Turnout beam real-time vibration response monitoring method and system based on neural network

Abstract

The application provides a kind of neural network-based turnout beam real-time vibration response monitoring method and system, by obtaining the original vibration response signal flow of turnout beam key section vibration sensing network;Original signal flow is reconstructed in time and space, mapped to the three-dimensional tensor space aligned with turnout beam physical coordinates, to obtain vibration response space-time distribution tensor;The space-time distribution tensor is input into the train load back propagation neural network, and the dynamic load transfer path mode and vibration energy attenuation topology are generated by load feature coding and time series propagation chain processing;Call beam body state evolution neural network to carry out multi-step state deduction, obtain beam body state evolution sequence and extract abnormal propagation path and structure weak link positioning;Accordingly, monitoring instruction stream is pushed to trackside monitoring and early warning terminal.The method realizes the accurate space-time modeling and future state deduction of turnout beam vibration response, effectively improves the early warning timeliness and positioning accuracy of turnout beam structure health monitoring.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically, to a method and system for real-time vibration response monitoring of turnout beams based on neural networks. Background Technology

[0002] With the development of railway transportation towards high speed and heavy load, turnout beams, as key load-bearing structures for track turning, directly affect structural safety and track stability due to their vibration response characteristics when trains pass. Real-time vibration response monitoring technology for turnout beams analyzes dynamic response signals collected by sensors at key sections to understand the real-time vibration state under train loads. Currently, time-domain statistics or frequency-domain spectral analysis are typically performed on the vibration acceleration signals of each sensor node to extract vibration amplitude or dominant frequency components and compare them with preset thresholds to determine if anomalies exist. This approach focuses on the instantaneous response intensity of a single point or isolated section, failing to reveal the transmission path of train loads within the beam and the distribution law of vibration energy attenuation at the overall structural level. This results in a lack of effective perception of the spatial propagation and evolution trend of vibration response, limiting the ability of monitoring methods to predict early changes in structural state. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide at least one method and system for real-time vibration response monitoring of turnout beams based on neural networks.

[0004] According to one aspect of the present invention, a method for real-time vibration response monitoring of turnout beams based on neural networks is provided, the method comprising: The original vibration response signal streams collected by each sensing node in the vibration sensing network deployed on the key section of the turnout beam structure are obtained. The original vibration response signal streams contain a continuous vibration acceleration signal sequence carrying timestamps that reflects the dynamic response of the turnout beam structure when the train passes. The original vibration response signal stream is reconstructed in a spatiotemporal manner, and the continuous vibration acceleration signal sequence is mapped to a three-dimensional tensor representation space aligned with the physical coordinates of the turnout beam. This results in a spatiotemporal distribution tensor of the turnout beam vibration response with three coordinate axes: time axis, first spatial distribution axis, and second spatial distribution axis. The element values ​​of the spatiotemporal distribution tensor of the turnout beam vibration response at any time and spatial coordinate represent the vibration acceleration amplitude at that moment and location. The spatiotemporal distribution tensor of the vibration response of the turnout beam is input into a pre-constructed train load backpropagation neural network. The train load backpropagation neural network performs load feature propagation and structural response mode generation on the spatiotemporal distribution tensor of the vibration response of the turnout beam, and obtains the dynamic load transfer path mode and vibration energy attenuation topology map describing the train wheelset at each key section when passing through the turnout beam. Based on the dynamic load transfer path pattern and vibration energy attenuation topology, a pre-constructed beam state evolution neural network is invoked to perform multi-step deduction of the vibration response propagation law of the turnout beam structure, thereby obtaining the beam state evolution sequence that characterizes the vibration response development trend of the turnout beam in the future period, and extracting the abnormal propagation path of vibration response and the location information of structural weak links based on the beam state evolution sequence. Based on the abnormal propagation path of vibration response and the location information of weak points in the structure, a real-time monitoring command stream for the turnout beam is generated, which includes the monitoring section identification and vibration response trend prediction description. The real-time monitoring command stream for the turnout beam is then pushed to the trackside monitoring and early warning terminal.

[0005] According to another aspect of the present invention, a computer system is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method described above.

[0006] This invention reconstructs the raw vibration response signal stream collected by a sensor network at key sections of a turnout beam into a spatiotemporal distribution tensor aligned with physical coordinates, thus providing a unified, structured representation of discrete sensor signals. The spatiotemporal distribution tensor is input into a train load backpropagation neural network, which performs load feature propagation and structural response pattern generation. This allows for the analysis of dynamic load transfer path patterns and vibration energy attenuation topology from the global vibration response, explicitly describing the load propagation mechanism and energy dissipation law. Based on this, a beam state evolution neural network is invoked to perform multi-step deduction of the vibration response propagation law, generating a beam state evolution sequence characterizing the future vibration response development trend. It also extracts information on abnormal propagation paths and structural weak points, achieving a shift from passive perception to proactive prediction. Finally, a monitoring command stream containing monitoring section identifiers and trend prediction descriptions is generated and pushed to a trackside early warning terminal, completing automated continuous tracking and targeted early warning. Because the entire process relies on neural networks to learn the inherent laws of data autonomously, there is no need to pre-set simplified physical models. It can still maintain stable monitoring performance when facing complex working conditions and individual structural differences, effectively improving the ability to identify early anomalies and the reliability of monitoring results.

[0007] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present invention. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of an application scenario provided by the present invention; Figure 2 This is a flowchart illustrating a real-time vibration response monitoring method for turnout beams based on neural networks provided by the present invention. Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] To facilitate a clearer understanding of this invention, we will first introduce the application scenarios of the neural network-based real-time vibration response monitoring method for turnout beams, such as... Figure 1 As shown, the application scenario of this invention includes a computer system 10 and a vibration sensing network. The vibration sensing network may include one or more sensing nodes; the number of sensing nodes is not limited here. Figure 1 As shown, the vibration sensing network may specifically include sensing node 1, sensing node 2, ..., sensing node n; it can be understood that sensing node 1, sensing node 2, sensing node 3, ..., sensing node n can all be network connected to computer system 10 so that each sensing node can interact with computer system 10 through network connection.

[0012] It is understood that computer system 10 may refer to a device that executes the real-time vibration response monitoring method for turnout beams based on neural networks provided in the embodiments of the present invention. The computer system 10 may be a server, such as a single physical server, or a server cluster or distributed system consisting of at least two physical servers.

[0013] Further, please see Figure 2 This is a flowchart illustrating a real-time vibration response monitoring method for turnout beams based on neural networks, provided in an embodiment of the present invention. Figure 2 As shown, this method can be derived from... Figure 1 The method for real-time vibration response monitoring of turnout beams based on neural networks is executed by a computer system 10, and may include the following steps: Step S100: Obtain the original vibration response signal stream collected by each sensing node in the vibration sensing network deployed on the key section of the turnout beam structure. The original vibration response signal stream contains a continuous vibration acceleration signal sequence carrying a timestamp and reflecting the dynamic response of the turnout beam structure when the train passes.

[0014] As a key load-bearing structure in rail transit systems enabling train track switching, the turnout beam exhibits significant nonlinear time-varying stress characteristics when train wheelsets pass over it. The key cross-sections of the turnout beam structure refer to the control cross-sectional locations that generate the maximum dynamic displacement or stress response under train load, pre-calibrated based on structural dynamics simulation analysis results. Specifically, these include the switch rail head width section, the frog rail width section, the long rail fastener support section, and the beam end support section. The vibration sensing network consists of multiple sensing nodes distributed at the key cross-sections of each turnout beam structure. Each sensing node integrates at least one triaxial microelectromechanical system (MEMS) accelerometer unit to synchronously acquire vibration acceleration response components along the longitudinal, transverse, and vertical orthogonal directions of the turnout beam. Each sensing node has a built-in high-precision clock synchronization circuit. By receiving timing signals from the Global Navigation Satellite System or network timing messages following a precise time protocol, it continuously calibrates its local clock crystal oscillator, ensuring that every vibration acceleration sample value acquired by each sensing node is timestamped and synchronized with the same absolute time reference. The sampling frequency of the sensor nodes is set to a value that satisfies the Nyquist sampling theorem and can completely reproduce the broadband vibration response characteristics of the turnout beam structure under high-frequency wheel-rail impact excitation when a train passes. When the train enters the turnout beam section, the wheelset passes through the switch, guide curve, and frog areas in sequence. The dynamic interaction force within the wheel-rail contact patch is transmitted to the turnout beam body through the rail and fastening system, exciting the turnout beam structure to produce a complex dynamic response process containing forced vibration components and freely decaying vibration components. During this period, each sensor node continuously senses and outputs a continuous vibration acceleration signal sequence reflecting the motion state of the turnout beam structure at each sampling instant. The signal streams output by all sensor nodes are converged through an industrial Ethernet ring network to jointly form the original vibration response signal stream carrying a unified timestamp.

[0015] Step S200: Perform spatiotemporal reconstruction of the original vibration response signal stream, map the continuous vibration acceleration signal sequence to a three-dimensional tensor representation space aligned with the physical coordinates of the turnout beam, and obtain the spatiotemporal distribution tensor of the turnout beam vibration response with three coordinate axes: time axis, first spatial distribution axis, and second spatial distribution axis. The element values ​​of the spatiotemporal distribution tensor of the turnout beam vibration response at any time coordinate and spatial coordinate represent the vibration acceleration amplitude at that moment and location.

[0016] In one implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Analyze the continuous vibration acceleration signal sequence collected by each sensor node in the original vibration response signal stream, and extract the vibration acceleration signal segment and the timestamp sequence corresponding to the vibration acceleration signal segment within the continuous time window of each sensor node.

[0017] Specifically, for the raw vibration response signal stream aggregated from the vibration sensing network, a signal stream parsing operation is first performed. The raw vibration response signal stream is transmitted in binary encoding at the physical layer, uses Ethernet frame encapsulation at the data link layer, and reassembles data packets according to the User Datagram Protocol (UDP) at the transport layer. The parsing process includes stripping the header information of the data packets, identifying the sensor node identifier in the load field, and separating and extracting the timestamp field and the sampled value field. Based on the unique hardware address of each sensor node, the interleaved multi-channel continuous vibration acceleration signal sequence is demultiplexed into an independent single-channel signal sequence corresponding to each sensor node. Subsequently, for each single-channel signal sequence, a continuous time window of a preset length is extracted along the positive and negative directions of the time axis, with the event triggering time of the train passing through the turnout beam as the origin of the time axis. The start time of this continuous time window covers the environmental vibration background acquisition stage before the front end of the train enters the turnout beam section, and the end time covers the stage after the rear end of the train has completely left the turnout beam section, when the free decay vibration of the turnout beam structure has basically subsided. The vibration acceleration sample values ​​extracted within this continuous time window constitute a vibration acceleration signal segment, which, together with the timestamps corresponding to each sampling point within the vibration acceleration signal segment, forms a timestamp sequence. Each element in the timestamp sequence is a high-precision time value, used as a reference for subsequent time-domain alignment operations.

[0018] Step S220: Based on the preset physical coordinate information of each sensing node on the key section of the turnout beam structure, construct a three-dimensional spatial coordinate system with the longitudinal mileage, lateral offset and vertical elevation of the turnout beam as coordinate axes, and map the physical coordinate information of each sensing node into the three-dimensional spatial coordinate system.

[0019] The calibration of the physical coordinates of the turnout beam is completed during the vibration sensor network deployment phase. For each sensor node, its actual spatial position on the turnout beam structure is measured using a total station or laser tracker during installation, recording three coordinate components: the longitudinal mileage value measured from a preset mileage starting point along the turnout beam's centerline, the lateral offset value measured outward from the centerline in a horizontal direction perpendicular to the centerline, and the vertical elevation value measured from the rail top surface elevation reference. These three coordinate components constitute the preset physical coordinate information of the sensor node. A right-handed spiral orthogonal three-dimensional spatial coordinate system is established, with the longitudinal mileage of the turnout beam as the first spatial axis, the lateral offset as the second spatial axis, and the vertical elevation as the third spatial axis. The origin of the three-dimensional spatial coordinate system can be selected as the intersection of the centerline of the turnout beam's starting mileage section and the rail top surface. The preset physical coordinate information of each sensing node is mapped one by one to the corresponding coordinate point in the three-dimensional spatial coordinate system, so that each sensing node obtains a three-dimensional spatial coordinate vector composed of three ordered values. This three-dimensional spatial coordinate vector accurately describes the installation position of the sensing node in the physical space of the turnout beam.

[0020] Step S230: Perform time-domain synchronization alignment on the vibration acceleration signal segments corresponding to each sensing node mapped to the three-dimensional spatial coordinate system. Taking the initial triggering time of the train passing through the turnout beam as the alignment benchmark, map the vibration acceleration amplitudes of different sensing nodes at the same time point to a unified time-domain alignment framework to obtain a set of time-domain aligned vibration acceleration synchronization sequences.

[0021] The method for determining the initial trigger time of the train passing the turnout beam is as follows: A pair of wheel axle sensors are installed on the track ahead of the turnout beam section. When the first wheelset of the train passes the wheel axle sensor, a pulse signal is triggered. The rising edge of this pulse signal is recorded as the initial trigger time of the train passing the turnout beam and distributed to all sensor nodes as the alignment reference time through the time synchronization mechanism of the vibration sensor network. Since the local clocks of each sensor node may have slight clock drift or sampling start deviation, the sampling time points of the vibration acceleration signal segments collected by each node are not strictly consistent on the absolute time axis. Therefore, time-domain resampling and alignment operations are required. Specifically, using the alignment reference time as the zero point of the time-domain alignment framework, a set of standard time grid points with a uniform sampling interval is defined. For each sensor node's vibration acceleration signal segment and its corresponding timestamp sequence, the estimated value of the vibration acceleration amplitude is recalculated on the standard time grid points using cubic spline interpolation, thereby synchronously mapping the sampling values ​​of each sensor node, originally located at discrete non-uniform time points, to a unified standard time grid. After the above processing, each sensor node outputs a time-domain aligned vibration acceleration synchronization sequence defined at the same standard time grid point. The time-domain aligned vibration acceleration synchronization sequences of all sensor nodes together constitute a time-domain aligned vibration acceleration synchronization sequence set. In this set, any time index points to the vibration acceleration amplitude of all sensor nodes at the same absolute time point.

