Railway track periodic intelligent monitoring method based on Internet of Things

Abstract

The invention provides a railway track periodical intelligent monitoring method based on the Internet of Things. The railway track periodical intelligent monitoring method comprises the steps that track geometric deformation data and track temperature distribution data are collected through sensing nodes deployed beside a track and on the track; extracting abnormal sections in the track geometric deformation data based on time window segmentation, and mapping the abnormal sections to a space grid of the track temperature distribution data to generate a deformation-temperature incidence matrix; calculating a track stress change gradient by combining the matrix and historical track temperature distribution data, and generating a track health degree evaluation index; and when the track health degree evaluation index meets a preset degradation condition, inputting the time sequence multi-source fusion data into a deep learning track state prediction module, and outputting a track failure risk level. According to the invention, the efficiency, accuracy and intelligent level of railway track monitoring can be improved, and the safety and reliability of railway transportation are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of railway track safety monitoring, and more particularly to a railway track periodic intelligent monitoring method based on Internet of Things. BACKGROUND

[0002] In the field of railway transportation, track safety monitoring has always been a key link to ensure the safety of train operation. The traditional track monitoring method mainly relies on manual inspection and regular maintenance, which is not only inefficient, but also difficult to find potential safety hazards of the track in real time. With the development of technology, some automated monitoring means have been gradually introduced, such as track inspection cars and track deformation monitoring systems. Although track inspection cars can accurately detect the geometric state of the track, their detection frequency is low and they cannot monitor the dynamic changes of the track in real time. While the existing track deformation monitoring system can monitor the geometric deformation of the track in real time, it often lacks comprehensive consideration of environmental factors such as track temperature, resulting in incomplete and inaccurate monitoring results.

[0003] In addition, with the rapid development of railway transportation, the complexity of railway lines and the diversity of operating environments are increasing. For example, high temperature, heavy load, busy trunk lines and small radius curves and other special scenarios pose higher requirements for the safety and stability of the track. The existing monitoring technology often has problems such as insufficient monitoring accuracy and delayed early warning when dealing with these complex scenarios. At the same time, the traditional monitoring system also has limitations in data processing and analysis, and cannot fully utilize multi-source data for comprehensive evaluation, making it difficult to achieve accurate prediction and risk warning of the track state.

[0004] In the implementation of the present application, there are at least the following problems or defects in the prior art: first, the existing monitoring technology cannot simultaneously monitor and analyze multi-source data such as track geometric deformation and track temperature, resulting in one-sided monitoring results and failing to fully reflect the actual health status of the track; second, the existing technology lacks targeted monitoring strategies and early warning mechanisms when dealing with complex operating scenarios, and cannot timely discover potential risks of the track; third, the data processing capacity of the existing monitoring system is limited, and it is difficult to efficiently process large-scale monitoring data and realize real-time evaluation and prediction of the track state. SUMMARY

[0005] The present application provides a railway track periodic intelligent monitoring method based on Internet of Things, comprising: acquiring track monitoring data, the track monitoring data being collected by a sensing node deployed beside the track and on the track, including track geometric deformation data collected by a track detector array and track temperature distribution data collected by a track temperature sensor array; Abnormal segment extraction is performed on track geometric deformation data based on time window segmentation, segments whose deformation values exceed a set threshold in a continuous monitoring period are screened out, and deformation abnormal segment data is generated; The deformation abnormal segment data is mapped to a spatial grid corresponding to the track temperature distribution data to generate a deformation-temperature correlation matrix; The track stress change gradient is calculated by combining the deformation-temperature correlation matrix and historical track temperature distribution data to generate a track health assessment index; When the track health assessment index meets the preset degradation condition, the constructed time series multi-source fusion data is input into the deep learning track state prediction module, and the track failure risk level is output; wherein the time series multi-source fusion data includes track temperature distribution data and deformation-temperature correlation matrix in a continuous monitoring period.

