attention score formula is as follows: ,in This indicates the strength of the association between dynamic features and each environmental feature. , This is the feature dimension of the query vector, used to prevent the inner product value from becoming too large and causing softmax saturation.

[0037] The attention weights are obtained by normalizing the score matrix using the softmax function. Weight The relative importance of environmental context to the current dynamic response was quantified. For example, in road sections with high settlement risk, the corresponding weight values ​​were significantly increased, allowing the model to focus on potential anomalies. A preliminary fusion output was generated using a weighted matrix. ,in This is a single-layer attention result. It is the dimension of the key vector and value vector, encoding the instantaneous modulation of dynamic features by the environmental context.

[0038] Feature interactions are enhanced by stacking multiple layers of attention mechanisms, each including residual connections and layer normalization to optimize training stability. The input to the second layer is... Repeat the above process to generate The Attention function reuses the same computational logic. Multi-layer processing allows the model to capture higher-order correlations, such as the deep association between geological fracture patterns and vibration anomalies. The final output is global orbital state features: ; in L This is the total number of floors (default). L =3), FFN is a feedforward network (consisting of two fully connected layers and ReLU activation). It integrates real-time response and historical environmental information. This design achieves adaptive environmental attention through dynamic weight allocation, which significantly improves the accuracy of disease location. For example, in high-risk subsidence areas, the model automatically amplifies the weight of abnormal signals, reducing the false negative rate, while avoiding the complex topology modeling overhead of traditional graph neural networks.

[0039] This invention relates to a track health status detection method based on high-speed railway space-air-ground collaborative sensing. The heterogeneous data spatial index database is constructed in the following manner: The target route is divided into multiple mileage segments, and a unique mileage identifier is assigned to each mileage segment as the primary key; Acquire surface deformation interferograms and UAV images covering the target route, process them to generate surface deformation interferograms and orthophotos respectively, segment them according to geographic coordinates, and associate each segmented image with its corresponding mileage segment through the mileage identifier; The vibration acceleration data collected by the onboard sensors of the train running on the target line and the track circuit monitoring data of the target line are obtained and associated with the corresponding mileage segment through the mileage identifier. A relational database is constructed with the mileage identifier as the primary key, so as to call the surface deformation interferogram, orthophoto, vibration acceleration data and track circuit monitoring data of the corresponding mileage segment based on the mileage identifier.

[0040] This embodiment provides a method for constructing a heterogeneous data spatial index database.

[0041] First, the track of the target line is divided into multiple mileage segments according to a preset interval. Alternatively, the mileage division method of the existing track line can be directly used.

[0042] In this embodiment, the total length of the line is L, and the basic segment length is... (For example, 200 meters), calculate the number of segments in the mileage segment: ; In the formula, For the floor function, each mileage segment j ( j =1,2,…, NAssign a unique identifier , This is the baseline mileage value for the starting point of the route, such as K0+000. Additional offsets are added for route feature points (such as bridge or tunnel entrances). To ensure the uniqueness of identifiers: ; In the formula, Calculated based on feature point coordinates. This segmentation rule enforces physical alignment, resolving the issue of spatial resolution differences in multi-source data and improving the accuracy of subsequent indexing.

[0043] Then, a spatial index database is built based on the mileage segment division structure: the database architecture is defined as a relational table, containing fields: mileage identifier K (e.g., K100+200, which represents a mileage segment 200 meters away from mileage point K100), latitude and longitude coordinates, and aerospace data index.

[0044] Then, GIS tools (such as ArcGIS) are used to map the aerospace imagery to specific mileage segments, and the coordinates are converted to mileage through georeferencing algorithms.

[0045] Specifically, during the data acquisition phase, satellite SAR data was collected, and surface subsidence information was obtained through interferometric processing. The spatial resolution was set at 5 meters, and surface deformation interferograms were generated to monitor millimeter-level deformation. UAV orthophotos were generated using a multi-rotor platform equipped with a high-resolution optical sensor, with a flight altitude controlled at 100 meters. The image resolution reached 0.05 meters, covering the entire slope area along the route, and was used to detect cracks and foreign object intrusion. The data was stored in GeoTIFF format and embedded in the WGS84 coordinate system to ensure geographic coordinate accuracy.