[0022] Step S240: Using the time-domain aligned vibration acceleration synchronization sequence set and the spatial coordinate information of each sensing node in the three-dimensional spatial coordinate system, construct a three-dimensional data container with the time dimension as the first axis, the first spatial distribution dimension of the sensing nodes as the second axis, and the second spatial distribution dimension of the sensing nodes as the third axis. Fill the corresponding grid positions of the three-dimensional data container with the vibration acceleration amplitude values ​​on each time-domain slice in the time-domain aligned vibration acceleration synchronization sequence set according to the corresponding first spatial distribution coordinates and second spatial distribution coordinates.

[0023] In one implementation, step S240 may specifically include the following steps S241 to S247: Step S241: Fill the vibration acceleration amplitude values ​​on each time-domain slice in the time-domain aligned vibration acceleration synchronization sequence set into the corresponding grid positions of the three-dimensional data container according to the corresponding first spatial distribution coordinates and second spatial distribution coordinates, so that the element value of the three-dimensional data container at any coordinate represents the vibration acceleration amplitude at the corresponding time point and spatial position.

[0024] The three-dimensional data container is a multi-dimensional array data structure stored contiguously in computer memory. Its index dimensions correspond to the time dimension, the first distribution dimension of the sensor node space, and the second distribution dimension of the sensor node space. The first distribution dimension of the sensor node space corresponds to the lateral offset coordinate direction in the three-dimensional spatial coordinate system, and the second distribution dimension of the sensor node space corresponds to the vertical elevation coordinate direction in the three-dimensional spatial coordinate system. In the time-domain aligned vibration acceleration synchronization sequence set, each standard time grid point corresponds to a time-domain slice, which contains the vibration acceleration amplitude of all sensor nodes at that time point. For the current time-domain slice, the vibration acceleration amplitude of each sensor node in the slice is traversed. Based on the lateral offset coordinate and vertical elevation coordinate of the sensor node in the three-dimensional spatial coordinate system, its grid index position in the latter two dimensions of the three-dimensional data container is determined, and the vibration acceleration amplitude is assigned to that grid position. After sequentially filling all time-domain slices, the element value stored in the three-dimensional data container at the coordinate position determined by any time index, any first spatial distribution index, and any second spatial distribution index is the vibration acceleration amplitude at the corresponding time point and spatial position.

[0025] Step S242: Traverse the time-domain aligned vibration acceleration synchronization sequence set, obtain the vibration acceleration amplitude and corresponding spatial coordinate information of each sensing node in each time-domain slice, and the time-domain slice corresponds to a single sampling moment in the process of the train passing through the turnout beam.

[0026] The time-domain aligned vibration acceleration synchronization sequence set can be understood as a two-dimensional data table structure with sensor nodes as rows and standard time grid points as columns. This two-dimensional data table structure is traversed column by column, and each traversal extracts a column vector corresponding to a standard time grid point. This column vector contains the vibration acceleration amplitude of all sensor nodes at that single sampling moment. Simultaneously, the lateral offset coordinates and vertical elevation coordinates of each sensor node in the three-dimensional spatial coordinate system are read from a pre-stored sensor node attribute table. The vibration acceleration amplitudes are then associated and bound with the spatial coordinate information, forming a list of spatial amplitude tuples corresponding to the current time-domain slice. Each time-domain slice uniquely corresponds to a single sampling moment in the process of a train passing through a turnout beam, and the time progression direction of the slice sequence is logically consistent with the projection of the train's travel direction in the time dimension.

[0027] Step S243: For each time domain slice, create a two-dimensional amplitude matrix with the row direction as the spatial distribution dimension of the sensor nodes and the column direction as the orthogonal direction of the spatial distribution dimension of the sensor nodes. The row index and column index of the two-dimensional amplitude matrix correspond to the lateral offset coordinate and vertical elevation coordinate of each sensor node in the three-dimensional spatial coordinate system, respectively.

[0028] The two-dimensional amplitude matrix is ​​a regular two-dimensional array, with its rows corresponding to the direction of change in the lateral offset coordinates and its columns corresponding to the direction of change in the vertical elevation coordinates. The row indices of the two-dimensional amplitude matrix cover the minimum to maximum values ​​of all sensing nodes in the lateral offset coordinate direction, and are discretized according to a preset lateral grid resolution, so that each row index corresponds to a lateral offset coordinate interval. Similarly, the column indices of the two-dimensional amplitude matrix cover the minimum to maximum values ​​of all sensing nodes in the vertical elevation coordinate direction, and are also discretized according to a preset vertical grid resolution, so that each column index corresponds to a vertical elevation coordinate interval. For each sensing node, its lateral offset coordinate value is used to find the lateral offset coordinate interval it falls into, determining the corresponding row index value; its vertical elevation coordinate value is used to find the vertical elevation coordinate interval it falls into, determining the corresponding column index value. The row and column indices in the two-dimensional amplitude matrix together determine a unique matrix element position, used to accommodate the vibration acceleration amplitude of that sensing node within the current time slice.

[0029] Step S244: Fill the corresponding row and column intersection positions of the two-dimensional amplitude matrix with the vibration acceleration amplitude values ​​of each sensor node in the current time domain slice. For the row and column intersection positions where no sensor nodes are deployed, use a preset empty marker to fill the position, and obtain the spatial amplitude snapshot matrix corresponding to the current time domain slice.

[0030] The filling operation is performed according to the row-column correspondence determined in step S243, assigning the vibration acceleration amplitude of each sensor node within the current time-domain slice to the matrix element specified by the corresponding row and column indices in the two-dimensional amplitude matrix. Since the number of sensor nodes deployed on the cross-section of the turnout beam is limited, most row-column intersections in the two-dimensional amplitude matrix do not correspond to actual physical sensor nodes. For these row-column intersections without deployed sensor nodes, no valid vibration acceleration amplitude is assigned; instead, a predefined missing marker with special identification meaning is filled in. This missing marker is identified as a missing value in subsequent processing, distinguishing it from any possible actual vibration acceleration amplitude. The matrix generated after filling is the spatial amplitude snapshot matrix corresponding to the current time-domain slice. This matrix visually represents a snapshot of the spatial distribution of vibration acceleration amplitudes at various cross-sectional locations of the turnout beam structure at a specific sampling moment during train passage.

[0031] Step S245: According to the time progression of the train passing through the turnout beam, the spatial amplitude snapshot matrices corresponding to each time domain slice are stacked sequentially along the time axis to form an initial three-dimensional data stack structure with time as the first axis and the rows and columns of the two-dimensional spatial amplitude snapshot matrix as the second and third axes.

[0032] The time progression order is determined by the timestamp values ​​of the standard time grid points in ascending order. For each time-domain slice in the time-domain aligned vibration acceleration synchronization sequence set, a corresponding spatial amplitude snapshot matrix is ​​generated according to the processing flow of steps S243 to S244. All spatial amplitude snapshot matrices corresponding to the time-domain slices are arranged sequentially along the newly added third dimension according to the time order of their corresponding standard time grid points. For example, the spatial amplitude snapshot matrix of the earliest moment is placed at the beginning of the stack index, and the spatial amplitude snapshot matrices of subsequent moments are superimposed on it or appended immediately after it. The resulting data structure has three index dimensions: the first index dimension points to different moments on the time axis, the second index dimension points to the row direction of the two-dimensional spatial amplitude snapshot matrix, and the third index dimension points to the column direction of the two-dimensional spatial amplitude snapshot matrix. This data structure is called the initial three-dimensional data stack structure.

[0033] Step S246: Perform tensor dimension normalization on the initial three-dimensional data stack structure, convert the vibration acceleration amplitude of each element in the initial three-dimensional data stack structure into a floating-point representation format with uniform dimensions, and explicitly declare the dimension size of the time axis, spatial row axis, and spatial column axis to obtain the normalized three-dimensional data container.

[0034] The vibration acceleration amplitudes stored in the initial 3D data stack structure may originate from different types of accelerometer units. The ranges of their output electrical signals, after amplification or attenuation by their respective signal conditioning circuits, differ from the reference level, resulting in incompletely unified numerical dimensions of the original sampled values. The primary operation of tensor dimension normalization is to convert the vibration acceleration amplitudes within each element into physical quantity values ​​expressed in SI units of acceleration, based on the sensitivity coefficients and zero-bias calibration parameters of each sensor node, and then store them in double-precision floating-point format. Subsequently, the three dimensions of the initial 3D data stack structure are explicitly declared, i.e., the dimension length values ​​of the time axis, the spatial row axis, and the spatial column axis are recorded in the metadata description field of the data container. The spatial row axis dimension length equals the number of discrete grid divisions in the horizontal offset coordinate direction, and the spatial column axis dimension length equals the number of discrete grid divisions in the vertical elevation coordinate direction. The data structure after dimension unification and explicit dimension declaration is the normalized 3D data container.

[0035] Step S247: Extend the boundary of the normalized three-dimensional data container. Add a virtual sensor node layer with a preset boundary width to the outer edge of the spatial row axis and spatial column axis. Assign the vibration acceleration amplitude of the virtual sensor node layer according to the vibration acceleration amplitude of the adjacent real sensor node and the spatial attenuation law to obtain the three-dimensional data container after boundary extension.

[0036] The purpose of boundary extension is to provide sufficient boundary neighborhood information for subsequent spatial interpolation, avoiding interpolation distortion or edge effects at the spatial edges of the 3D data container. Virtual sensor node columns with a number of columns equal to the preset boundary width are added to the outer edges of both sides of the spatial row axis; virtual sensor node rows with a number of rows equal to the preset boundary width are added to the outer edges of both sides of the spatial column axis. The length of the virtual sensor node layer in the time axis dimension is consistent with the real data. For each grid position within a virtual sensor node layer, the vibration acceleration amplitude is assigned based on the vibration acceleration amplitude of one or more real sensor nodes spatially closest to that virtual grid position at the same time point. A spatial attenuation law is introduced during the assignment process; that is, the vibration acceleration amplitude of the virtual sensor node layer is the result of multiplying the vibration acceleration amplitude of adjacent real sensor nodes by an attenuation factor that monotonically decreases with increasing spatial distance. The attenuation factor can be calculated using a negative exponential function with distance as the independent variable. The parameters of this function are pre-calibrated based on the damping characteristics and wave propagation attenuation characteristics of the turnout beam structural material. After the above boundary extension process, the length of the spatial row axis dimension of the normalized three-dimensional data container is increased by twice the preset boundary width, the length of the spatial column axis dimension is increased by twice the preset boundary width, and the length of the time axis dimension remains unchanged, thus obtaining the boundary-extended three-dimensional data container.

[0037] Step S250: Perform spatial dimension interpolation and temporal dimension smoothing on the filled three-dimensional data container. Use inverse distance weighted interpolation based on the spatial distance of sensor nodes to estimate the vibration acceleration amplitude of the spatial coordinates of the undeployed sensor nodes, and perform moving mean filtering on the vibration acceleration amplitude of adjacent time-domain slices in the temporal dimension to obtain the interpolated and smoothed three-dimensional data container.

[0038] The filled 3D data container is the 3D data container with boundary extension obtained in step S247. The purpose of spatial dimension interpolation is to assign reasonable vibration acceleration amplitude estimates to the un-deployed sensor node grid positions in the container that were originally filled with missing markers, thereby forming a continuous and complete vibration response distribution field in space. The specific execution logic of the inverse distance weighted interpolation method is as follows: For each target grid position that needs to be interpolated in the 3D data container, in the two-dimensional plane formed by the spatial row axis and the spatial column axis, identify all real sensor node positions within a certain search radius around it. Calculate the spatial Euclidean distance between the target grid position and each neighboring real sensor node position. If the spatial Euclidean distance is zero, the vibration acceleration amplitude of the real sensor node is directly used as the interpolation result. If the spatial Euclidean distance is greater than zero, the reciprocal of the spatial Euclidean distance is used as a weighting factor to perform a weighted average of the vibration acceleration amplitudes of each neighboring real sensor node in the current time-domain slice, and the weighted average result is assigned to the target grid position. This interpolation process is performed independently for each time-domain slice. After spatial interpolation, temporal smoothing is performed on the 3D data container along the time axis. Temporal smoothing employs a moving mean filter, which involves defining a sliding window along the time axis encompassing the center time and several adjacent times. The arithmetic mean of the vibration acceleration amplitudes at the same spatial grid position within this window is calculated, and this arithmetic mean replaces the original vibration acceleration amplitude at the center time. The sliding window moves along the time axis moment by moment, repeating the mean replacement operation for all spatial grid positions on each time slice. The data structure resulting from the combined spatial interpolation and temporal smoothing processes constitutes the interpolated and smoothed 3D data container.

[0039] Step S260: Define the interpolated and smoothed three-dimensional data container as the spatiotemporal distribution tensor of the turnout beam vibration response. The spatiotemporal distribution tensor of the turnout beam vibration response has three coordinate axes: time axis, first spatial distribution axis, and second spatial distribution axis. The coordinates on the time axis correspond to a specific time point, and the coordinates on the first spatial distribution axis and the second spatial distribution axis together determine a specific spatial location. The element values ​​of the spatiotemporal distribution tensor of the turnout beam vibration response at any time coordinate and spatial coordinate represent the vibration acceleration amplitude at that specific time point and specific spatial location.

[0040] In one implementation, step S260 may be followed by steps S261 to S266: Step S261: Perform tensor decomposition on the spatiotemporal distribution tensor of the turnout beam vibration response. Slice the spatiotemporal distribution tensor of the turnout beam vibration response along the spatial distribution dimension of the sensing nodes to obtain the single-position vibration response time series vector corresponding to different spatial coordinate positions.