[0006] Further, the generation of the deformation-temperature correlation matrix comprises: Performing Gaussian-Krueger projection conversion on the deformation abnormal segment data to calculate the geodetic coordinates of the abnormal segment; Based on the grid resolution of the track temperature sensor array, bilinear interpolation is performed to determine the mapping unit of the deformation segment in the track temperature grid; An empty matrix with the same dimension as the track temperature distribution data is constructed to initialize the deformation-temperature correlation matrix; Fill in the temperature gradient value at the mapping unit coordinate to generate a weighted correlation matrix; The unmapped units in the correlation matrix are set to invalid values; wherein the track detector array and the track temperature sensor array are distributedly deployed at a fixed interval.

[0007] Further, after generating the deformation-temperature correlation matrix, it further comprises: Extracting the area where the temperature gradient change exceeds the set rate in the correlation matrix to generate a track temperature anomaly mask; Analyzing the second derivative of the track geometric deformation data to screen segments with positive deformation acceleration; Performing morphological dilation operation on the deformation acceleration segment to generate a deformation trend mask; Calculate the spatial overlap area ratio of the track temperature anomaly mask and the deformation trend mask; When the overlap area ratio exceeds the set proportion, a multi-source alarm trigger factor is generated.

[0008] Further, the generation of the track health assessment index comprises: Extracting the maximum stress concentration coefficient from the deformation-temperature correlation matrix; Retrieving the stress reference value in the historical track temperature distribution data at the same location; Dynamically correcting the stress reference value using the exponential weighted moving average algorithm; Calculate the relative deviation of the current stress concentration coefficient and the corrected reference value; Map the health degree evaluation index based on the deviation value, and the index value interval is 0-1.

[0009] Further, the deep learning track state prediction module performs operations including: Extract the spatial temperature field features of the track temperature distribution data through a three-dimensional convolution layer; Extract the stress distribution features of the deformation-temperature correlation matrix through a three-dimensional convolution layer; Cascade the spatial temperature field features and the stress distribution features at the channel level to generate a fusion feature map; Use a long short-term memory network to analyze the time evolution law of the fusion feature map; Generate a state feature vector based on attention mechanism weighted fusion of spatio-temporal features; Output five types of track states and corresponding confidence through a fully connected layer classifier.

[0010] Further, before inputting the time series multi-source fusion data into the deep learning track state prediction module, a data expansion process is further included: Monitor the number of consecutive periods when the track health degree evaluation index exceeds the warning threshold; When the number of consecutive over-limit periods reaches a set value, automatically trigger the data expansion mechanism; Retrieve a complete data set of a set number of historical monitoring periods before the current time point from the edge storage database, which includes track geometric deformation data and corresponding track temperature distribution data; Read the track temperature distribution data of a set number of real-time monitoring periods after the current time point from the real-time data buffer; Align the historical data set and the real-time data set in the time dimension, and unify the data sampling frequency by timestamp matching; Perform spatial grid filling on the aligned data, and use the temperature gradient change rate of adjacent grids to interpolate and complete the missing area data; Construct an expanded time series data set containing historical and real-time data, with a time span greater than the original monitoring period.

[0011] Further, the preset degradation condition is dynamically configured according to the operation scenario: During the summer high-temperature period monitoring period, when the single-period decline of the track health degree evaluation index exceeds the high-temperature scenario set threshold, trigger the degradation determination; During the monitoring period of heavy-load freight lines, when the cumulative decline of the track health degree evaluation index for three consecutive periods exceeds the heavy-load scenario set threshold, trigger the degradation determination; For the busy trunk line with total weight exceeding the standard, the deterioration determination is triggered when the track health evaluation index continuously falls below the safety threshold for a set number of consecutive periods; For the curve section with a curve radius less than a set value, the deterioration determination is triggered when the range of the track health evaluation index in a set length range exceeds the set threshold for the curve scene.