[0046] A digital surface model is generated by orthorectifying the optical image sequence collected by the UAV through motion recovery structure domain. Based on this model, the original images are orthorectified one by one to eliminate perspective distortion. The corrected images are then automatically stitched together along the seam line to form a large orthorectified image mosaic covering the entire line.

[0047] Based on this, the surface deformation interferogram The large-scale orthophoto mosaic was cropped according to mileage segments to obtain the surface deformation interferogram and UAV orthophoto corresponding to each mileage segment.

[0048] Specifically, a rectangular region of segment j is defined, and the image grid is cropped according to a preset mileage segment: ; In the formula, This is a coordinate-to-line mileage conversion mapping function. and For section jThe minimum and maximum mileage are determined. The cropped image is stored and then downsampled to a uniform size and associated with the corresponding mileage identifier.

[0049] For vehicle-to-ground data, train axle box vibration sensors are used to collect train axle box vibration acceleration sequences. As train vibration data, track circuit impedance is collected. As track circuit monitoring data t This is a timestamp. Mileage tags are recorded synchronously during data collection. Mileage can be corrected by combining GPS (Global Positioning System) and RFID (Radio Frequency Identification) technologies. ; In the formula, The weighting factor (calculated based on signal strength). and The mileage was measured using both GPS and RFID. Data point attribution was ensured to be accurate, with an error of less than 0.5 meters. Based on the corrected mileage... Determine the mileage segment to which the data point belongs and associate the data point with the primary key of the corresponding mileage segment.

[0050] Then organize the database using a relational structure, with the primary key as... The database includes a static environment layer for storing aerospace data and a dynamic response layer for storing vehicle-to-ground data entries. The vehicle-to-ground data also includes a train identifier (TrainID) to retrieve the data of the corresponding train when data is retrieved based on the detection mode.

[0051] This invention constructs a heterogeneous data spatial index database and achieves seamless integration of heterogeneous data through mileage-based forced alignment, supporting efficient queries and laying the foundation for subsequent asynchronous interactions. The beneficial effects are that the physical alignment mechanism avoids complex graph neural network modeling, simplifies topological associations, and improves the accuracy of hazard location.

[0052] This invention relates to a track health status detection method based on high-speed railway air-space-ground collaborative sensing. The step of extracting static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto specifically includes: Spatial registration was performed between the surface deformation interferograms and UAV orthophotos of the target mileage segment, and they were aligned to a unified coordinate system. The aligned surface deformation interferogram is used as the geological texture channel, and the aligned UAV orthophoto is used as the appearance texture channel. Channel fusion is performed to generate a static environment tensor. The generated static environment tensor is input into a visual extraction network to extract deep environment features as the static environment features.

[0053] In this embodiment, the surface deformation interferogram and UAV orthophoto of the target mileage segment are spatially registered and channel fused. Then, a deep environmental feature tensor that can characterize the geological conditions and surface morphology of the target mileage segment is extracted through a visual Transformer network.

[0054] Specifically, based on the unique mileage identifier of the target mileage segment, the system initiates an asynchronous query request to the database deployed in the cloud, and calls the surface deformation interferogram and UAV orthophoto of the target mileage segment.

[0055] Then, spatial registration is performed on the above images: the surface deformation interferogram and the UAV orthophoto are aligned to a unified coordinate system using a geographic coordinate transformation matrix to eliminate scale differences. The registration formula is as follows: ; In the formula, It is the tensor of the surface deformation interferogram image. It is the orthophoto tensor of the UAV. T It is an affine transformation matrix, calculated based on the coordinates of the corner points of the mileage segment.

[0056] Aligned surface deformation interferograms are used as geological texture channels to characterize subsidence risk; aligned UAV orthophotos are used as appearance texture channels to characterize surface morphology. Both are unified to the same spatial dimension through bilinear interpolation to generate a multi-channel static environment tensor. ,in H and W It has a uniform height and width. C This is the number of channels (default C=4, including RGB and depth). This tensor is input into a visual Transformer network (such as the Swing Transformer) to extract static environment features. Where SwinT represents the forward propagation of the network, and the output feature tensor has a dimension of . .