[0041] Tensor decomposition refers to the process of cutting a high-dimensional tensor along one or more dimensions to extract low-dimensional sub-tensors or vectors. For the spatiotemporal distribution tensor of the vibration response of a turnout beam, several spatial grid positions are defined along the first and second spatial distribution axes. The slicing decomposition operation traverses each binary combination formed by the index values ​​of the first and second spatial distribution axes. For each specific spatial grid position, the indices of the first and second spatial distribution axes are fixed, and the vibration acceleration amplitudes at all times along the time axis are extracted. These amplitudes are arranged in chronological order to form a one-dimensional vector, which is the single-position vibration response time series vector corresponding to that specific spatial coordinate position. The length of the single-position vibration response time series vector is equal to the dimension length of the spatiotemporal distribution tensor of the turnout beam vibration response along the time axis, and each element of the vector corresponds to the vibration acceleration amplitude at a specific time point.

[0042] Step S262: Analyze the amplitude fluctuation characteristics of the vibration response time series vector at each single location during the entire time of train passage, extract the peak acceleration occurrence time and peak acceleration amplitude of the vibration response time series vector at each single location, and construct a peak response distribution map indexed by spatial coordinates.

[0043] The entire time interval for train passage refers to the period from the initial triggering moment of the first wheelset to the departure of the last wheelset from the turnout beam section and the free decay vibration of the turnout beam structure essentially ceases. For each single-location vibration response time series vector, all element values ​​are iterated, and the element with the largest absolute value is determined by comparison and sorting. The value corresponding to this largest absolute value element is the peak acceleration amplitude at that spatial coordinate location, and the time corresponding to the position index of this element on the time axis is the moment when the peak acceleration occurs. All spatial grid locations are iterated, and the four fields of the first spatial distribution axis coordinates, the second spatial distribution axis coordinates, the peak acceleration amplitude, and the moment when the peak acceleration occurs for each spatial grid location are combined into a single data record. All data records for all spatial grid locations together constitute the peak response distribution map. The peak response distribution map can be expressed in the form of contour maps, cloud maps, or tables. Indexed by spatial coordinates, it reflects the maximum vibration intensity experienced by various parts of the turnout beam structure during train passage and their temporal sequence.

[0044] Step S263: Based on the chronological order of the peak acceleration occurrence times at adjacent spatial coordinate positions within the peak response distribution map, determine the main transmission direction of the train load on the turnout beam structure and the load transmission speed characterization quantity.

[0045] Adjacent spatial coordinate positions refer to a pair of grid points that are spatially adjacent in the peak response distribution map. For a series of spatial grid positions arranged sequentially along the longitudinal mileage direction of the turnout beam, the chronological order of their peak acceleration occurrence times is compared. If the occurrence times of peak acceleration are successively delayed along the mileage increasing direction, it indicates that the train load is transmitted along that direction. Dividing the longitudinal mileage difference between adjacent spatial grid positions by the corresponding difference in peak acceleration occurrence times yields a value with velocity dimensions, which reflects the average transmission speed of the train load within the interval of adjacent grid positions. The above velocity dimension value is calculated for multiple pairs of adjacent grid positions in the longitudinal direction of the turnout beam, and their statistical median value is obtained. This statistical result is the load transmission speed characterization quantity. The load transmission speed characterization quantity can be used to evaluate the material continuity, stiffness distribution uniformity, and whether there are abnormal wave velocity changes caused by structural damage in the turnout beam structure. The main transmission direction is determined based on the spatial gradient direction at the moment when the peak acceleration occurs. The gradient vector is calculated by finite differences in the longitudinal mileage increase direction, the lateral offset direction, and the vertical elevation direction. The opposite direction of the gradient vector is the main transmission direction of the train load on the turnout beam structure.

[0046] Step S264: Calculate the degree of cross-relation between vibration response time series vectors at different spatial coordinate positions within the spatiotemporal distribution tensor of the turnout beam vibration response. Construct a spatial correlation network for the vibration response based on the degree of cross-relation. In the spatial correlation network, nodes represent the spatial positions of sensing nodes, and the weights of the connecting edges represent the degree of similarity of the vibration response waveforms between nodes.

[0047] The degree of cross-correlation is described using the cross-correlation coefficient. For any two different spatial coordinate positions corresponding to single-position vibration response time series vectors, the normalized cross-correlation coefficient is calculated over the entire time span of train passage. Specifically, the mean of each time series vector is subtracted to obtain a zero-mean sequence. The two zero-mean sequences are multiplied element-wise and summed. Simultaneously, the square root of the sum of squares of the two zero-mean sequences is calculated. The sum is divided by the product of the square roots of the two sums, and the result is the cross-correlation coefficient. The cross-correlation coefficient ranges between -1 and +1. The closer the value is to +1, the more similar the waveforms of the two time series vectors are, indicating that the vibration response modes of the two spatial positions are more consistent during train passage. The cross-correlation coefficient is calculated for each pair of spatial coordinate positions. Each spatial coordinate position is abstracted as a node in the spatial correlation network of vibration response. A connection edge is created between two nodes, and the corresponding cross-correlation coefficient is assigned to the connection edge as its weight. If the cross-correlation coefficient is lower than a preset correlation threshold, the corresponding connection edge can be omitted to form a sparse network. The vibration response spatial correlation network is stored in the form of a graph data structure, where the node list stores the identifier and coordinate information of the spatial location of each sensing node, and the edge list stores the node pairs with connections and the corresponding description of the degree of mutual relationship.

[0048] Step S265: Perform community detection on the vibration response spatial association network, and classify the sensor nodes whose mutual relationship description exceeds the preset association threshold into the same vibration response community. Sensor nodes in the same vibration response community exhibit similar vibration response patterns when the train passes.

[0049] Community detection can be performed using the Louvain algorithm based on modularity optimization. This algorithm takes the node connection edge weight matrix of the vibration response spatial correlation network as input and iteratively optimizes the network's modularity evaluation index to automatically divide the nodes in the network into several subsets with tightly connected internal connections and sparse external connections. Each subset of nodes is a vibration response community. During the partitioning process, the degree of interrelationship is used as the core basis for the edge weights; edges with higher weight values ​​tend to retain the nodes at both ends in the same community. After the algorithm converges, each sensing node in the vibration response spatial correlation network is assigned a community label identifier. Sensing nodes with the same community label identifier are assigned to the same vibration response community. When a train passes over a turnout beam, the waveform envelope, peak distribution, and spectral composition of the vibration response time series vectors of sensing nodes within the same vibration response community show a high degree of consistency, indicating that these spatial locations are in the same structural stress zone or vibration mode dominance region.

[0050] Step S266: Based on the vibration response community division results and the load transfer velocity characterization quantity, the spatiotemporal distribution tensor of the turnout beam vibration response is compressed and encoded to generate a low-dimensional spatiotemporal distribution tensor of the turnout beam vibration response that retains the core vibration response propagation characteristics. The low-dimensional spatiotemporal distribution tensor of the turnout beam vibration response is used as the input data of the train load backpropagation neural network.

[0051] The goal of compression coding is to significantly reduce the data dimensionality of the turnout beam vibration response tensor while preserving as much of the core vibration response propagation characteristics as possible. This is to accommodate the data size limitations of the input layer of the train load backpropagation neural network and improve computational efficiency. The first operation of compression coding is spatial dimensionality reduction based on the vibration response community partitioning results: For all spatial grid locations belonging to the same vibration response community, principal component analysis is used to calculate the first principal component direction of the vibration response time series vectors at all single locations within the community on the plane formed by the first and second spatial distribution axes. The projection coefficients along this principal component direction are used as the representative vibration response time series vector for that community, thus compressing the time series vectors of multiple spatial grid locations within the community into a single representative time series vector. The second operation of compression coding is time dimensionality reduction based on load transfer velocity characteristics: The delay time required for the train load to propagate in the longitudinal adjacent grid intervals of the turnout beam is estimated based on the load transfer velocity characteristics. The minimum sampling interval for time-domain downsampling is determined accordingly. Vibration acceleration amplitudes at certain time points are extracted from the original time series using equal-interval sampling, reducing the time axis dimension length while ensuring no loss of load propagation temporal characteristics. After the aforementioned spatial and temporal dimensionality reduction operations, the generated data structure is a low-dimensional spatiotemporal distribution tensor of the turnout beam vibration response that retains the core vibration response propagation characteristics. The lengths of the three coordinate axes of the low-dimensional turnout beam vibration response spatiotemporal distribution tensor are all significantly smaller than the corresponding lengths of the original turnout beam vibration response spatiotemporal distribution tensor. This low-dimensional turnout beam vibration response spatiotemporal distribution tensor will be used as the input data for the train load backpropagation neural network, passing it into the network's input layer interface.

[0052] Step S300: Input the spatiotemporal distribution tensor of the vibration response of the turnout beam into the pre-constructed train load backpropagation neural network. Perform load feature propagation and structural response mode generation on the spatiotemporal distribution tensor of the vibration response of the turnout beam through the train load backpropagation neural network to obtain the dynamic load transfer path mode and vibration energy attenuation topology map describing the train wheelset at each key section when passing through the turnout beam.

[0053] In one implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Input the spatiotemporal distribution tensor of the turnout beam vibration response into the input layer of the train load backpropagation neural network, perform batch normalization and dimension reshaping operations on the spatiotemporal distribution tensor of the turnout beam vibration response, and convert the spatiotemporal distribution tensor of the turnout beam vibration response into a standard input data stream that conforms to the input specification of the hidden layer of the train load backpropagation neural network.

[0054] The train load backpropagation neural network is a deep convolutional recurrent hybrid architecture neural network. Its input layer includes a batch normalization sublayer and a dimension reshaping sublayer. When the spatiotemporal distribution tensor of the turnout beam vibration response is input into the input layer, its data shape is a three-dimensional array, corresponding to the time dimension, the first spatial distribution dimension, and the second spatial distribution dimension, respectively. The batch normalization operation calculates the mean and standard deviation of the vibration response time series at each spatial grid location along the time dimension of the turnout beam vibration response spatiotemporal distribution tensor. The original vibration acceleration amplitude is subtracted from the mean and then divided by the standard deviation, so that the normalized numerical distribution has the characteristics of zero mean and unit standard deviation. This operation effectively eliminates the problem of baseline drift and scale inconsistency of vibration response amplitude caused by differences in installation location, local structural stiffness, or individual sensor sensitivity among different sensing nodes. The dimension reshaping operation changes the dimensional arrangement order and dimensional size combination of the three-dimensional tensor after batch normalization while keeping the total amount of data unchanged. Specifically, the first and second spatial distribution dimensions are flattened and merged into a single one-dimensional spatial feature channel dimension, forming a two-dimensional matrix format with the time dimension as the first dimension and the spatial feature channels as the second dimension. This two-dimensional matrix format meets the requirements of the input data shape for subsequent convolutional and recurrent layers in the train load backpropagation neural network. The output data after the dimension reshaping operation is the standard input data stream.

[0055] Step S320: The load feature encoder of the train load backpropagation neural network performs multi-level load feature extraction on the standard input data stream. The vibration acceleration amplitude gradient change features between adjacent sensing nodes in the standard input data stream are extracted through cascaded convolutional noise reduction layers to generate a local load action domain feature map that characterizes the local load influence range near the action position of the train wheelset.

[0056] In one implementation, step S320 may specifically include the following steps S321 to S326: Step S321: Apply a first-level temporal convolution kernel array to the standard input data stream. The first-level temporal convolution kernel array performs a sliding convolution operation on the standard input data stream along the time dimension to extract the abrupt rising edge and decay falling edge morphology of the vibration acceleration amplitude during the train passing through the turnout beam within a short time window, generating a primary temporal load feature map containing temporal edge response features.

[0057] The load feature encoder is the front-end feature extraction subnetwork of the train load backpropagation neural network, consisting of multiple cascaded convolutional layers. The first-level temporal convolutional kernel array contains several sets of one-dimensional convolutional kernels, each performing a sliding convolution operation along the time dimension of the standard input data stream. The size of the one-dimensional convolutional kernel is a window covering several consecutive time steps along the time dimension, and the weight parameters of the convolutional kernel are optimized through the backpropagation algorithm during the neural network pre-training process. In the sliding convolution operation, the convolutional kernel is sequentially multiplied element-wise with each time window segment in the standard input data stream, and the results are summed to obtain the convolutional response value of the kernel at the current time position. Different sets of one-dimensional convolutional kernels, with different weight configurations, selectively respond to temporal edge patterns such as fast rising edges, slow rising edges, fast falling edges, and slow falling edges in the vibration acceleration signal. Among them, the weight distribution of some convolutional kernels exhibits an alternating positive and negative pattern, which is sensitive to the first-order difference changes of the signal, thus effectively capturing the rapid increase in vibration acceleration amplitude within a very short time when the train wheelset passes directly under the sensing node; the weight distribution of other convolutional kernels focuses on tracking the envelope change of the signal attenuation segment, extracting the amplitude attenuation rate information of the free vibration attenuation stage. The output responses of all time-domain convolutional kernels are concatenated along the newly added feature channel dimension to form a primary time-domain load feature map.

[0058] Step S322: Apply a second-level spatial convolution kernel array to the primary time-domain load feature map. The second-level spatial convolution kernel array performs a sliding convolution operation on the primary time-domain load feature map along the spatial distribution dimension of the sensing nodes to extract the spatial gradient distribution features and spatial correlation features of the vibration acceleration amplitude between adjacent sensing nodes, and generate a secondary spatial load feature map containing a description of the spatial gradient direction.