[0012] Further, the acoustic verification after outputting the track failure risk level includes: When the risk level is determined to be level four or five, the acoustic sensor array installed on the rail waist is automatically awakened; Collect the original vibration signal of the steel rail and perform signal preprocessing, and use wavelet transform algorithm to filter out the interference of locomotive running noise; Three types of feature parameters are extracted from the denoised vibration signal: fundamental frequency amplitude decay rate, third harmonic distortion, and spectral energy entropy value; The feature parameters are combined and input into the pre-trained steel rail crack recognition model, which is based on deep neural network; The internal implicit crack propagation probability value and the fracture risk index of the steel rail are output, and the fracture risk index is a continuous value of 0-100%; According to the preset interval of the fracture risk index, the corresponding disposal level instruction is generated.

[0013] Further, the execution process of the grading disposal instruction includes: When the fracture risk index is in the first preset interval, a first level speed limit instruction is sent to the train dispatching control system, requiring the train speed through the section to be reduced to a set value; When the fracture risk index is in the second preset interval, a red warning maintenance work order with geographic coordinates is automatically generated and pushed to the emergency terminal of the track maintenance section and the countdown disposal process is started; When the fracture risk index exceeds the preset fracture threshold, the railway signal blocking system control protocol is triggered to forcibly block the train operation right of the affected section; Synchronously activate the high-definition camera group beside the track to shoot the steel rail surface image at a set frame rate per second; Run the steel rail surface defect detection algorithm to identify the crack, fracture, and block physical damage features in the image; Cross-verify the image recognition result with the acoustic detection conclusion, and when they are consistent, confirm the final disposal instruction.

[0014] Further, the core operations performed by the edge computing gateway include: Real-time receive the original deformation data stream uploaded by the track detector array, and use the sliding window filtering algorithm to remove outliers and smooth the data fluctuations; Synchronously process the temperature data uploaded by the track temperature sensor array, and perform real-time compensation for temperature drift based on the sensor calibration parameters; The morphing abnormal section extraction algorithm is executed on the local computing resource, and the time window length is dynamically adjusted to adapt to the change of train operation density; A morphing-temperature matrix associated with space and time is constructed, and distributed memory computing technology is used to accelerate the matrix operation process; The track health evaluation index is calculated according to the monitoring period, and the historical data of the index is compressed and stored in the local solid state memory; A light version of the deep learning track state prediction model is loaded, and the risk level prediction calculation is directly performed on the edge side; According to the prediction result, it is decided whether to trigger the acoustic verification and hierarchical disposal process, and the cloud data transmission delay is reduced.

[0015] The above embodiments according to the present application have at least the following beneficial effects: 1. By deploying sensing nodes to collect track geometric deformation data and track temperature distribution data, and extracting morphing abnormal sections based on time window segmentation, comprehensive monitoring and accurate positioning of track state are realized. This method not only can monitor the geometric deformation of the track in real time, but also can analyze the track temperature distribution data comprehensively, effectively solving the problem of single monitoring data in traditional monitoring methods, which cannot fully reflect the track health status, improving the accuracy and reliability of the monitoring results.

[0016] 2. The deformation-temperature correlation matrix is combined with the historical track temperature distribution data to calculate the track stress change gradient and generate the track health evaluation index. This innovative method can dynamically reflect the health status change of the track under different environmental conditions, solving the problem of insufficient monitoring accuracy and untimely warning in existing technologies under complex operation scenarios. At the same time, the deep learning track state prediction module analyzes the time series multi-source fusion data, realizing the early warning of track failure risk, and providing strong support for the safety maintenance of railway track.