[0057] This invention relates to a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing. The step of extracting dynamic response features based on the train vibration data and track circuit monitoring data specifically includes: The train vibration data and the track circuit monitoring data are preprocessed respectively; The preprocessed train vibration data and track circuit monitoring data are input into a cascaded model of a one-dimensional convolutional neural network and a long short-term memory network to extract temporal features as the dynamic response features.

[0058] In this embodiment, the extracted vibration acceleration time-series signal and track circuit parameter sequence are preprocessed, and then a one-dimensional convolutional neural network and a long short-term memory network are used to extract the time-series feature vector characterizing the dynamic response of the track structure.

[0059] Specifically, first, the retrieved train vibration data... Track circuit monitoring Preprocessing: A Butterworth low-pass filter (cutoff frequency 5Hz) is applied to remove high-frequency noise, and Z-score normalization is performed, as shown in the formula below. ,in and These are the sequence mean and standard deviation.

[0060] Optionally, when calling single train passage data in the detection mode, the preprocessed data of that single train is used directly; when calling multiple train passage data in the detection mode, a sliding window is applied to the historical multiple passage data (e.g., the most recent 10 passages), with the window size w configurable (default 200 samples), and statistical features such as the mean are calculated. Sum of variance: ; Forming strong timing input , L It is the sequence length. F This is the number of features (including vibration and electrical parameters). This input is fed into a cascaded model of a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM): the 1D-CNN uses convolutional kernels to extract local features, and the LSTM captures long-term dependencies, outputting a dynamic response feature vector. .

[0061] The dimension of the dynamic response feature vector is .

[0062] This invention significantly improves efficiency through parallel processing, with static and dynamic streams running independently, and reduces overall latency through an asynchronous mechanism. After feature extraction, and As input for subsequent interactive modules. In terms of beneficial effects, the use of sliding window statistical processing of historical data enhances feature robustness, smooths random noise, and highlights trend anomalies; at the same time, the hierarchical attention mechanism of the visual Transformer effectively captures subtle geological changes (such as cracks or subsidence), improving feature discrimination power compared to traditional CNNs and supporting accurate location of diseased sections.

[0063] This invention relates to a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing, wherein the cross-attention operation includes: The first acquisition time of the surface deformation interferogram and the second acquisition time of the UAV orthophoto are obtained, and the earlier of the two is used as the reference acquisition time. The time difference between the reference acquisition time and the current time is calculated, and a timeliness weight is calculated according to a preset attenuation function. The timeliness weight is negatively correlated with the time difference. The timeliness weight is used to adjust the contribution of the static environmental features in performing the cross-attention operation.

[0064] Furthermore, since the update frequency of aerospace data is low, older aerospace data may provide incorrect environmental priors. Therefore, in this embodiment, a time difference is introduced to construct a timeliness weight, and the contribution of static environmental features based on aerospace data in the cross-attention operation is dynamically adjusted so that static environmental features closer to the evaluation time have higher weights.

[0065] Specifically, when the detection inference is initiated, the system retrieves the latest aerospace imagery data and determines the first acquisition time of the surface deformation interferogram and the second acquisition time of the UAV orthophoto.

[0066] Optionally, in this embodiment, the time when the surface deformation interferogram and the UAV orthophoto are stored in the database is used as their corresponding acquisition time.

[0067] Then, the acquisition time that is further from the current time is used as the baseline acquisition time. Calculate the baseline acquisition time and the current time. Time difference .

[0068] A predefined decay function is defined based on the exponential function: In the formula, The attenuation coefficient is adjustable. This indicates the weight of timeliness.

[0069] This decay function makes Increased imagery, meaning that when images become old, they rapidly decay to near zero, causing decisions to favor vehicle-to-ground data with better timeliness.

[0070] The static environment features are weighted using timeliness weights, so that the adjusted key vector and value vector are respectively... and ,in, and These are the original key vectors and value vectors, i.e., static environment features.

[0071] Based on this, attention scores are obtained by performing cross-attention operations on the weighted key vectors and value vectors.

[0072] This invention effectively prevents misjudgments caused by outdated prior information by using dynamic weights based on time differences, thereby improving the robustness of the system in dynamic environments.