[0059] The second-level spatial convolutional kernel array acts on the spatial feature channel dimension of the primary temporal load feature map. In the primary temporal load feature map, adjacent index positions in the spatial feature channel dimension correspond to adjacent sensor node spatial grid positions in the first or second spatial distribution dimension. The spatial convolutional kernel array contains several sets of one-dimensional convolutional kernels, which perform sliding convolution operations along the spatial feature channel dimension. The size of the spatial convolutional kernels is designed to cover several adjacent sensor node positions in space. In the sliding convolution operation, the kernel weight distribution is trained to be sensitive to the spatial difference of the vibration response amplitude of adjacent sensor nodes. The weight arrangement of some spatial convolutional kernels is equivalent to the spatial first derivative operator, and its output response reflects the magnitude and sign of the gradient of the vibration acceleration amplitude along the lateral offset direction or the vertical elevation direction; the weight arrangement of some spatial convolutional kernels is equivalent to the spatial second derivative operator, and its output response reflects the degree of concentration or diffusion of vibration energy in space. After the spatial convolution operation is completed, each time position in the primary temporal load feature map corresponds to a set of feature vectors containing spatial gradient magnitude and spatial gradient direction information. The output responses of all spatial convolution kernels are concatenated along the feature channel dimension and merged or replaced with the original feature channels of the primary temporal load feature map to form the secondary spatial load feature map. The secondary spatial load feature map contains descriptions of the severity of spatial changes in vibration acceleration amplitude between adjacent sensing nodes and the direction of the spatial change trend.

[0060] Step S323: Apply a third-level spatiotemporal joint convolution kernel array to the secondary spatial load feature map. The third-level spatiotemporal joint convolution kernel array simultaneously performs three-dimensional sliding convolution operations on the secondary spatial load feature map in both the time dimension and the spatial distribution dimension of the sensing nodes, extracting the linkage change pattern of the train load in the process of temporal advancement and spatial propagation, and generating a deep spatiotemporal load feature map containing spatiotemporal correlation texture.

[0061] The third-level spatiotemporal joint convolutional kernel array is implemented using a three-dimensional convolutional kernel. The three dimensions of the three-dimensional convolutional kernel correspond to the time dimension, the first spatial distribution dimension, and the second spatial distribution dimension, respectively. In the secondary spatial load feature map, the spatial feature channel dimensions are rearranged to restore a two-dimensional planar layout of the first and second spatial distribution dimensions, which, together with the time dimension, constitutes a three-dimensional data structure. The three-dimensional convolutional kernel slides simultaneously along both the time and spatial axes on this three-dimensional data structure, calculating a weighted sum within the kernel's coverage area at each spatiotemporal location. The weight parameters of the three-dimensional convolutional kernel learn specific spatiotemporal coupling patterns through the training process. For example, there is a response enhancement pattern that propagates forward along the time axis and moves forward along the spatial longitudinal direction. This pattern corresponds to the physical process where the vibration response caused by the train wheelset load moving forward on the turnout beam is delayed in the time domain and propagates in space. Another example is a response pattern that diffuses in the spatial lateral direction and decays along the time axis. This pattern corresponds to the physical process where vibration energy diffuses laterally from the wheel-rail contact point to both sides of the turnout beam and dissipates over time. After the 3D sliding convolution operation is completed, a new set of feature maps is output. Each feature channel in this set of feature maps corresponds to a specific train load temporal propagation and spatial propagation linkage change mode. The output response of all 3D convolution kernels constitutes a deep spatiotemporal load feature map. The deep spatiotemporal load feature map contains rich spatiotemporal correlation textures, including implicit encoding representations of vibration wavefront propagation direction, propagation speed, and energy diffusion range.

[0062] Step S324: Perform feature sparsification on the deep spatiotemporal load feature map, set the feature positions with response intensity lower than the preset activation threshold in the deep spatiotemporal load feature map to zero, retain the significant feature positions with response intensity higher than the preset activation threshold, and generate a sparsified deep spatiotemporal load feature map.

[0063] Feature sparsity is achieved through a nonlinear activation function threshold mechanism. The preset activation threshold is a scalar value adaptively determined based on the statistical distribution of the feature map during neural network training. In each training iteration batch, the mean and standard deviation of all element values ​​within the deep spatiotemporal load feature map are calculated. The preset activation threshold is set as the mean minus the standard deviation multiplied by a fixed coefficient. For each element position in the deep spatiotemporal load feature map, its absolute value is compared with the preset activation threshold. If the absolute value of the element is less than the preset activation threshold, the load feature response intensity corresponding to that feature position is considered weak, possibly due to sensor background noise or environmental micro-vibration interference, and does not contain meaningful train load propagation mode information; therefore, the element value is forcibly modified to zero. If the absolute value of the element is greater than or equal to the preset activation threshold, the feature position is considered to have captured significant load propagation spatiotemporal features, and its original value is retained unchanged. Feature sparsification effectively suppresses low-amplitude noise responses and irrelevant background features in deep spatiotemporal load feature maps, highlighting the significant spatiotemporal texture caused by train wheel-set loads. Simultaneously, it provides a sparser input representation for subsequent network layers, reducing computational complexity. The feature map after feature sparsification is the sparsed deep spatiotemporal load feature map.

[0064] Step S325: Perform feature pyramid pooling on the sparse deep spatiotemporal load feature map. Use a pooling window with a first size to perform downsampling operation on the sparse deep spatiotemporal load feature map to generate a first pooled feature map, and use a pooling window with a second size to perform downsampling operation on the sparse deep spatiotemporal load feature map to generate a second pooled feature map. The first size is larger than the second size. The first pooled feature map represents the overall load distribution trend of the turnout beam, and the second pooled feature map represents the local load concentration phenomenon near the wheelset action point.

[0065] Feature pyramid pooling is a parallel multi-scale pooling structure containing at least two parallel pooling branches. The first pooling branch employs a pooling window of a first size, which covers a large receptive field in both the first and second spatial distribution dimensions, such as covering almost the entire spatial grid region corresponding to the cross-section of the turnout beam. The pooling window of the first size slides on a sparse deep spatiotemporal load feature map, taking the maximum value of all elements within the coverage area of ​​each pooling window as the pooling output for that region. After downsampling through the first-size pooling window, the spatial dimensional resolution of the feature map is significantly reduced, and each pooling output value summarizes the overall load characteristic response intensity over a large spatial range. The resulting first pooling feature map has a relatively coarse spatial texture, but it can clearly reflect the differences in the overall load distribution trend between different longitudinal sections of the turnout beam, such as the comparison of the overall load degree between the front and rear regions of the turnout beam, or the difference in the overall vibration energy level between the switch rail side and the frog rail side. The second pooling branch employs a pooling window of a second size, which covers a smaller local receptive field in the spatial dimension, corresponding only to the local spatial region near a single wheel-rail contact patch. The second-sized pooling window also performs max pooling, but the downsampled feature map retains higher spatial resolution, enabling precise characterization of local load concentration near the wheelset contact point, including the spatial boundary and morphology of the high-amplitude vibration region directly below the wheel-rail contact point. The first and second pooling feature maps have the same number of channels in the feature channel dimension, but differ in their spatial dimensions.

[0066] Step S326: Perform channel concatenation and fusion of the first pooling feature map and the second pooling feature map, and use pointwise convolution to perform cross-channel information interaction and dimensionality reduction on the concatenated feature channels to generate a local load domain feature map that integrates multi-scale load distribution information.

[0067] The channel concatenation and fusion operation first adjusts the first pooling feature map to have the same spatial dimension as the second pooling feature map through upsampling interpolation, ensuring a one-to-one correspondence between the two feature maps on the spatial grid. Then, the resized first and second pooling feature maps are concatenated along the feature channel dimension, arranging all feature channels of the two feature maps sequentially to form a concatenated feature map with the sum of the number of channels in the two feature maps. Next, pointwise convolution is used to process the concatenated feature map. Pointwise convolution refers to a 2D convolution operation with a 1×1 kernel size. This convolution operation performs linear combination and non-linear activation only along the feature channel dimension, without changing the spatial dimension of the feature map. The number of pointwise convolution kernels is set to a target number of channels less than the total number of channels in the concatenated feature map. Each pointwise convolution kernel is responsible for learning a cross-channel information fusion weight combination. During the convolution calculation, each pointwise convolution kernel performs a weighted sum of all channel values ​​at the same spatial location within the concatenated feature map and generates a new feature value through a non-linear activation function. Different pointwise convolutional kernels, due to their varying weight parameters, can extract different fusion information patterns from multi-scale pooling features in the spliced ​​feature map. For example, some pointwise convolutional kernels focus on superimposing and enhancing the large-scale overall load distribution trend with small-scale local load concentration features, highlighting dangerous areas that simultaneously possess a high overall load background and local stress concentration; others focus on calculating the differences between features at different scales to identify locations of abrupt changes in load distribution. The outputs of all pointwise convolutional kernels collectively constitute a local load domain feature map that fuses multi-scale load distribution information. The feature vector of each spatial location in the local load domain feature map comprehensively reflects the macroscopic role of that location in the overall load transfer path and its microscopic response characteristics under local wheel-rail contact action.

[0068] Step S330: Input the local load domain feature map into the time-series load propagation chain processing layer of the train load backpropagation neural network, which includes a recurrent memory unit array that expands along the time dimension. The recurrent memory unit array performs progressive propagation path tracking on the local load domain feature map according to the time causality law of the turnout beam vibration response, and generates a load propagation state sequence describing the gradual transmission of the train wheelset load along the longitudinal direction of the turnout beam.

[0069] The temporal load propagation chain processing layer is a sub-network structure in the train load backpropagation neural network specifically responsible for modeling the temporal dependencies of load transmission. This layer receives the local load domain feature map from the load feature encoder as the input sequence, the length of which is equal to the time dimension of the spatiotemporal distribution tensor of the turnout beam vibration response. The core computational unit of the temporal load propagation chain processing layer is an array of recurrent memory units expanded along the time dimension. This array consists of multiple recurrent memory units with identical structures but shared time step parameters connected sequentially, each corresponding to a time step in the input sequence. The recurrent memory unit uses a gated recurrent unit as its basic computational module. The gated recurrent unit contains two gating mechanisms: an update gate and a reset gate. The update gate controls the proportion of hidden state information from the previous time step retained in the current time step, while the reset gate controls the proportion of hidden state information from the previous time step used in the calculation of the candidate hidden state of the current time step. At each time step, the recurrent memory unit receives a slice of the local load domain feature map of the current time step as an input vector, and simultaneously receives the hidden state vector output from the recurrent memory unit of the previous time step. The input vector and the previous hidden state vector are linearly transformed and then input into the update gate and reset gate, respectively. The gate value is compressed to the 0-1 range using an S-shaped growth curve function. The gate value output by the update gate is multiplied element-wise by the previous hidden state vector, and its complement is multiplied element-wise by the candidate hidden state vector. The two are added together to obtain the hidden state vector at the current time step. The gate value output by the reset gate is multiplied element-wise by the previous hidden state vector, concatenated with the current input vector, and subjected to a linear transformation and hyperbolic tangent activation to obtain the candidate hidden state vector. The circular memory unit array performs the above calculations sequentially in time order, collecting the hidden state vectors output at each time step to form a load propagation state sequence. Since the circular memory unit has the ability to remember historical information, each hidden state vector in the load propagation state sequence not only contains the local load distribution information at the current time step, but also implicitly encodes the historical load transfer path information from the start time to the current time through circular concatenation.

[0070] Step S340: The load propagation state sequence is mapped to a graph structure through the energy attenuation topology reconstruction layer of the train load backpropagation neural network. Each sensor node is regarded as a graph node, and the vibration energy transfer relationship between nodes is regarded as a graph edge. By comparing the differences in the node state values ​​at each time step in the load propagation state sequence, the directed edge weights between adjacent graph nodes are generated. The directed edge weights characterize the vibration energy transfer efficiency of the train load between adjacent sensor nodes.

[0071] In one implementation, step S340 may be followed by steps S341 to S346: Step S341: Extract the graph Laplacian matrix at each time step from the dynamic directed graph. The diagonal elements of the graph Laplacian matrix represent the out-degree or in-degree strength of each graph node, and the off-diagonal elements represent the negative values ​​of the directed edge weights between adjacent graph nodes.

[0072] The dynamic directed graph is a graph sequence data generated by the energy attenuation topology reconstruction layer after performing graph structure mapping on the load propagation state sequence. Each time step corresponds to a directed graph. In each directed graph, the set of graph nodes corresponds to the spatial grid position of the sensing nodes in the spatiotemporal distribution tensor of the turnout beam vibration response. Each graph node has a node state value, which is extracted from the hidden state vector at the corresponding time step in the load propagation state sequence. The set of directed edges consists of connections between adjacent graph nodes. The adjacency relationship is determined based on the adjacency relationship of the sensing nodes in the first and second spatial distribution dimensions. For any pair of adjacent graph nodes, the node state values ​​of the two nodes at the current time are compared. If the node state value of the source node is greater than that of the sink node, a directed edge is established from the source node to the sink node; if the node state value of the source node is less than that of the sink node, a directed edge is established from the sink node to the source node. The weight of a directed edge is calculated by dividing the difference between the state values ​​of the source and sink nodes by the sum of their state values. This ratio reflects the magnitude of the driving potential difference that drives the transfer of vibrational energy from high-response nodes to low-response nodes. For each directed graph, its graph Laplace matrix is ​​constructed. The graph Laplace matrix is ​​a square matrix whose number of rows and columns equals the total number of nodes in the directed graph. The diagonal elements of the graph Laplace matrix are the sum of the out-degree and in-degree of the node. Off-diagonal elements are the negative of the weight of a directed edge if one exists; otherwise, they are zero.

[0073] Step S342: Perform eigenvalue decomposition on the graph Laplace matrix at each time step to calculate the eigenvalue sequence and eigenvector matrix of the graph Laplace matrix at each time step. The eigenvalue sequence reflects the distribution ratio of vibration energy in each vibration mode of the turnout beam structure at the corresponding time step, and the eigenvector matrix reflects the degree of participation of each graph node in the dominant vibration mode at the corresponding time step.

[0074] Eigenvalue decomposition can be performed using the QR iterative algorithm. The calculated eigenvalues ​​are arranged in ascending order to form an eigenvalue sequence, with each eigenvalue corresponding to an eigenvector. All eigenvectors are arranged column-wise to form an eigenvector matrix. In graph signal processing theory, smaller eigenvalues ​​correspond to low-frequency vibration modes in the graph structure, representing a smooth distribution of vibration energy on a larger spatial scale; larger eigenvalues ​​correspond to high-frequency vibration modes in the graph structure, representing local oscillations of vibration energy on a smaller spatial scale. The relative magnitudes of each eigenvalue in the eigenvalue sequence directly reflect the proportion of vibration energy distributed in the corresponding vibration mode to the total energy.