[0017] 3. The present application introduces an edge computing gateway to process the data uploaded by the sensing nodes in real time, execute the morphing abnormal section extraction algorithm, and calculate the track health evaluation index and risk level prediction locally. This method effectively reduces the cloud data transmission delay, improves the response speed and real-time performance of the monitoring system. At the same time, combined with the execution process of acoustic verification and hierarchical disposal instructions, the accuracy of the monitoring results is further ensured, and appropriate disposal measures can be taken in time according to the actual health status of the track, improving the intelligent level and safety guarantee ability of railway track monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and other objects, features and advantages of the present application will become readily apparent from the following detailed description read in conjunction with the accompanying drawings. In the drawings, which represent several embodiments of the present application in example, but not limiting, manner: Figure 1 A flowchart of a railway track periodic intelligent monitoring method based on the Internet of Things is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0020] As shown in Figure 1 The present application proposes a railway track periodic intelligent monitoring method based on the Internet of Things, comprising: S1, acquiring track monitoring data, the track monitoring data being collected by a sensing node deployed on a trackside and a track, including track geometric deformation data collected by a track detector array and track temperature distribution data collected by a track temperature sensor array; S2, extracting an abnormal section of the track geometric deformation data based on time window segmentation, screening out a section whose deformation value exceeds a set threshold in a continuous monitoring period, and generating deformation abnormal section data; S3, mapping the deformation abnormal section data to a spatial grid corresponding to the track temperature distribution data, and generating a deformation-temperature correlation matrix; S4, calculating a track stress change gradient in combination with the deformation-temperature correlation matrix and historical track temperature distribution data, and generating a track health degree evaluation index; S5, when the track health degree evaluation index meets a preset degradation condition, inputting the constructed time series multi-source fusion data into a deep learning track state prediction module, and outputting a track failure risk level; wherein the time series multi-source fusion data includes track temperature distribution data and deformation-temperature correlation matrix in a continuous monitoring period.

[0021] The present application constructs a track health evaluation system through multi-source data fusion and spatio-temporal correlation analysis, and realizes accurate monitoring of track state in complex operation scenarios by combining dynamic threshold judgment and deep learning prediction. Specifically, it synchronously collects geometric deformation and track temperature data and establishes a spatial mapping relationship, uses time window segmentation and stress gradient calculation to realize dynamic identification of abnormal sections, and finally forms a closed-loop monitoring system from data collection to risk warning by fusing spatio-temporal features through time series data expansion and three-dimensional convolutional neural network.

[0022] Specifically, first, track monitoring data is collected by sensor nodes deployed on the trackside and track, including track geometric deformation data collected by track detector arrays and track temperature distribution data collected by track temperature sensor arrays. These data reflect the real-time state and environmental conditions of the track.

[0023] Next, the track geometric deformation data is segmented by time window and abnormal section is extracted. By setting a threshold, the section whose deformation exceeds the threshold in the continuous monitoring period is selected to generate deformation abnormal section data. This step can effectively identify potential track problem areas.

[0024] Then, the deformation abnormal section data is mapped to the spatial grid corresponding to the track temperature distribution data to generate a deformation-temperature correlation matrix. This step establishes the correlation between track deformation and temperature distribution, providing a basis for subsequent analysis.

[0025] Based on the deformation-temperature correlation matrix and historical track temperature distribution data, the track stress change gradient is calculated to generate a track health degree evaluation index. This index considers deformation and temperature factors and can more accurately reflect the health status of the track.

[0026] When the track health degree evaluation index meets the preset degradation condition, the system inputs the constructed time series multi-source fusion data into the deep learning track state prediction module. These fusion data include track temperature distribution data and deformation-temperature correlation matrix in the continuous monitoring period, containing rich spatio-temporal information.

[0027] Finally, the deep learning module outputs the track failure risk level based on the input fusion data. This prediction result can provide an important reference for railway operation and maintenance.

[0028] Specifically, track monitoring data refers to track state information collected by sensor nodes installed on the trackside and track, including track geometric shape change data collected by track detector arrays and temperature distribution data collected by track temperature sensor arrays. Distributed deployment of sensor nodes realizes multi-dimensional data collection, solving the problem of single monitoring data in traditional monitoring.

[0029] The time window segmentation refers to a processing method of dividing the continuously monitored track geometric deformation data according to a fixed time interval. Specifically, a sliding window algorithm can be used to dynamically intercept data segments, and a threshold value is set to screen the abnormal sections of the deformation variable in the continuous period, thereby improving the timeliness of the abnormality detection. The deformation-temperature correlation matrix refers to a multi-dimensional data set formed by spatially mapping the geometric deformation abnormal section and the track temperature distribution data. Specifically, the spatial coordinates are converted by the Gauss-Krueger projection and the bilinear interpolation is performed to realize the grid alignment, thereby revealing the physical correlation between the temperature gradient and the deformation.