[0073] This invention relates to a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing, wherein the detection modes include a real-time monitoring mode and a target time period monitoring mode. In the real-time monitoring mode, the train vibration data of the most recent train passing through the target mileage section and the track circuit monitoring data of the most recent preset time window are called up, as well as the most recently stored surface deformation interferogram and UAV orthophoto. In the target time period monitoring mode, train vibration data and track circuit monitoring data passing through the target mileage section within the target time period are retrieved, as well as surface deformation interferograms and UAV orthophotos stored at the most recent historical time from the target time period.

[0074] In this embodiment, the detection mode includes a real-time monitoring mode, which is used for daily inspection of the track and to evaluate the track health status of the target line by calling the latest stored space-ground-vehicle data of the target line.

[0075] Specifically, under the real-time monitoring model, the most recently stored surface interference deformation interferogram and UAV orthophoto, as well as the latest train vibration data, are called up; and the most recently preset time window, such as the track circuit monitoring data of the most recent hour, is called up to extract static environmental features and dynamic response features.

[0076] It also includes a target time period monitoring mode, which is used to assess the track health status of the track line within a user-specified time window.

[0077] In the target time period monitoring mode, for vehicle-to-ground data, the system calls up all train vibration data and track circuit monitoring data within the target time period; for space-air data, it calls up the surface deformation interferogram and UAV orthophotos stored in the historical time closest to the target time period.

[0078] This invention relates to a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing. The classifier is trained in the following manner: Acquire historical data, including historical surface deformation interferograms of the line mileage section, historical UAV orthophotos, historical train vibration data, and historical track circuit monitoring data; The track quality index was extracted from historical train vibration data, and the environmental risk level was extracted based on historical surface deformation interferograms and historical UAV orthophotos. Based on preset threshold rules, and based on the track quality index and environmental risk level of the line mileage section, a health status label is generated for the historical data of the line mileage section. The health status label includes health status, potential disease status, and disease status. On a dataset constructed from historical data with defined labels, the classifier is trained using the historical data as input and its corresponding health status labels as supervisory signals.

[0079] The classifier used in this embodiment is a multilayer perceptron classifier. Unlike conventional binary classification tasks, this invention combines the orbital quality index threshold from dynamic response features with the environmental risk classification from static environmental features to define a fine-grained health status matrix.

[0080] The health status matrix is ​​divided into three fine-grained levels: healthy status, indicating no risk; potential disease status, indicating monitoring; and disease status, indicating that immediate intervention is required.

[0081] The above three levels are jointly determined based on the orbital quality index (TQI) threshold in the dynamic response characteristics and the environmental risk level in the static environmental data.

[0082] The TQI index is extracted from train vibration data, and the environmental risk level E is extracted based on surface deformation interferograms and UAV orthophotos. E is discretized into 1 (low risk), 2 (medium risk), and 3 (high risk).

[0083] Based on the following pre-set threshold rules, the true label for each piece of space-air-ground data is determined: If T≤T low If E≤1, then the label is "healthy". If T > T high If E ≥ 2, then the label is "disease state"; Otherwise, it is labeled as a potential disease condition.

[0084] Among them, T low and T high It is based on the pre-set thresholds calibrated according to engineering standards.

[0085] The historical surface deformation interferograms, historical UAV orthophotos, historical train vibration data, and historical track circuit monitoring data collected are labeled using the above method to form historical space-ground vehicle-ground data-label pairs. The classifier is trained using the historical space-ground vehicle-ground data as input and its corresponding health status response label as the supervision signal.

[0086] In this embodiment, the classifier contains two fully connected layers.

[0087] The first layer uses the ReLU activation function, and the output is... ,in This is the weight matrix. For bias vectors, h The number of hidden units; The second layer outputs a logits vector. , , .

[0088] Finally, the softmax function is applied to generate the probability distribution. , This represents the probability of being categorized as healthy, potentially diseased, or diseased. The architecture captures the complex relationships between features through non-linear mapping and supports multi-class outputs.

[0089] Preferably, to address the class imbalance problem where faulty samples are far fewer than normal samples in practice, a focused loss function is used during model training. This function automatically reduces the weight of easily classified normal samples, while focusing training on potentially diseased samples that are difficult to classify, thereby improving the model's sensitivity in identifying early and minor diseases. During model inference, the specific health level and corresponding confidence level of the target mileage segment are output, achieving a leap from "overall assessment" to "precise localization".