[0075] Step S343: Extract the eigenvector corresponding to the largest eigenvalue in the eigenvalue sequence as the principal mode distribution vector. The magnitude of each element in the principal mode distribution vector reflects the dominant position of the corresponding sensing node in the vibration energy propagation process at the current moment.

[0076] In the physical meaning of graph signal analysis, the eigenvector corresponding to the largest eigenvalue represents the vibration mode with the most dramatic changes and the most concentrated energy in the graph structure, i.e., the dominant vibration mode. The eigenvector corresponding to the largest eigenvalue is extracted separately from the eigenvector matrix. The dimension of this eigenvector is equal to the total number of nodes in the directed graph, and each element in the vector corresponds to a specific sensor node's spatial grid position. This extracted eigenvector is denoted as the dominant mode distribution vector. In the dominant mode distribution vector, the absolute value of each element reflects the participation coefficient of the corresponding sensor node in the dominant vibration mode. The larger the participation coefficient, the more significantly dominant the sensor node plays in the vibration energy propagation process at the current moment. The relative order of the elements in the dominant mode distribution vector provides a priority sequence regarding the importance of each sensor node in the vibration energy propagation process.

[0077] Step S344: Analyze the evolution trajectory of the principal mode distribution vector over time, extract the moment when the relative order of elements in the principal mode distribution vector changes abruptly, and mark the moment when the change occurs as the occurrence time of the load transfer path switching event.

[0078] The time-varying trajectory refers to the process of observing the changes in the values ​​of each element and their relative order over time after arranging the principal modal distribution vectors at multiple consecutive moments in chronological order. For each pair of adjacent moments, the order of elements in the principal modal distribution vector at the previous moment is compared with the order of elements in the principal modal distribution vector at the next moment. The order is descending, with the element with the largest participation coefficient in the first position. When the order of the next moment changes significantly relative to the previous moment, for example, a sensor node that was originally in the first position moves to the third position or later, a load transfer path switching event is determined to have occurred between those adjacent moments. The specific quantitative criterion for a sudden change in order is the decrease in the Kendall rank correlation coefficient. A sudden change is confirmed when the correlation coefficient falls below a preset threshold. The moment when the sudden change occurs is marked as the occurrence time of the load transfer path switching event.

[0079] Step S345: Based on the occurrence time of the load transfer path switching event and the corresponding principal mode distribution vector, perform path level marking on the set of load transfer path description lines in the dynamic load transfer path mode, assign level identifiers to the corresponding load transfer path description lines according to the size sort of each element in the principal mode distribution vector, and determine the direction pointing mark of the load transfer path description lines according to the direction of element sorting change.

[0080] The dynamic load transfer path pattern consists of a series of load transfer path description lines that spatially connect adjacent sensor nodes. During the time interval between two adjacent load transfer path switching events, the main load transfer path remains relatively stable. For each moment of the principal mode distribution vector within this time interval, the sensor node with the largest participation coefficient in the vector is taken as the starting point of the main path. Adjacent sensor nodes with the second largest participation coefficients are then connected sequentially along a spatial direction consistent with the gradient direction of the principal mode distribution vector, forming one or more continuous load transfer path description lines. Based on the ranking number of each sensor node in the principal mode distribution vector, a hierarchical identifier is assigned to the load transfer path description line passing through that sensor node. The path description line containing the first-ranked sensor node is assigned a first-level identifier; the path description line containing the second-ranked sensor node, if different from the first-level path, is assigned a second-level identifier, and so on. The direction pointing mark is determined based on the changing direction of the order of adjacent sensing nodes in the principal mode distribution vector on the path description line. If the order number increases along the extension direction of the path description line, the direction pointing mark is defined as positive, indicating that the vibration energy propagates from the high participation node to the low participation node.

[0081] Step S346: Input the load transfer path description line set carrying hierarchical identifiers and direction pointing marks and the vibration energy attenuation topology map into the output mapping layer of the train load backpropagation neural network. Through the output mapping layer, map the graph node coordinates in the load transfer path description line set carrying hierarchical identifiers and direction pointing marks and the graph node coordinates in the vibration energy attenuation topology map back to the physical coordinate system of the turnout beam, and generate the final dynamic load transfer path pattern and vibration energy attenuation topology map that correspond one-to-one with the key section positions of the turnout beam structure.

[0082] The output mapping layer is the output processing layer in the train load backpropagation neural network responsible for mapping the graphical representation back to the original physical space coordinate system. Internally, this layer maintains a coordinate mapping lookup table, which records the correspondence between each graphical node index and its longitudinal mileage coordinates, lateral offset coordinates, and vertical elevation coordinates in the three-dimensional spatial coordinate system of the turnout beam. For each path description line in the load transfer path description line set, its contained graphical node index sequence is traversed. The coordinate mapping lookup table converts the graphical node index sequence into a sequence of spatial coordinate points in the physical coordinate system of the turnout beam, and adjacent spatial coordinate points are connected by straight line segments to form a physically locatable load transfer path geometric description line. Simultaneously, the hierarchical identifier and direction indicator carried by the path description line are attached as attribute information to the corresponding geometric description line. The hierarchical identifier is represented by line width or color coding, and the direction indicator is represented by an arrow symbol attached to the geometric description line. For the vibration energy attenuation topology diagram, the coordinates of the graph nodes are also mapped to the spatial positions in the physical coordinate system of the turnout beam through a coordinate mapping lookup table. The weights of the directed edges of the graph are mapped to the connecting lines between spatial positions and the energy transfer efficiency represented by the color intensity.

[0083] Step S350: Perform topological feature aggregation on the dynamic directed graph composed of graph nodes and directed edge weights, perform graph convolution operation on the dynamic directed graph at each time along the time dimension, extract the importance ranking information of graph nodes and the activity change trend of graph edges at different times, and generate dynamic load transfer path pattern and vibration energy attenuation topology graph that evolves over time.

[0084] Topological feature aggregation is performed on graph sequence data, using spatial graph convolution. For each time-series dynamic directed graph, the graph convolutional layer receives the current state feature vector of the graph node and the graph's adjacency matrix as input. The adjacency matrix is ​​filled with off-diagonal positions by directed edge weights, with self-loop edge weights set to 1. The graph convolution operation aggregates the state feature information of its neighboring graph nodes for each graph node. The aggregation method involves symmetrically normalizing the adjacency matrix, multiplying it by the node state feature matrix, and then performing a linear transformation and non-linear activation through a learnable weight matrix. After stacking several graph convolutional layers, the feature vector output by each graph node integrates the topological information and node state information of its multi-order neighborhood. The magnitude of the graph node feature vector output by the graph convolution is calculated; a larger magnitude indicates a higher importance ranking for the graph node at the current time step. The changes in edge weights output by the edge attention mechanism module in the graph convolutional layer are analyzed. The rate of change of weights of the same directed edge between adjacent time steps is calculated. Edges with a rate of change exceeding a preset rate of change threshold are marked as active edges. The changes in the number and distribution of active edges over time constitute the trend of edge activity in the graph.

[0085] Step S360: Visualize and encode the dynamic load transfer path pattern and the vibration energy attenuation topology map to generate a set of load transfer path description lines with the physical coordinates of the turnout beam as the reference system and a set of energy attenuation layer description lines with the sensor nodes as the reference positions. The set of load transfer path description lines is used to describe the dynamic load transfer path pattern at each key section, and the set of energy attenuation layer description lines is used to describe the vibration energy attenuation topology map at each key section.

[0086] Visualization encoding conversion is the process of transforming abstract data structures into a graphical description language that conforms to engineering visualization specifications. For a set of load transfer path description lines, the encoding conversion rules are as follows: each load transfer path description line is defined by a starting point coordinate, an ending point coordinate, and several intermediate control point coordinates. The coordinate values ​​are represented by a triplet of longitudinal mileage, lateral offset, and vertical elevation in the physical coordinate system of the turnout beam. On the visualization terminal, the load transfer path description line is presented as a continuous broken line segment with arrows. The arrow direction is consistent with the load transfer direction marker, and the line thickness is inversely proportional to the layer identifier. For a set of energy attenuation layer description lines, with each sensor node as the reference center point and all adjacent sensor nodes connected by its directed edges as the radiation direction, a gradient color line extending outward from the center point is drawn. The intensity of the gradient color encodes the vibration energy transfer efficiency represented by the directed edge weight in that direction; the darker the color, the higher the vibration energy transfer efficiency. The energy attenuation layer description lines are divided into several layer intervals based on the numerical values ​​of the directed edge weights. Description lines within the same layer interval use the same color coding, while different layer intervals exhibit a stepped color transition. The load transfer path description line set and the energy attenuation layer description line set can be overlaid on the 3D model or 2D projection plan of the turnout beam structure.

[0087] Step S400: Based on the dynamic load transfer path pattern and vibration energy attenuation topology, the pre-constructed beam state evolution neural network is invoked to perform multi-step deduction of the vibration response propagation law of the turnout beam structure, so as to obtain the beam state evolution sequence that characterizes the vibration response development trend of the turnout beam in the future period, and extract the abnormal propagation path of vibration response and the location information of weak links in the structure based on the beam state evolution sequence.

[0088] In one implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Convert the dynamic load transfer path pattern into a load transfer state sequence represented by a graph node state vector sequence, convert the vibration energy attenuation topology graph into a vibration energy attenuation relation matrix represented by an adjacency matrix, and use the load transfer state sequence and the vibration energy attenuation relation matrix as the initial input data of the beam state evolution neural network.

[0089] In the dynamic load transfer path model, each time step corresponds to a directed graph. Each node in this directed graph is associated with a node state value extracted from the corresponding time step in the load propagation state sequence. The node state values ​​of all nodes are arranged into a column vector according to a fixed node index order, which is the node state vector for that time step. Arranging the node state vectors for all time steps in chronological order constitutes the load transfer state sequence. Similarly, the vibration energy attenuation topology graph is based on the directed graph at each time step, and the adjacency matrix of this directed graph is extracted as the vibration energy attenuation relation matrix. The element values ​​in the vibration energy attenuation relation matrix reflect the attenuation ratio of vibration energy during the transfer from the source node to the sink node. Before being input into the beam state evolution neural network, the load transfer state sequence and the vibration energy attenuation relation matrix need to undergo missing value imputation and numerical normalization preprocessing.

[0090] Step S420: Through the state initialization layer of the beam state evolution neural network, a coupling mapping operation is performed on the initial moment state vector of the load transfer state sequence and the vibration energy attenuation relationship matrix. The load state values ​​of each graph node in the initial moment state vector are distributed to the adjacent graph nodes according to the attenuation ratio defined in the vibration energy attenuation relationship matrix, thereby generating an initial state transition vector that characterizes the vibration response distribution prediction of the turnout beam structure at the next moment.

[0091] The state initialization layer of the beam state evolution neural network receives the graph node state vectors corresponding to the initial time step in the load transfer state sequence and the vibration energy attenuation matrix at the initial time step as input. The computational logic of the coupling mapping operation is as follows: For each graph node in the initial time step state vector, its source value is multiplied by the attenuation ratio coefficient in the vibration energy attenuation matrix, where the graph node is the source node and each adjacent graph node is the sink node, to obtain the energy component transferred from the source node to each adjacent sink node. Each graph node is then treated as a sink node, and the energy components transferred to it from all source nodes are accumulated, along with the untransferred energy portion retained by the graph node itself, to obtain the predicted node state value for the next time step. All the predicted node state values ​​for the next time step are arranged in the same graph node index order as the initial time step state vector to form the initial state transition vector.

[0092] Step S430: Input the initial state transition vector into the recurrent evolution unit chain of the beam state evolution neural network. The recurrent evolution unit chain consists of multiple gated recurrent units arranged in series. Each gated recurrent unit receives the hidden state vector of the previous time step and the input state vector of the current time step. Through the internal gating mechanism, it performs selective forgetting and selective updating operations on the historical state information and the current input information to generate the hidden state vector of the current time step and pass it to the next gated recurrent unit.

[0093] The recurrent evolutionary unit chain is the core recursive inference component of the beam state evolution neural network, consisting of several gated recurrent units with identical structures stacked in series along the time step direction. The computational process of each gated recurrent unit within a single time step includes the following sub-steps: First, the hidden state vector from the previous time step is concatenated with the input state vector from the current time step along the feature dimension to form a concatenated vector. This concatenated vector is then passed through three independent linear transformation layers to generate the vector before activation by the update gate, the vector before activation by the reset gate, and the vector before activation by the candidate hidden state. The vectors before activation by the update gate and the vector before activation by the reset gate are activated to the 0-1 interval using an sigmoid growth curve function. The reset gate vector is then multiplied element-wise by the hidden state vector from the previous time step, and the product is concatenated again with the input state vector from the current time step. After passing through a fourth linear transformation layer, it is activated by a hyperbolic tangent function to obtain the candidate hidden state vector. Finally, the updated gate vector is multiplied element-wise with the hidden state vector of the previous time step, and the complement of the updated gate vector is multiplied element-wise with the candidate hidden state vector. The result of the element-wise addition of the two is the hidden state vector at the current time step.

[0094] Step S440: In each iteration of the cyclic evolution unit chain, the beam state evolution neural network dynamically adjusts the vibration energy attenuation relation matrix according to the hidden state vector output at the previous moment. Based on the relative magnitude of the state values ​​of each graph node in the hidden state vector, it enhances the attenuation weight of the corresponding edge of the graph node with the higher state value and weakens the attenuation weight of the corresponding edge of the graph node with the lower state value, thereby generating a dynamic attenuation relation matrix that is adapted to the current vibration response state of the beam.

[0095] In one implementation, step S440 may specifically include the following steps S441 to S446: Step S441: Analyze the hidden state vector output from the previous moment, extract the state values ​​corresponding to each graph node in the hidden state vector, and the state values ​​represent the vibration response intensity of the corresponding sensing node position at the previous moment.