[0030] The track health assessment index refers to a quantitative parameter reflecting the safety state of the track structure. Specifically, the index value is generated by calculating the dynamic deviation of the current stress concentration coefficient and the historical reference value. The exponential weighted moving average algorithm is used to correct the reference value, thereby realizing the accurate evaluation of the track state change. The time series multi-source fusion data refers to a composite data set containing the track temperature distribution data and the deformation-temperature correlation matrix in the continuous monitoring period. Specifically, the spatial features are extracted by the three-dimensional convolution layer, and the time evolution law is analyzed by combining the long short-term memory network, thereby solving the problem of time and space feature separation in the traditional method. The deep learning track state prediction module refers to a neural network model based on the spatio-temporal feature fusion. Specifically, the temperature field features and the stress distribution features are fused in a channel cascading manner, and the state feature vector is generated by combining the attention mechanism weighting, thereby realizing the multi-dimensional prediction of the track failure risk.

[0031] The preset degradation condition refers to a threshold decision rule for triggering the track state prediction. Specifically, the differentiated conditions such as the single-period descending threshold value in the high-temperature period and the cumulative descending threshold value of the heavy-load line are dynamically configured according to different operation scenarios, thereby enhancing the adaptability to complex scenarios. As a preferred embodiment, the scheme of the present application is implemented as follows: On a certain high-speed rail line, a track detector array and a track temperature sensor array are arranged along the line. The track detector array collects track geometric deformation data at a frequency of 10 Hz, including track gauge, level, height, and track orientation parameters. The track temperature sensor array collects track temperature distribution data at an interval of 1 minute.

[0032] The system uses 15 minutes as a time window to segment and extract abnormal data from the track geometric deformation data. The track gauge change threshold is set to ±2 mm, the level change threshold is set to ±5 mm, the height change threshold is set to ±4 mm, and the track orientation change threshold is set to ±3 mm. When any parameter exceeds the threshold in two consecutive time windows, the section is marked as a deformation abnormal section.

[0033] The deformation abnormal section data is mapped to the track temperature distribution data grid through spatial coordinates, generating a 100m×100m resolution deformation-temperature correlation matrix. Each cell in the matrix contains the deformation variable and the temperature value at that position.

[0034] The system retrieves the track temperature distribution data of the past 7 days, combines the current deformation-temperature correlation matrix, and calculates the track stress change gradient. The exponential weighted moving average algorithm is used to generate a track health evaluation index between 0 and 1.

[0035] The health evaluation index degradation condition is set as being lower than 0.8 for 3 consecutive time windows. When this condition is met, the system automatically triggers the deep learning prediction module. The input of the module includes track temperature distribution data in the last 24 hours and the deformation-temperature correlation matrix.

[0036] The deep learning module uses a three-dimensional convolutional neural network structure to extract spatio-temporal features and outputs a 5-level track failure risk prediction result through a fully connected layer.

[0037] The application further proposes a method for generating a deformation-temperature correlation matrix, including performing Gaussian-Krueger projection conversion on deformation anomaly segment data to calculate geodetic coordinates, determining a mapping unit through bilinear interpolation based on the grid resolution of the track temperature sensor array, constructing an empty matrix of the same dimension and initializing the correlation matrix, filling the temperature gradient value to generate a weighted matrix at the mapping unit coordinates, and setting the non-mapping unit to an invalid value.

[0038] The Gaussian-Krueger projection conversion converts the track linear coordinates into the geographic coordinate system, eliminating the spatial projection error of the track curve segment. The bilinear interpolation performs weighted calculation based on the temperature values of the four adjacent nodes of the track temperature grid, adapting to the data resolution difference caused by different sensor spacings. The dimension of the empty matrix is strictly consistent with the track temperature distribution data, ensuring the dimension matching of the subsequent matrix operation. The temperature gradient value is filled with the weighted average value of the temperature change rate of the adjacent grid, reflecting the dynamic characteristics of the local temperature field. The invalid value setting marks the non-mapping area to avoid noise data interference in the matrix operation.