[0090] The loss function is defined as follows: ; In the formula, It is a true label encoded in one-hot encoding. To predict probabilities for the model, It is a category weighting factor. Focus parameters. Settings. For example, inverse class frequency , 8. To increase the weighting of Grade B (potential disease status) and Grade C (disease status), Enhanced penalty for difficult-to-classify samples. This loss automatically reduces the contribution of high-probability, easily-classified samples, such as normal Grade A (healthy state), and focuses on low-probability potential disease samples (such as Grade B and C), significantly improving the sensitivity of early minor diseases and reducing the false negative rate.

[0091] The classifier inference outputs the health level and confidence score. For the target mileage segment, the input... ,calculate Predicting health level At the same time, output the confidence level. .

[0092] Confidence level reflects the reliability of the model's decisions and is used for subsequent maintenance decisions. For example, if... and If this is detected, an early warning will be triggered. The entire process achieves a leap from overall assessment to precise positioning through fine-grained classification and confidence quantification.

[0093] This invention includes a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing, and further comprises: Monitor the test results and the feedback information from manual review; When the frequency of the detection results indicating disease status exceeds a preset threshold, or when the manual review feedback information confirms a missed report, a data update instruction is automatically generated. In response to the data update command, a drone or satellite is scheduled to retake images of the target mileage segment, and the heterogeneous data spatial index database is updated based on the retaken images.

[0094] In this implementation, the system monitors the evaluation results (such as health status level) and manual review feedback in real time, and builds a feedback log database.

[0095] Based on alarm data in the feedback log library, for any mileage segment, if the frequency of detection results indicating a diseased state exceeds a preset threshold, or if manual verification confirms a missed report (i.e., manual verification shows the actual health state is a diseased state), a data update command is automatically generated. The command triggering condition can be formulated as follows: ; In the formula, Alarm frequency, The preset threshold is (e.g., 0.1 times / day). I For indicator functions, This indicates a missed report flag. The repeat shooting command calls the UAV or satellite mission scheduling system via API, prioritizes updating the target area data, processes the newly acquired images, and sends them back to the database, completing closed-loop optimization.

[0096] The aforementioned re-shooting triggering process achieves adaptive data freshness maintenance, reduces the risk of missed reports through high-frequency re-shooting, and improves overall monitoring accuracy. Regarding beneficial effects, the timeliness-weighted mechanism (…) (Attenuation) solves the bottleneck of slow updates in aerospace data, ensuring that the fusion model remains reliable in the event of sudden environmental changes; closed-loop feedback optimizes resource allocation by intelligently triggering re-scanning, avoids redundant scanning across the entire line, and reduces operating costs.

[0097] This invention constructs a unified heterogeneous data spatial index database using railway mileage segments as the unique primary key. Static environmental data, such as surface deformation maps and digital surface models generated by space-air remote sensing, are linked with dynamic response data, such as vehicle vibration and track circuit data corrected by GPS and RFID joint positioning, all indexed by mileage segment identifiers. This forms a two-layer data architecture of "static environment" and "dynamic response." When a specific mileage segment needs to be evaluated, the corresponding multimodal data is asynchronously retrieved through the mileage index. Environmental features and dynamic response features are extracted separately, and then fused using an environmental context attention mechanism to quantify the correlation between the train's dynamic response and the static environment of that segment. Subsequently, combining the track quality index and environmental risk classification, a focused loss function is used to solve the category imbalance problem, achieving a refined health status assessment. Finally, the system is deployed in the cloud, introducing a data timeliness weighting mechanism. The environmental feature weights are dynamically adjusted based on the difference between the space-air data acquisition time and the current operating time. Simultaneously, assessment feedback triggers drone or satellite re-enhancing, forming a complete closed loop of data acquisition, fusion assessment, and feedback optimization.

[0098] The track health status detection device based on high-speed railway air-space-vehicle-ground collaborative sensing provided by the present invention is described below. The track health status detection device based on high-speed railway air-space-vehicle-ground collaborative sensing described below can be referred to in correspondence with the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing described above.