[0096] The hidden state vector is a vector with the same dimension as the total number of graph nodes, where each element corresponds one-to-one with a specific graph node index. The extraction operation involves reading the value of the corresponding element in the hidden state vector sequentially according to the graph node indices, and assigning this value as the state value of the corresponding graph node. The magnitude of the state value is directly proportional to the predicted amplitude of the vibration acceleration response at the corresponding sensor node location at the previous time step.

[0097] Step S442: Calculate the proportion of each graph node's state value to the total number of graph node state values. Mark graph nodes with proportions exceeding a preset proportion threshold as high-response dominant nodes and graph nodes with proportions below the preset proportion threshold as low-response subordinate nodes.

[0098] When calculating the proportional descriptor, the state value of a single graph node is divided by the sum of the state values ​​of all graph nodes. The quotient is the proportional descriptor of that node. The sum of the proportional descriptors of all graph nodes equals 1. The preset proportional threshold is a boundary determined based on the upper quartiles of the proportional descriptors of all graph nodes. The proportional descriptor of each graph node is compared with the preset proportional threshold; nodes exceeding the threshold are marked as high-response dominant nodes, and nodes below or equal to the threshold are marked as low-response subordinate nodes.

[0099] Step S443: For the set of incoming edges pointing to the dominant node with high response in the vibration energy attenuation relationship matrix, increase the attenuation weight of each edge in the set of incoming edges proportionally according to the proportion of the source node state value to the total state value of all nodes. The larger the proportion, the greater the increase in attenuation weight.

[0100] The incoming edge set refers to the set of all directed edges with a high-response dominant node as the sink node. For each directed edge in the incoming edge set, its source node is identified and its proportional descriptor is obtained. The original decay weight is multiplied by a boosting factor greater than 1, which is positively correlated with the proportional descriptor of the source node.

[0101] Step S444: For the set of outgoing edges in the vibration energy attenuation relationship matrix that points from the high-response dominant node to the low-response subordinate node, maintain the original attenuation weight of each edge in the outgoing edge set unchanged or perform a slight downward adjustment.

[0102] The outgoing edge set refers to the set of all directed edges with high-response dominant nodes as source nodes and low-response subordinate nodes as sink nodes. For each directed edge in this outgoing edge set, the strategy is to maintain the original decay weight unchanged, or to multiply the original decay weight by a fine-tuning coefficient slightly less than 1 and slightly reduce it.

[0103] Step S445: For the set of edges connecting two low-response subordinate nodes in the vibration energy attenuation relation matrix, perform a proportional reduction operation on the attenuation weight of the edges based on the product of the proportion of the state values ​​of the two low-response subordinate nodes to the total state values ​​of all nodes.

[0104] Calculate the proportional descriptors of the two low-response subordinate nodes connected by the edge, multiply the two proportional descriptors, and obtain a smaller value than the individual proportional descriptors. Use this product as a reduction factor, and multiply the original attenuation weight by this reduction factor to obtain the reduced attenuation weight. This operation effectively suppresses redundant cyclic propagation of vibration energy in the low-response region.

[0105] Step S446: The vibration energy attenuation relation matrix after the above differential adjustment operation is used as the dynamic attenuation relation matrix, and the dynamic attenuation relation matrix is ​​used in the hidden state vector update calculation process of the next gated loop unit in the current iteration step.

[0106] Step S450: Input the hidden state vector and dynamic decay relationship matrix output by the cyclic evolution unit chain at the current moment into the state decoder of the beam state evolution neural network. The state decoder converts the hidden state vector into a spatial state distribution map with the same dimension as the original sensor node spatial distribution through deconvolution mapping operation. The spatial state distribution map represents the predicted vibration acceleration amplitude distribution of each sensor node position of the turnout beam at the corresponding future moment.

[0107] The state decoder's network structure employs a deconvolutional architecture symmetrical to the payload feature encoder. The deconvolutional mapping operation comprises several cascaded deconvolutional layers, each consisting of an upsampling operation and a learnable transposed convolutional kernel. The hidden state vector is first reshaped into a low-resolution feature map corresponding to the spatial distribution dimension of the original sensor nodes. The dynamic decay relation matrix is ​​transformed into a graph topology prior feature map with the same number of channels as the low-resolution feature map through a graph embedding layer, and then concatenated with the low-resolution feature map along the channel dimension. The concatenated feature map is then passed through multiple deconvolutional layers, with the spatial resolution progressively increasing until it is finally restored to a spatial state distribution map that perfectly matches the spatial distribution dimension of the original sensor nodes.

[0108] Step S460: Arrange the spatial state distribution maps output at each moment during the continuous multi-step simulation in chronological order to form a beam state evolution sequence that characterizes the development trend of the vibration response of the turnout beam in the future time period. Each frame of the spatial state distribution map in the beam state evolution sequence corresponds to the vibration response prediction state at a future moment.

[0109] As one implementation method, step S400, which extracts the vibration response anomaly propagation path and structural weak point location information based on the beam state evolution sequence, may specifically include the following steps S470 to S4130: Step S470: Perform background vibration trend separation on the spatial state distribution map of each frame in the beam state evolution sequence. Use time-domain high-pass filtering to filter out the low-frequency baseline drift component that reflects the normal vibration of the turnout beam in the spatial state distribution map, and retain the high-frequency abnormal vibration component that reflects instantaneous impact and abnormal fluctuations to obtain the high-frequency abnormal vibration spatial distribution sequence.

[0110] For each fixed spatial grid location, the predicted vibration acceleration amplitude is extracted along the time axis across all frames of the beam state evolution sequence, forming the predicted vibration acceleration time series for that spatial grid location. A Butterworth high-pass filter is used, and the filtering operation is implemented in the time domain through iterative difference equations. The high-pass filtered predicted vibration acceleration time series is then reassembled back into the grid layout of the spatial state distribution map. The high-frequency anomalous vibration spatial distribution maps of all frames are arranged in their original time sequence to form a high-frequency anomalous vibration spatial distribution sequence.

[0111] Step S480: Calculate the cumulative energy descriptor of the high-frequency abnormal vibration components at each spatial location within the high-frequency abnormal vibration spatial distribution sequence in the time dimension, and generate a spatial distribution map of the cumulative energy of abnormal vibration. The spatial location with a larger energy descriptor in the spatial distribution map of the cumulative energy of abnormal vibration corresponds to a more significant duration of abnormal vibration.

[0112] For a given spatial grid location, the high-frequency anomalous vibration component values ​​across all time frames of the high-frequency anomalous vibration spatial distribution sequence are extracted to form a time series vector. The instantaneous energy value sequence is obtained by squaring each element in this time series vector. All instantaneous energy values ​​are summed, and the resulting sum is the cumulative energy descriptor for that spatial grid location. The cumulative energy descriptors for all spatial grid locations are then filled into a two-dimensional grid according to their spatial coordinates. This two-dimensional grid is the spatial distribution map of the anomalous vibration cumulative energy.

[0113] Step S490: Perform spatial clustering analysis on the spatial distribution map of abnormal vibration accumulated energy, and aggregate the sensing nodes whose abnormal vibration accumulated energy descriptor exceeds the preset energy threshold and are spatially adjacent into abnormal vibration clusters, and extract the mean energy descriptor, cluster area descriptor and cluster centroid coordinates of each abnormal vibration cluster.

[0114] First, threshold binarization is performed on each grid element in the spatial distribution map of accumulated energy from anomalous vibrations. Grid elements with accumulated energy descriptors greater than a preset energy threshold are marked as high-energy grids. Then, all high-energy grids are traversed, and a recursive region growing process aggregates all spatially connected high-energy grids into an independent anomalous vibration cluster. For each anomalous vibration cluster, the arithmetic mean of the accumulated energy descriptors of all grids within the cluster is calculated as the cluster's average energy descriptor. The total number of grids in the anomalous vibration cluster is counted and multiplied by the actual physical area corresponding to a single grid to obtain the cluster area descriptor. Finally, the energy-weighted average of the coordinates of all grids within the anomalous vibration cluster is calculated to obtain the cluster's centroid coordinates.

[0115] Step S4100: Analyze the centroid coordinate migration trajectory of each abnormal vibration cluster between different time frames of the beam state evolution sequence, determine the propagation direction and propagation path of the abnormal vibration on the turnout beam structure based on the centroid coordinate migration trajectory, and generate the vibration response abnormal propagation path including the abnormal propagation start position, the propagation passage position and the propagation termination position.

[0116] In one implementation, step S4100 may specifically include the following steps S4101 to S4106: Step S4101: Extract the set of cluster centroid coordinates of the same abnormal vibration cluster in a continuous time frame sequence, and arrange the set of cluster centroid coordinates into a coordinate sequence vector according to the order of the time frames.

[0117] Step S4102: Perform a first-order difference operation on the coordinate sequence vector to generate a displacement vector of the cluster centroid coordinates between adjacent time frames. The displacement vector includes the displacement direction angle and the displacement distance length.

[0118] Subtracting the cluster centroid coordinates of the previous time frame from the cluster centroid coordinates of the later time frame in the coordinate sequence vector yields the first and second displacement components. The resulting two-dimensional vector is the displacement vector. The displacement direction angle and displacement distance are then calculated based on these two displacement components.

[0119] Step S4103: Map the displacement direction angle of the displacement vector to a preset discrete direction sector. The discrete direction sector is based on the longitudinal mileage increase direction of the turnout beam and is divided into the full circumference direction range at equal angular intervals.

[0120] For example, the entire circumference is evenly divided into 8 discrete directional sectors, each covering a 45-degree angle range. The longitudinal mileage increase direction of the turnout beam is defined as the reference direction, and the corresponding displacement direction angle is set to 0 degrees.

[0121] Step S4104: Statistically analyze the distribution frequency of the displacement vector corresponding to the coordinate sequence vector in each discrete direction sector, and take the direction corresponding to the discrete direction sector with the highest distribution frequency as the main propagation direction of the abnormal vibration.

[0122] Step S4105: Project the coordinate sequence vector along the main propagation direction, calculate the projected mileage description of each coordinate point in the main propagation direction, and extract the monotonically increasing interval of the projected mileage description as the effective propagation path interval.

[0123] The unit direction vector is composed of the cosine and sine values ​​of the displacement direction angle corresponding to the main propagation direction. The projected mileage sequence is calculated by multiplying the coordinates of the first spatial distribution axis of the coordinate points by the first component of the unit direction vector, and adding the product of the coordinates of the second spatial distribution axis and the second component of the unit direction vector. The time interval in the projected mileage descriptor that continuously increases is extracted as the effective propagation path interval.

[0124] Step S4106: Define the sequence of physical locations of sensor nodes contained within the effective propagation path interval as the vibration response anomaly propagation path, and take the sensor node position corresponding to the initial projected mileage descriptor of the effective propagation path interval as the anomaly propagation start position, and the sensor node position corresponding to the final projected mileage descriptor as the anomaly propagation termination position.

[0125] Step S4110: Extract the vibration response prediction amplitude envelope of each sensing node on the abnormal propagation path of the vibration response within the beam state evolution sequence, and calculate the amplitude growth trend descriptor and amplitude fluctuation descriptor of the vibration response prediction amplitude envelope in the future time period.

[0126] For each sensing node along the abnormal propagation path of the vibration response, the predicted vibration acceleration amplitude at that node's location is extracted from all spatial state distribution maps of the beam's state evolution sequence, forming a time series. A Hilbert transform is applied to this time series to calculate the analytic signal; the magnitude of the analytic signal is the instantaneous envelope amplitude. The curve formed by connecting the envelope amplitudes at each time point is the vibration response prediction amplitude envelope. Linear regression is performed on the envelope; the slope of the fitted line is the descriptor of the amplitude growth trend. The sum of squared residuals between the envelope and the fitted line is calculated as the descriptor of the amplitude fluctuation.

[0127] Step S4120: Perform dimensionless processing on the amplitude growth trend description and amplitude fluctuation degree description respectively to obtain dimensionless growth trend characteristic value and fluctuation degree characteristic value.

[0128] Dividing the magnitude growth trend descriptor by the absolute value of the predicted vibration acceleration amplitude of the sensing node in the first frame of the beam state evolution sequence yields a dimensionless growth trend characteristic value. Dividing the magnitude fluctuation degree descriptor by the mean of the predicted amplitude envelope of the vibration response of the sensing node yields a dimensionless fluctuation degree characteristic value.

[0129] Step S4130: Based on the dimensionless growth trend characteristic value and fluctuation degree characteristic value, combined with the preset material fatigue characteristic description curve of the turnout beam structure, the structural weakness rating of each sensing node position on the abnormal propagation path of vibration response is performed, and the position with the highest weakness ranking in the rating result is marked as the structural weak link location information.

[0130] In one implementation, step S4130 may specifically include the following steps S4131 to S4136: Step S4131: Extract the stress amplitude and fatigue life relationship description curve corresponding to the current turnout beam material grade from the preset material fatigue characteristic description curve of the turnout beam structure. The stress amplitude and fatigue life relationship description curve reflects the number of cycles required for the material to reach fatigue failure under a specific stress amplitude cycle.

[0131] Step S4132: Convert the maximum amplitude of the vibration response prediction amplitude envelope at each sensing node location along the abnormal vibration response propagation path into the equivalent fatigue stress amplitude. The conversion process is based on the geometric characteristics of the turnout beam section and the constitutive relationship of material mechanics.

[0132] Based on the moment of inertia, material modulus of elasticity, and mode participation factor of the turnout beam at the sensing node, the acceleration amplitude is converted into the bending moment amplitude. Then, the bending modulus of the turnout beam is calculated based on its geometric dimensions. The bending moment amplitude is divided by the bending modulus to obtain the bending normal stress amplitude of the edge fibers. The rainflow counting method is used to cyclically count the stress time history, and representative stress amplitude levels are extracted as the equivalent fatigue stress amplitude.

[0133] Step S4133: By using the equivalent fatigue stress amplitude, look up the corresponding allowable cycle number reference descriptor on the stress amplitude and fatigue life relationship curve. The smaller the allowable cycle number reference descriptor, the more likely the corresponding sensing node location is to accumulate fatigue damage under continuous vibration.