[0039] Specifically, after the geodetic coordinates of the abnormal segment are converted by the Gaussian-Krueger projection, the geographic spatial position and the coordinate system of the track temperature grid are unified. The bilinear interpolation process calculates the relative position weight coefficient of the deformation segment center point in the temperature grid according to the actual distribution density of the track temperature sensor nodes, ensuring the spatial alignment accuracy of the temperature data and the deformation data. The initialized empty matrix builds a complete spatial correlation model through unit-by-unit assignment, and the temperature gradient value is filled with a sliding window algorithm to calculate the local temperature change trend. The invalid value processing of the non-mapping unit sets a special identifier, which automatically skips the calculation of these areas in the subsequent data processing process. This process effectively solves the spatial matching problem of heterogeneous sensor data and provides an accurate input data basis for subsequent stress gradient calculation.

[0040] As a preferred embodiment, the scheme of the application is implemented as follows: Generating the deformation-temperature correlation matrix comprises the following steps: Firstly, the Gaussian-Krueger projection conversion is performed on the deformation anomaly section data to calculate the geodetic coordinates of the anomaly section. Specifically, the WGS84 coordinate system is used as the reference to convert the longitude and latitude coordinates of the anomaly section into X and Y coordinate values in the plane rectangular coordinate system.

[0041] Further, based on the grid resolution of the track temperature sensor array, a bilinear interpolation is performed to determine the mapping unit of the deformation section in the track temperature grid. For example, the track temperature sensor array is arranged at an interval of 10 meters to form a 100x100 temperature grid. Through the bilinear interpolation algorithm, the anomaly section coordinates are mapped into the nearest four temperature grid units.

[0042] Next, an empty matrix of the same dimension as the track temperature distribution data is constructed to initialize the deformation-temperature correlation matrix. The number of rows and columns of the matrix is consistent with the number of rows and columns of the temperature grid.

[0043] Thus, the temperature gradient values are filled in at the mapping unit coordinates to generate a weighted correlation matrix. The temperature gradient value is calculated by the temperature difference of adjacent grid units divided by the grid interval. The weight value is set according to the deformation anomaly degree, and the higher the anomaly degree, the greater the weight value.

[0044] Finally, the unmapped units in the correlation matrix are set to invalid values. The track detector array and the track temperature sensor array are deployed at a fixed interval. For example, the track detector array is arranged with a detection unit every 50 meters, and the track temperature sensor array is arranged with a temperature sensor every 10 meters.

[0045] The present application further proposes that after generating the deformation-temperature correlation matrix, it further comprises: extracting the area where the temperature gradient change exceeds the set rate to generate a track temperature anomaly mask; analyzing the second derivative of the track geometric deformation data to filter the sections where the deformation acceleration is continuously positive; performing a morphological dilation operation on the deformation acceleration section to generate a deformation trend mask; calculating the spatial overlap area ratio of the track temperature anomaly mask and the deformation trend mask; and generating a multi-source alarm trigger factor when the overlap area ratio exceeds the set proportion.

[0046] The set range of the temperature gradient change rate can be configured to 0.5°C to 2°C per minute, and the specific value is adjusted according to the thermal expansion coefficient of the steel rail material. The determination condition that the deformation acceleration is continuously positive requires that the second derivative is greater than zero in the continuous three sampling periods. The morphological dilation operation uses a 3x3 rectangular structure element, and the dilation number is set to two. The set proportion of the spatial overlap area ratio is dynamically adjusted according to the line grade, for example, the dry line railway is set to 60%, and the branch line railway is set to 40%.