[0099] like Figure 2 As shown, the track health status detection device based on high-speed railway air-space-vehicle-ground collaborative perception includes a calling module 201, an extraction module 202, a fusion module 203, and a detection module 204. The calling module 201 is used to call the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment in the detection mode from a pre-built heterogeneous data spatial index database based on the detection mode. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. Extraction module 202 is used to extract static environmental features of the target mileage section based on the surface deformation interferogram and the UAV orthophoto, and to extract dynamic response features based on the train vibration data and track circuit monitoring data; The fusion module 203 is used to perform cross-attention operation with the dynamic response features as query vectors and the static environment features as key vectors and value vectors to fuse and obtain global orbital state features. The detection module 204 is used to input the global track state features into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

[0100] In order to fully utilize multi-source, heterogeneous, and asynchronous space-ground, air-vehicle-ground data and improve the accuracy of track health status assessment, this invention pre-constructs a relational database with the track mileage segment as the unique primary key, so as to call the space-ground, air-vehicle-ground data corresponding to the mileage segment and realize the spatial alignment of multi-source data.

[0101] The data from space, air, vehicle, and ground includes satellite synthetic aperture radar (SAR) imagery, UAV optical image sequences, train vibration data monitored by onboard vibration sensors, and track circuit monitoring data. Specifically, continuous multi-phase SAR satellite imagery reflects the settlement changes of the soil around the track, UAV optical image sequences reflect the surface morphology around the track, and train vibration data and track circuit monitoring data capture the impact of subtle changes, such as cracks, on the track.

[0102] Preferably, in this embodiment, the surface deformation interferogram is obtained by processing two consecutive satellite synthetic aperture radar images, and the UAV optical image sequence is processed into UAV orthophotos. Then, the images are associated with their corresponding route mileage segments for storage, so as to reduce the amount of data that the heterogeneous data spatial index database needs to store.

[0103] By constructing the heterogeneous data spatial index database, various applications for track health status detection can be performed. For example, during routine track inspections, recently stored space-ground-vehicle-ground data can be used to replace manual inspections and assess the track's health status. Another example is the prediction and analysis of disease evolution trends for specific track mileage sections based on long-term accumulated multi-period data. Furthermore, after extreme weather or geological disasters, historical images and vibration data of affected mileage sections can be quickly retrieved to compare and evaluate the impact of the disaster on track health status.

[0104] Optionally, multiple different detection modes are predefined based on different detection applications. When assessing the track health status, the corresponding space-air-ground data is retrieved according to the data retrieval rules pre-configured for each detection mode. For example, if the detection mode is a routine inspection of a certain mileage segment, the vibration data of the 10 trains that recently passed through that mileage segment is retrieved for evaluation; if the detection mode is a historical status assessment, the vibration data of multiple trains that passed through this mileage segment in the historical period, as well as the space-air data stored in the historical period, are retrieved for evaluation.

[0105] Furthermore, since satellite data based on satellite synthetic aperture radar and UAV orthophoto data collected by UAVs have long update cycles, usually on a weekly or monthly basis, while train vibration data and track circuit monitoring data are collected in real time as the train runs, with sampling frequencies reaching the millisecond level, there is a huge difference in time scale between space-air data and vehicle-to-ground data.

[0106] Based on this, if we forcibly merge the four types of data (space, air, vehicle, and ground) at the timestamp level, the amount of space data within the same mileage segment is far less than that of vehicle-ground data, making effective alignment difficult. Therefore, in this embodiment, space data and vehicle-ground data are processed separately.

[0107] Specifically, features are extracted from aerospace data, namely surface deformation interferograms and UAV orthophotos of the target mileage section, to obtain static environment features; features are extracted from vehicle-to-ground data, namely train vibration data and track circuit monitoring data of the target mileage section, to obtain dynamic response features. Then, through a cross-attention mechanism, using dynamic response features as queries and static environment features as keys and values, semantic alignment is achieved to obtain global track state features.

[0108] This approach eliminates the need to forcibly integrate space-air data and vehicle-to-ground data into the same time grid, thereby fully utilizing the environmental background information provided by space-air data and the high-frequency dynamic response provided by vehicle-to-ground data to achieve efficient fusion and accurate evaluation of heterogeneous asynchronous data.