[0134] Step S4134: Construct a two-dimensional weak point assessment coordinate system with dimensionless growth trend feature value as the first coordinate axis and dimensionless fluctuation degree feature value as the second coordinate axis, and map the dimensionless growth trend feature value and fluctuation degree feature value corresponding to each sensor node position to the two-dimensional weak point assessment coordinate system.

[0135] Step S4135: Draw a cluster of equal weakness description lines in the two-dimensional weakness assessment coordinate system. The cluster of equal weakness description lines is formed by connecting coordinate points with the same allowable cycle number reference description value. The equal weakness description line closer to the origin of the coordinate system corresponds to a longer allowable cycle number reference description value.

[0136] The drawing of the line clusters describing the degree of weakness is based on a pre-established weakness mapping model. This mapping model uses a multivariate nonlinear regression method to fit the functional relationship between the dimensionless growth trend feature value, the dimensionless fluctuation feature value, and the allowable cycle number reference descriptor.

[0137] Step S4136: Based on the relative positional relationship between the landing point of each sensor node in the two-dimensional weak assessment coordinate system and the weakness description line cluster, determine the weakness rating of each sensor node position. Sensor nodes whose landing points are close to the weakness description line cluster on the outer edge of the coordinate system will receive a higher weakness rating. Positions with weakness ratings exceeding the preset rating threshold will be output as structural weak link location information.

[0138] Step S500: Based on the abnormal propagation path of vibration response and the location information of weak structural links, generate a real-time monitoring command stream for the turnout beam that includes the monitoring section identification and vibration response trend prediction description, and push the real-time monitoring command stream for the turnout beam to the trackside monitoring and early warning terminal.

[0139] In one implementation, step S500 may specifically include the following steps S510 to S560: Step S510: Analyze the sequence of physical locations of sensor nodes contained in the abnormal propagation path of vibration response, extract the mileage number of the turnout beam section corresponding to each sensor node in the sequence of physical locations of sensor nodes, and use the mileage number of the turnout beam section as the basic coding element of the monitoring section identifier.

[0140] The vibration response anomaly propagation path consists of a series of sequentially arranged physical locations of sensor nodes, each with a unique longitudinal mileage coordinate value. The parsing operation queries the configuration database and retrieves the corresponding turnout beam section mileage marker based on the identifier of each sensor node in the sequence of physical locations. All retrieved turnout beam section mileage markers are then arranged in order along the vibration response anomaly propagation path, forming the basic coding element sequence for the monitoring section identification.

[0141] Step S520: Prioritize the sensor node locations included in the structural weak link location information, mark the top preset number of sensor node locations with the highest weakness rating as priority monitoring sections, and assign priority monitoring identification prefixes to priority monitoring sections that are different from ordinary monitoring sections.

[0142] Priority ranking is performed by sorting all records in descending order of their vulnerability severity rating. If the vulnerability severity ratings are the same, a weighted composite score is further compared between the dimensionless growth trend characteristic value and the dimensionless fluctuation characteristic value. A predetermined number of sensor node locations at the top of the ranked list are selected as priority monitoring sections, and a priority monitoring identifier prefix string is generated for each.

[0143] Step S530: Extract the vibration response prediction amplitude envelope of the sensing node corresponding to the priority monitoring section in the beam state evolution sequence, and analyze the amplitude change trend characteristics of the vibration response prediction amplitude envelope in the future time period. The amplitude change trend characteristics include upward trend type, stable trend type and downward trend type.

[0144] The envelope is divided into an early period and a later period on the time axis. The mean of the envelope for the early period and the mean of the envelope for the later period are calculated separately. The ratio of the later mean to the earlier mean is calculated as the amplitude change ratio. If the amplitude change ratio is greater than a preset upper threshold, it is determined to be an upward trend; if it is less than a preset lower threshold, it is determined to be a downward trend; if it is between the two, it is determined to be a stable trend.

[0145] Step S540: Generate a vibration response trend prediction description text template for each priority monitoring section based on the amplitude change trend characteristics. The vibration response trend prediction description text template includes an amplitude trend type field, a predicted peak occurrence time field, and a predicted peak amplitude range field.

[0146] The amplitude trend type field is filled with standard terminology based on the trend type determined in step S530. The predicted peak occurrence time field is filled with the time offset corresponding to the local maximum point of the envelope with the largest amplitude in the future time period. The predicted peak amplitude range field is filled with a description of the numerical range formed by the maximum and minimum values ​​of the envelope in the future time period.

[0147] Step S550: The basic coding elements of the monitoring section identifier, the priority monitoring identifier prefix, and the vibration response trend prediction description text template are encapsulated in a structured manner to generate a turnout beam real-time monitoring instruction data package containing instruction header field, section identifier field, trend description field, and verification tail field.

[0148] The instruction header field includes the frame synchronization header identifier, protocol version number, total data packet length, and source and destination device addresses. The section identifier field stores the complete monitoring section identifier string, composed of a priority monitoring identifier prefix and basic coded elements. The trend description field stores a text template string describing the vibration response trend prediction. The checksum field stores the checksum calculated by performing a cyclic redundancy check on all bytes of the aforementioned fields.

[0149] Step S560: According to the preset instruction sending cycle, the real-time monitoring instruction data packets of the turnout beam are combined into a continuously sent real-time monitoring instruction stream for the turnout beam, and the real-time monitoring instruction stream for the turnout beam is pushed to the designated data receiving port of the trackside monitoring and early warning terminal.

[0150] At the arrival of each instruction transmission cycle, the system collects newly generated vibration response anomaly propagation paths and structural weak point location information since the previous transmission cycle, and generates a real-time monitoring instruction data packet for the turnout beam in the current cycle. The data packets generated cycle by cycle are sequentially arranged to form a continuous real-time monitoring instruction stream for the turnout beam, and pushed to the data receiving port specified by the preset Internet Protocol address and transport layer port number of the trackside monitoring and early warning terminal through a pre-established transmission control protocol socket connection.

[0151] For example, a trackside monitoring and early warning terminal refers to a dedicated industrial electronic device installed beside the railway track to receive and display real-time monitoring command streams for turnout beams. This terminal can be deployed in a signal equipment cabinet or dedicated monitoring chassis near the section where the turnout beam is located, establishing a communication connection with the data aggregation equipment and monitoring system server of the vibration sensor network via a network interface. The trackside monitoring and early warning terminal internally includes a data processing unit, a display and interaction unit, and an audible and visual alarm unit. The data processing unit is responsible for parsing the real-time monitoring command data packets of the turnout beam received through a designated data receiving port, extracting the monitoring section identifier, priority monitoring identifier prefix, and vibration response trend prediction description text template, and driving the display and interaction unit to display the location distribution of priority monitoring sections, the trajectory line of the abnormal vibration response propagation path, and the location markers of structural weak points on the LCD screen using a graphical interface. When the audible and visual alarm unit receives information about the location of a high-risk structural weak point whose weakness rating exceeds a preset rating threshold, it triggers a buzzer to emit an audio alarm and drives a warning light to flash, alerting trackside maintenance personnel to immediately pay attention to the abnormal state of the designated section of the turnout beam structure. The trackside monitoring and early warning terminal also has historical data storage and query functions. It can replay the beam state evolution sequence and corresponding monitoring instructions generated during past train passing events according to the time index, providing data support for trend analysis of the health status of turnout beam structures and maintenance decisions.

[0152] Please see details. Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Figure 3As shown, the computer system 1000 described above may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer system 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 3 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0153] exist Figure 3 In the computer system 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

[0154] It should be understood that the computer system 1000 described in the embodiments of the present invention can execute the foregoing text. Figure 2 The implementation principle and beneficial effects of the real-time vibration response monitoring method for turnout beams based on neural networks described in the corresponding embodiments will not be repeated here.

Claims

1. A method for real-time vibration response monitoring of turnout beams based on neural networks, characterized in that, The method includes: The original vibration response signal stream collected by each sensing node in the vibration sensing network deployed on the key section of the turnout beam structure is obtained. The original vibration response signal stream includes a continuous vibration acceleration signal sequence carrying a timestamp and reflecting the dynamic response of the turnout beam structure when the train passes. The original vibration response signal stream is reconstructed in a spatiotemporal manner, and the continuous vibration acceleration signal sequence is mapped to a three-dimensional tensor representation space aligned with the physical coordinates of the turnout beam. This results in a spatiotemporal distribution tensor of the turnout beam vibration response with three coordinate axes: a time axis, a first spatial distribution axis, and a second spatial distribution axis. The element values ​​of the spatiotemporal distribution tensor of the turnout beam vibration response at any time and spatial coordinate represent the vibration acceleration amplitude at that moment and location. The spatiotemporal distribution tensor of the vibration response of the turnout beam is input into a pre-constructed train load backpropagation neural network. The train load backpropagation neural network performs load feature propagation and structural response mode generation on the spatiotemporal distribution tensor of the vibration response of the turnout beam, thereby obtaining the dynamic load transfer path mode and vibration energy attenuation topology map describing the train wheelset at each key section when passing through the turnout beam. Based on the dynamic load transfer path pattern and the vibration energy attenuation topology, a pre-constructed beam state evolution neural network is invoked to perform multi-step deduction of the vibration response propagation law of the turnout beam structure, thereby obtaining a beam state evolution sequence that characterizes the vibration response development trend of the turnout beam in the future period, and extracting abnormal vibration response propagation path and structural weak link location information based on the beam state evolution sequence. Based on the abnormal propagation path of the vibration response and the location information of the weak link in the structure, a real-time monitoring command stream for the turnout beam, which includes the monitoring section identification and vibration response trend prediction description, is generated and pushed to the trackside monitoring and early warning terminal.

2. The method according to claim 1, characterized in that, The process of reconstructing the spatiotemporal vibration response of the original vibration response signal stream maps the continuous vibration acceleration signal sequence to a three-dimensional tensor representation space aligned with the physical coordinates of the turnout beam, resulting in a spatiotemporal distribution tensor of the turnout beam vibration response with three coordinate axes: a time axis, a first spatial distribution axis, and a second spatial distribution axis. The element values ​​of this spatiotemporal distribution tensor at any time and spatial coordinate represent the vibration acceleration amplitude at that moment and location, including: The continuous vibration acceleration signal sequence collected by each sensing node in the original vibration response signal stream is analyzed, and the vibration acceleration signal segment of each sensing node in the continuous time window and the timestamp sequence corresponding to the vibration acceleration signal segment are extracted. Based on the pre-set physical coordinate information of each sensing node on the key section of the turnout beam structure, a three-dimensional spatial coordinate system is constructed with the longitudinal mileage, lateral offset and vertical elevation of the turnout beam as coordinate axes, and the physical coordinate information of each sensing node is mapped into the three-dimensional spatial coordinate system. The vibration acceleration signal segments corresponding to each sensing node mapped to the three-dimensional spatial coordinate system are synchronized in the time domain. Taking the starting trigger time of the train passing through the turnout beam as the alignment benchmark, the vibration acceleration amplitudes of different sensing nodes at the same time point are mapped to a unified time domain alignment framework to obtain a set of synchronized vibration acceleration sequences after time domain alignment. Using the time-domain aligned vibration acceleration synchronization sequence set and the spatial coordinate information of each sensing node in the three-dimensional spatial coordinate system, a three-dimensional data container is constructed with the time dimension as the first axis, the first spatial distribution dimension of the sensing nodes as the second axis, and the second spatial distribution dimension of the sensing nodes as the third axis. The vibration acceleration amplitude on each time-domain slice in the time-domain aligned vibration acceleration synchronization sequence set is filled into the corresponding grid position of the three-dimensional data container according to the corresponding first spatial distribution coordinate and second spatial distribution coordinate. Spatial dimension interpolation and temporal dimension smoothing are performed on the filled three-dimensional data container. The vibration acceleration amplitude is estimated for the spatial coordinate position of the un-deployed sensor node using an inverse distance weighted interpolation method based on the spatial distance of the sensor node. The vibration acceleration amplitude of adjacent time-domain slices in the temporal dimension is then filtered by a moving mean to obtain the interpolated and smoothed three-dimensional data container. The interpolated and smoothed three-dimensional data container is defined as the spatiotemporal distribution tensor of the vibration response of the turnout beam.

3. The method according to claim 2, characterized in that, The set of vibration acceleration synchronization sequences aligned in the time domain and the spatial coordinate information of each sensing node in the three-dimensional spatial coordinate system are used to construct a three-dimensional data container with the time dimension as the first axis, the first spatial distribution dimension of the sensing nodes as the second axis, and the second spatial distribution dimension of the sensing nodes as the third axis, including: The vibration acceleration amplitude values ​​on each time-domain slice in the time-domain aligned vibration acceleration synchronization sequence set are filled into the corresponding grid positions of the three-dimensional data container according to the corresponding first spatial distribution coordinates and second spatial distribution coordinates, so that the element value of the three-dimensional data container at any coordinate represents the vibration acceleration amplitude at the corresponding time point and spatial position. Traverse the time-domain aligned vibration acceleration synchronization sequence set to obtain the vibration acceleration amplitude and corresponding spatial coordinate information of each sensing node in each time-domain slice. The time-domain slice corresponds to a single sampling moment during the process of the train passing through the turnout beam. For each time-domain slice, a two-dimensional amplitude matrix is ​​created with the spatial distribution dimension of the sensor nodes as the row direction and the orthogonal direction of the spatial distribution dimension of the sensor nodes as the column direction. The row index and column index of the two-dimensional amplitude matrix correspond to the lateral offset coordinate and vertical elevation coordinate of each sensor node in the three-dimensional spatial coordinate system, respectively. The vibration acceleration amplitude of each sensor node in the current time domain slice is filled into the corresponding row and column intersection position of the two-dimensional amplitude matrix. For the row and column intersection position where no sensor node is deployed, a preset empty marker is used to occupy the position, so as to obtain the spatial amplitude snapshot matrix corresponding to the current time domain slice. According to the time sequence of the train passing through the turnout beam, the spatial amplitude snapshot matrices corresponding to each time domain slice are stacked sequentially along the time axis to form an initial three-dimensional data stack structure with time as the first axis and the rows and columns of the two-dimensional spatial amplitude snapshot matrix as the second and third axes. The initial three-dimensional data stack structure is normalized by tensor dimension, and the vibration acceleration amplitude of each element in the initial three-dimensional data stack structure is converted into a floating-point representation format with uniform dimensions. The dimensions of the time axis, spatial row axis, and spatial column axis are explicitly declared to obtain the normalized three-dimensional data container. The normalized three-dimensional data container is extended by boundary extension. A virtual sensing node layer with a preset boundary width is added to the outer edge of the spatial row axis and spatial column axis. The vibration acceleration amplitude of the virtual sensing node layer is assigned according to the vibration acceleration amplitude of the adjacent real sensing node and the spatial attenuation law to obtain the three-dimensional data container after boundary extension.