[0047] Specifically, the track temperature anomaly mask filters the mutation area in the correlation matrix by setting a temperature gradient change rate threshold, for example, when the temperature gradient of a certain grid cell rises by more than 5°C within five minutes, it is marked as abnormal. The deformation trend mask identifies the section with continuously increasing deformation rate by calculating the second derivative of the track geometry deformation data, and then expands the detection range to adjacent sections by morphological dilation. The calculation of the spatial overlap area ratio uses raster processing, projects the track temperature anomaly mask and the deformation trend mask into the same coordinate system, and counts the proportion of overlapping grid cells to the total number of grid cells. When the proportion exceeds the preset threshold, a multi-source alarm trigger factor is triggered, indicating that there is a significant correlation between temperature mutation and deformation acceleration in spatial distribution, and the composite early warning mechanism needs to be started. This process cross- verifies multidimensional data, effectively reduces the probability of false judgment of a single indicator, and improves the reliability of the early warning signal. As a preferred embodiment, the scheme of the present application is implemented as follows: After generating the deformation-temperature correlation matrix, the following steps are further performed: First, extract the area where the temperature gradient change exceeds the set rate in the correlation matrix to generate a track temperature anomaly mask. Specifically, the temperature gradient change rate threshold can be set to 2°C / m / h, the temperature gradient of each grid cell in the correlation matrix is calculated, and the grid cells exceeding the threshold are marked as 1, and the rest are marked as 0, thereby generating a binary track temperature anomaly mask.

[0048] Secondly, analyze the second derivative of the track geometry deformation data and select the section with continuously positive acceleration. For example, the central difference method can be used to calculate the second derivative of the track geometry deformation data, and a time window such as 30 minutes is set. Within this window, the sign of the second derivative is continuously detected, and the section with continuously positive sign is marked.

[0049] Then, perform morphological dilation operation on the deformation acceleration section to generate a deformation trend mask. Thus, a 3x3 structure element can be used to dilate the selected deformation acceleration section to expand its influence range and form a deformation trend mask.

[0050] Then, calculate the spatial overlap area ratio of the track temperature anomaly mask and the deformation trend mask. The specific operation is to perform pixel-by-pixel AND operation on the two masks, count the number of 1 in the result, and divide the total number of mask pixels to obtain the overlap area ratio.

[0051] Finally, when the overlap area ratio exceeds the set proportion, a multi-source alarm trigger factor is generated. For example, the set proportion can be set to 20%, and when the calculated overlap area ratio exceeds this threshold, a multi-source alarm trigger factor is generated, which is used for subsequent early warning processing.

[0052] The application further proposes generating the track health degree evaluation index, including: extracting the maximum stress concentration coefficient from the deformation-temperature correlation matrix; calling the stress reference value in the historical track temperature distribution data at the same position; dynamically correcting the stress reference value by using the exponential weighted moving average algorithm; calculating the relative deviation of the current stress concentration coefficient and the corrected reference value; and mapping the health degree evaluation index based on the deviation value, with the index value interval being 0-1.

[0053] In the formula, the maximum stress concentration coefficient is extracted by traversing the stress gradient values of each grid element in the deformation-temperature correlation matrix, and the grid coordinates with the largest gradient change are selected; when the stress reference value in the historical track temperature distribution data is called, the stress records of the same spatial position at different time periods in the edge storage database are accessed; when the exponential weighted moving average algorithm is used, the decay factor parameter is set to control the influence weight of the historical data on the reference value, and the decay factor value range is 0.6-0.9; the relative deviation is calculated by using the percentage difference algorithm, and the current stress concentration coefficient is divided by the dynamically corrected reference value after difference operation; when the health degree evaluation index is mapped, the linear or nonlinear function relationship between the deviation percentage and the index value is established to ensure that the index value monotonically changes in the 0-1 interval.

[0054] Specifically, in the process of generating the track health degree evaluation index, first, the area with the largest stress gradient is located from the deformation-temperature correlation matrix, and the stress concentration coefficient of the area is obtained. When the historical stress reference value at the position is called, the historical data records under the same spatial coordinates are retrieved by time stamp matching. The historical reference value is dynamically corrected by using the exponential weighted moving average algorithm, and by adjusting the decay factor parameter, the recent data obtains a higher weight, thereby reflecting the cumulative effect of temperature change. When the relative deviation of the current stress concentration coefficient and the corrected reference value is calculated, the deviation degree is quantified by using the difference percentage formula. Finally, the deviation value is mapped to the 0-1 interval, when the deviation value increases, the index value tends to 0, indicating that the health degree deteriorates; when the deviation value decreases, the index value tends to 1, indicating that the health degree is good. The mapping process is realized by a preset conversion function, and the function type includes a segmented linear function or an S-shaped curve function, and the specific parameters are set according to the track material characteristics.