[0109] Finally, the obtained global track state features are input into a pre-trained classifier to obtain the health assessment results of the target mileage segment under the selected detection mode, which are: healthy state, potential disease state, or disease state.

[0110] Among them, "healthy status" means that the track geometry is good and the environment is safe; "potentially defective status" means that the track geometry has slight deviations but the environment is extremely dangerous; and "defective status" means that the track geometry exceeds the limit and requires immediate repair.

[0111] This invention constructs a heterogeneous data spatial index database with the line mileage segment as the unique primary key, thereby achieving forced alignment of space-air-vehicle-ground data in the spatial dimension. Furthermore, addressing the temporal scale differences between space-air and vehicle-ground data, it extracts static environmental features and dynamic response features respectively, performing cross-attention operations to obtain global features. This achieves temporal alignment at the semantic level, enabling the detection of track health status to fully utilize the environmental background provided by space-air imagery and leverage the high-frequency dynamic response advantage of the more timely vehicle-ground data. This achieves efficient fusion and accurate assessment of heterogeneous asynchronous data, effectively improving the accuracy and robustness of track health status detection.

[0112] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing. This method includes: based on a detection mode, calling from a pre-constructed heterogeneous data spatial index database the corresponding surface deformation interferogram, UAV orthophoto, train vibration data, and track circuit monitoring data for the target mileage segment under the detection mode; the heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key; extracting static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto, and extracting dynamic response features based on the train vibration data and track circuit monitoring data; performing cross-attention operation using the dynamic response features as a query vector and the static environmental features as a key vector and value vector to fuse and obtain global track status features; inputting the global track status features into a classifier to obtain the track health status detection result of the target mileage segment as healthy, potentially defective, or defective.

[0113] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative perception provided by the above methods. The method includes: based on a detection mode, calling the surface deformation interferogram, UAV orthophoto, train vibration data, and track circuit monitoring data corresponding to the target mileage segment in the detection mode from a pre-constructed heterogeneous data spatial index database, wherein the heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key; extracting static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto, and extracting dynamic response features based on the train vibration data and track circuit monitoring data; performing cross-attention operation with the dynamic response features as query vector and the static environmental features as key vector and value vector to fuse and obtain global track status features; inputting the global track status features into a classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential defect status, or defect status.

[0115] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing provided by the above methods. The method includes: based on a detection mode, retrieving from a pre-constructed heterogeneous data spatial index database the corresponding surface deformation interferogram, UAV orthophoto, train vibration data, and track circuit monitoring data of the target mileage segment under the detection mode, wherein the heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key; extracting static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto, and extracting dynamic response features based on the train vibration data and track circuit monitoring data; performing cross-attention operation using the dynamic response features as a query vector and the static environmental features as a key vector and value vector to fuse and obtain global track status features; inputting the global track status features into a classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential defect status, or defect status.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting track health status based on high-speed railway air-space-vehicle-ground collaborative sensing, characterized in that, include: Based on the detection mode, the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment under the detection mode are called from the pre-built heterogeneous data spatial index database. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. Static environmental features of the target mileage section are extracted based on the surface deformation interferogram and the UAV orthophoto, and dynamic response features are extracted based on the train vibration data and track circuit monitoring data. Using the dynamic response features as the query vector and the static environment features as the key and value vectors, a cross-attention operation is performed to fuse and obtain the global orbital state features. The global track state features are input into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

2. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to claim 1, characterized in that, The heterogeneous data spatial index database is constructed in the following manner: The target route is divided into multiple mileage segments, and a unique mileage identifier is assigned to each mileage segment as the primary key; Acquire surface deformation interferograms and UAV images covering the target route, process them to generate surface deformation interferograms and orthophotos respectively, segment them according to geographic coordinates, and associate each segmented image with its corresponding mileage segment through the mileage identifier; The vibration acceleration data collected by the onboard sensors of the train running on the target line and the track circuit monitoring data of the target line are obtained and associated with the corresponding mileage segment through the mileage identifier. A relational database is constructed with the mileage identifier as the primary key, so as to call the surface deformation interferogram, orthophoto, vibration acceleration data and track circuit monitoring data of the corresponding mileage segment based on the mileage identifier.

3. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to claim 1, characterized in that, The step of extracting the static environmental features of the target mileage segment based on the surface deformation interferogram and the UAV orthophoto specifically includes: Spatial registration was performed between the surface deformation interferograms and UAV orthophotos of the target mileage segment, and they were aligned to a unified coordinate system. The aligned surface deformation interferogram is used as the geological texture channel, and the aligned UAV orthophoto is used as the appearance texture channel. Channel fusion is performed to generate a static environment tensor. The generated static environment tensor is input into a visual extraction network to extract deep environment features as the static environment features.

4. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to claim 1, characterized in that, The step of extracting dynamic response features based on the train vibration data and track circuit monitoring data specifically includes: The train vibration data and the track circuit monitoring data are preprocessed respectively; The preprocessed train vibration data and track circuit monitoring data are input into a cascaded model of a one-dimensional convolutional neural network and a long short-term memory network to extract temporal features as the dynamic response features.

5. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to claim 1, characterized in that, The cross-attention operation includes: The first acquisition time of the surface deformation interferogram and the second acquisition time of the UAV orthophoto are obtained, and the earlier of the two is used as the reference acquisition time. The time difference between the reference acquisition time and the current time is calculated, and a timeliness weight is calculated according to a preset attenuation function. The timeliness weight is negatively correlated with the time difference. The timeliness weight is used to adjust the contribution of the static environmental features in performing the cross-attention operation.

6. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to any one of claims 1-5, characterized in that, The detection modes include real-time monitoring mode and target time period monitoring mode; In the real-time monitoring mode, the train vibration data of the most recent train passing through the target mileage section and the track circuit monitoring data of the most recent preset time window are called up, as well as the most recently stored surface deformation interferogram and UAV orthophoto. In the target time period monitoring mode, train vibration data and track circuit monitoring data passing through the target mileage section within the target time period are retrieved, as well as surface deformation interferograms and UAV orthophotos stored at the most recent historical time from the target time period.

7. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to any one of claims 1-5, characterized in that, The classifier is trained in the following manner: Acquire historical data, including historical surface deformation interferograms of the line mileage section, historical UAV orthophotos, historical train vibration data, and historical track circuit monitoring data; The track quality index was extracted from historical train vibration data, and the environmental risk level was extracted based on historical surface deformation interferograms and historical UAV orthophotos. Based on preset threshold rules, and based on the track quality index and environmental risk level of the line mileage section, a health status label is generated for the historical data of the line mileage section. The health status label includes health status, potential disease status, and disease status. On a dataset constructed from historical data with defined labels, the classifier is trained using the historical data as input and its corresponding health status labels as supervisory signals.

8. The track health status detection method based on high-speed railway air-space-vehicle-ground collaborative sensing according to any one of claims 1-5, characterized in that, Also includes: Monitor the test results and the feedback information from manual review; When the frequency of the detection results indicating disease status exceeds a preset threshold, or when the manual review feedback information confirms a missed report, a data update instruction is automatically generated. In response to the data update command, a drone or satellite is scheduled to retake images of the target mileage segment, and the heterogeneous data spatial index database is updated based on the retaken images.

9. A track health status detection device based on high-speed railway air-space-vehicle-ground collaborative sensing, characterized in that, include: The calling module is used to call the surface deformation interferogram, UAV orthophoto, train vibration data and track circuit monitoring data corresponding to the target mileage segment in the detection mode from a pre-built heterogeneous data spatial index database based on the detection mode. The heterogeneous data spatial index database is a relational database with the line mileage segment as the unique primary key. The extraction module is used to extract static environmental features of the target mileage section based on the surface deformation interferogram and the UAV orthophoto, and to extract dynamic response features based on the train vibration data and track circuit monitoring data. The fusion module is used to perform cross-attention operation with the dynamic response features as query vectors and the static environment features as key vectors and value vectors to fuse and obtain global orbital state features. The detection module is used to input the global track state features into the classifier to obtain the track health status detection result of the target mileage segment as healthy status, potential disease status, or disease status.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the track health status detection method based on high-speed railway air-space-vehicle-ground collaborative perception as described in any one of claims 1 to 6.

Images