4. The method according to claim 2, characterized in that, After defining the interpolated and smoothed three-dimensional data container as the spatiotemporal distribution tensor of the turnout beam vibration response, the method further includes: Tensor decomposition is performed on the spatiotemporal distribution tensor of the vibration response of the turnout beam. The tensor of the spatiotemporal distribution of the vibration response of the turnout beam is sliced ​​and decomposed along the spatial distribution dimension of the sensing nodes to obtain the single-position vibration response time series vector corresponding to different spatial coordinate positions. The amplitude fluctuation characteristics of the vibration response time series vector at each single location during the entire time of train passage are analyzed. The peak acceleration occurrence time and peak acceleration amplitude of the vibration response time series vector at each single location are extracted, and a peak response distribution map indexed by spatial coordinates is constructed. Based on the chronological order of the peak acceleration occurrence times at adjacent spatial coordinate positions within the peak response distribution spectrum, the main transmission direction of the train load on the turnout beam structure and the load transmission speed characterization quantity are determined. The degree of cross-relationship between vibration response time series vectors at different spatial coordinate positions within the spatiotemporal distribution tensor of the turnout beam vibration response is calculated. Based on the degree of cross-relationship, a spatial correlation network of vibration response is constructed. In the spatial correlation network of vibration response, nodes represent the spatial positions of sensing nodes, and the weights of connecting edges represent the degree of similarity of vibration response waveforms between nodes. Community detection is performed on the vibration response spatial association network. Sensor nodes whose mutual relationship description exceeds a preset association threshold are assigned to the same vibration response community. Sensor nodes in the same vibration response community exhibit similar vibration response patterns when a train passes. Based on the vibration response community division results and the load transfer velocity characterization quantity, the spatiotemporal distribution tensor of the turnout beam vibration response is compressed and encoded to generate a low-dimensional spatiotemporal distribution tensor of the turnout beam vibration response that retains the core vibration response propagation characteristics.

5. The method according to claim 1, characterized in that, The process involves inputting the spatiotemporal distribution tensor of the turnout beam's vibration response into a pre-constructed train load backpropagation neural network. This network then performs load feature propagation and structural response mode generation on the turnout beam's spatiotemporal distribution tensor, resulting in a dynamic load transfer path pattern and vibration energy attenuation topology map describing the dynamic load transfer path pattern and vibration energy attenuation at key sections when the train wheelset passes over the turnout beam. This includes: The spatiotemporal distribution tensor of the turnout beam vibration response is input into the input layer of the train load backpropagation neural network. Batch normalization and dimension reshaping operations are performed on the spatiotemporal distribution tensor of the turnout beam vibration response to convert the spatiotemporal distribution tensor of the turnout beam vibration response into a standard input data stream that conforms to the input specification of the hidden layer of the train load backpropagation neural network. The load feature encoder of the train load backpropagation neural network extracts multi-level load features from the standard input data stream. The vibration acceleration amplitude gradient change features between adjacent sensing nodes in the standard input data stream are extracted through cascaded convolutional noise reduction layers to generate a local load action domain feature map that characterizes the local load influence range near the action position of the train wheelset. The local load domain feature map is input into the time-series load propagation chain processing layer of the train load backpropagation neural network, which includes a cyclic memory unit array that expands along the time dimension. The cyclic memory unit array performs progressive propagation path tracking on the local load domain feature map according to the time causality law of the turnout beam vibration response, and generates a load propagation state sequence describing the gradual transmission of the train wheelset load along the longitudinal direction of the turnout beam. The load propagation state sequence is mapped to a graph structure through the energy attenuation topology reconstruction layer of the train load backpropagation neural network. Each sensing node is regarded as a graph node, and the vibration energy transfer relationship between nodes is regarded as a graph edge. By comparing the differences in the node state values ​​at each time step in the load propagation state sequence, directed edge weights between adjacent graph nodes are generated. Topological feature aggregation is performed on the dynamic directed graph composed of the graph nodes and the directed edge weights. Graph convolution operation is performed on the dynamic directed graph at each time along the time dimension to extract the importance ranking information of the graph nodes and the activity change trend of the graph edges at different times, and to generate a dynamic load transfer path pattern and vibration energy attenuation topology graph that evolves over time. The dynamic load transfer path pattern and vibration energy attenuation topology are visualized and encoded to generate a set of load transfer path description lines with the physical coordinates of the turnout beam as the reference system and a set of energy attenuation layer description lines with the sensor node as the reference position.

6. The method according to claim 5, characterized in that, The multi-level load feature extraction of the standard input data stream by the load feature encoder through the train load backpropagation neural network includes: A first-level temporal convolution kernel array is applied to the standard input data stream. The first-level temporal convolution kernel array performs a sliding convolution operation on the standard input data stream along the time dimension to extract the abrupt rising edge and decay falling edge morphology of the vibration acceleration amplitude during the train passing through the turnout beam within a short time window, generating a primary temporal load feature map containing temporal edge response features. A second-level spatial convolution kernel array is applied to the primary time-domain load feature map. The second-level spatial convolution kernel array performs a sliding convolution operation on the primary time-domain load feature map along the spatial distribution dimension of the sensing nodes to extract the spatial gradient distribution features and spatial correlation features of the vibration acceleration amplitude between adjacent sensing nodes, and generate a secondary spatial load feature map containing a description of the spatial gradient direction. A third-level spatiotemporal joint convolution kernel array is applied to the secondary spatial load feature map. The third-level spatiotemporal joint convolution kernel array performs three-dimensional sliding convolution operation on the secondary spatial load feature map in both the time dimension and the spatial distribution dimension of the sensing nodes, extracting the linkage change pattern of the train load in the temporal propagation and spatial propagation process, and generating a deep spatiotemporal load feature map containing spatiotemporal correlation texture. The deep spatiotemporal load feature map is subjected to feature sparsification. Feature positions in the deep spatiotemporal load feature map with response intensity lower than a preset activation threshold are set to 0, while significant feature positions with response intensity higher than the preset activation threshold are retained, thereby generating a sparsified deep spatiotemporal load feature map. The sparse deep spatiotemporal load feature map is subjected to feature pyramid pooling. A pooling window with a first size is used to downsample the sparse deep spatiotemporal load feature map to generate a first pooling feature map, and a pooling window with a second size is used to downsample the sparse deep spatiotemporal load feature map to generate a second pooling feature map. The first size is larger than the second size. The first pooling feature map represents the overall load distribution trend of the turnout beam, and the second pooling feature map represents the local load concentration phenomenon near the wheelset action point. The first pooling feature map and the second pooling feature map are concatenated and fused through channels. The concatenated feature channels are then subjected to cross-channel information interaction and dimensionality reduction using pointwise convolution to generate the local load domain feature map that integrates multi-scale load distribution information.

7. The method according to claim 5, characterized in that, After performing graph structure mapping processing on the load propagation state sequence through the energy attenuation topology reconstruction layer of the train load backpropagation neural network, the process further includes: The graph Laplacian matrix at each time step is extracted from the dynamic directed graph. The diagonal elements of the graph Laplacian matrix represent the out-degree or in-degree strength of each graph node, and the off-diagonal elements represent the negative values ​​of the directed edge weights between adjacent graph nodes. Perform eigenvalue decomposition on the graph Laplacian matrix at each time step to calculate the eigenvalue sequence and eigenvector matrix of the graph Laplacian matrix at each time step; The eigenvector corresponding to the largest eigenvalue in the eigenvalue sequence is extracted as the principal mode distribution vector. The magnitude of each element in the principal mode distribution vector reflects the dominant position of the corresponding sensing node in the vibration energy propagation process at the current moment. Analyze the evolution trajectory of the principal mode distribution vector over time, extract the moment when the relative order of elements in the principal mode distribution vector changes abruptly, and mark the moment when the change occurs as the moment when the load transfer path switching event occurs; Based on the occurrence time of the load transfer path switching event and the corresponding principal mode distribution vector, a path level label is executed on the set of load transfer path description lines in the dynamic load transfer path mode. The corresponding load transfer path description lines are assigned level identifiers according to the size of each element in the principal mode distribution vector, and the direction pointing mark of the load transfer path description lines is determined according to the direction of element sorting change. The set of load transfer path description lines carrying hierarchical identifiers and direction pointing marks, along with the vibration energy attenuation topology map, are input into the output mapping layer of the train load backpropagation neural network. The output mapping layer then maps the graph node coordinates in the set of load transfer path description lines carrying hierarchical identifiers and direction pointing marks, as well as the graph node coordinates in the vibration energy attenuation topology map, back to the physical coordinate system of the turnout beam, generating the final dynamic load transfer path pattern and vibration energy attenuation topology map that correspond one-to-one with the key section positions of the turnout beam structure.

8. The method according to claim 1, characterized in that, Based on the dynamic load transfer path pattern and the vibration energy attenuation topology, a pre-constructed beam state evolution neural network is invoked to perform multi-step deduction of the vibration response propagation law of the turnout beam structure, obtaining a beam state evolution sequence characterizing the vibration response development trend of the turnout beam in the future time period, including: The dynamic load transfer path pattern is converted into a load transfer state sequence represented by a graph node state vector sequence, the vibration energy attenuation topology is converted into a vibration energy attenuation relation matrix represented by an adjacency matrix, and the load transfer state sequence and the vibration energy attenuation relation matrix are used as the initial input data of the beam state evolution neural network. Through the state initialization layer of the beam state evolution neural network, a coupling mapping operation is performed on the initial moment state vector of the load transfer state sequence and the vibration energy attenuation relationship matrix. The load state values ​​of each graph node in the initial moment state vector are distributed to adjacent graph nodes according to the attenuation ratio defined in the vibration energy attenuation relationship matrix, thereby generating an initial state transition vector that characterizes the vibration response distribution prediction of the turnout beam structure at the next moment. The initial state transition vector is input into the recurrent evolution unit chain of the beam state evolution neural network. The recurrent evolution unit chain consists of multiple gated recurrent units arranged in series. Each gated recurrent unit receives the hidden state vector of the previous time step and the input state vector of the current time step. Through the internal gating mechanism, it performs selective forgetting and selective updating operations on the historical state information and the current input information to generate the hidden state vector of the current time step and pass it to the next gated recurrent unit. In each iteration of the cyclic evolution unit chain, the beam state evolution neural network dynamically adjusts the vibration energy attenuation relationship matrix according to the hidden state vector output at the previous moment. Based on the relative magnitude of the state values ​​of each graph node in the hidden state vector, it enhances the attenuation weight of the edge corresponding to the graph node with the higher state value and weakens the attenuation weight of the edge corresponding to the graph node with the lower state value, thereby generating a dynamic attenuation relationship matrix that is adapted to the current vibration response state of the beam. The hidden state vector output by the cyclic evolution unit chain at the current time and the dynamic decay relationship matrix are input into the state decoder of the beam state evolution neural network. The state decoder converts the hidden state vector into a spatial state distribution map with the same dimension as the original sensor node spatial distribution through deconvolution mapping operation. The spatial state distribution map represents the predicted vibration acceleration amplitude distribution of each sensor node position of the turnout beam at the corresponding future time. The spatial state distribution maps output at each moment during the continuous multi-step simulation process are arranged in chronological order to form the beam state evolution sequence that characterizes the vibration response development trend of the turnout beam in the future time period. Each frame of the spatial state distribution map in the beam state evolution sequence corresponds to a vibration response prediction state at a future moment.

9. The method according to claim 8, characterized in that, In each iteration of the cyclic evolution unit chain, the beam state evolution neural network dynamically adjusts the vibration energy attenuation relationship matrix based on the hidden state vector output at the previous moment, including: The hidden state vector output at the previous moment is parsed, and the state values ​​corresponding to each graph node in the hidden state vector are extracted. The state values ​​represent the vibration response intensity of the corresponding sensing node position at the previous moment. Calculate the proportion of each graph node's state value to the total number of graph node state values. Graph nodes with a proportion exceeding a preset proportion threshold are marked as high-response dominant nodes, and graph nodes with a proportion below the preset proportion threshold are marked as low-response subordinate nodes. For the set of incoming edges pointing to the high-response dominant node in the vibration energy attenuation relationship matrix, the attenuation weight of each edge in the set of incoming edges is increased proportionally according to the proportion of the source node state value to the total state value of all nodes. The larger the proportion, the greater the increase in attenuation weight. For the set of outgoing edges in the vibration energy attenuation relationship matrix that points from the high-response dominant node to the low-response subordinate node, the original attenuation weight of each edge in the set of outgoing edges remains unchanged or is slightly reduced. For the set of edges connecting two low-response subordinate nodes in the vibration energy attenuation relationship matrix, the attenuation weight of the edges is proportionally reduced based on the product of the proportion of the state values ​​of the two low-response subordinate nodes to the total state values ​​of all nodes. The vibration energy attenuation relation matrix after the above-mentioned differential adjustment operation is used as the dynamic attenuation relation matrix, and the dynamic attenuation relation matrix is ​​used in the hidden state vector update calculation process of the next gated loop unit in the current iteration step.

10. A computer system, characterized in that, include: processor; And a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 9.

Images