[0055] As a preferred embodiment, the scheme of the application is implemented as follows: Generating the track health degree evaluation index includes the following steps: First, the maximum stress concentration coefficient is extracted from the deformation-temperature correlation matrix. Specifically, by scanning each element in the deformation-temperature correlation matrix, the stress concentration coefficient with the largest value is found.

[0056] Secondly, the stress reference value in the historical track temperature distribution data at the same position is called. For example, the average stress value of the position in the past 30 days is obtained from the database as the reference.

[0057] Further, the stress reference value is dynamically corrected by using an exponential weighted moving average algorithm. In specific implementation, the weight factor can be set to 0.2, and the last 5 historical data points are weighted and averaged to obtain the corrected reference value.

[0058] Subsequently, the relative deviation of the current stress concentration coefficient and the corrected reference value is calculated. In this way, the degree of change of the current stress level relative to the historical average level can be quantified.

[0059] Finally, the health degree evaluation index is mapped based on the deviation value, and the index value interval is 0-1. For example, the index value can be set to 1 when the deviation is 0, and the index value is 0 when the deviation is 100%, and the intermediate value is linearly interpolated.

[0060] The application further proposes that the spatial temperature field features of the track temperature distribution data are extracted by a three-dimensional convolution layer; the stress distribution features of the deformation-temperature correlation matrix are extracted by a three-dimensional convolution layer; the spatial temperature field features and the stress distribution features are concatenated at the channel level to generate a fusion feature map; the time evolution law of the fusion feature map is analyzed by using a long short-term memory network; the spatio-temporal features are weighted and fused based on an attention mechanism to generate a state feature vector; and five types of track states and corresponding confidence degrees are output by a fully connected layer classifier.

[0061] Among them, the three-dimensional convolution layer uses convolution kernels of different sizes to perform sliding window calculation on the three-dimensional spatial grid of the track temperature distribution data, to extract the local spatial correlation pattern of the temperature field; when performing convolution operation on the three-dimensional structure of the deformation-temperature correlation matrix, the step parameter is used to control the feature map downsampling ratio, to retain the gradient change information of the stress distribution. The channel concatenation operation concatenates the temperature field feature map and the stress feature map in the channel dimension to form a four-dimensional tensor containing multi-physical field information. The long short-term memory network sets the number of hidden units to match the number of channels of the fusion feature map, and iteratively processes the feature sequence of the continuous monitoring period through time steps. The attention mechanism dynamically weights the spatio-temporal features by using a learnable weight matrix, and focuses on the correlation relationship between the temperature mutation area and the stress concentration area. The fully connected classifier sets five output nodes, respectively corresponding to different track failure risk levels, and each node output value is normalized and converted into a confidence probability.

[0062] Specifically, the three-dimensional convolution kernel of the rail temperature distribution data slides along the longitude, latitude and temperature gradient directions to capture the temperature change trend between adjacent grids. In the three-dimensional convolution operation of the deformation-temperature correlation matrix, the depth of the convolution kernel is aligned with the time dimension of the matrix to extract the time evolution law of the stress distribution. The fused feature map after channel concatenation is unfolded in the long short-term memory network according to the time step, and the hidden state of each time step carries historical feature information. The attention weight matrix automatically learns the correlation strength of the temperature field and the stress field through the training process, and suppresses the interference of the noise region. The cross-entropy loss function is used in the full connection layer to optimize the classification boundary, and the confidence output reflects the certainty degree of the model to the prediction result. Through the multi-level feature extraction and fusion mechanism, the spatio-temporal correlation modeling capability of the track state prediction is significantly improved.

[0063] As a preferred embodiment, the scheme of the application is implemented as follows: The operation of the deep learning track state prediction module includes the following steps: First, the spatial temperature field features of the rail temperature distribution data are extracted through the three-dimensional convolution layer. Specifically, a 3x3x3 convolution kernel is used to convolve the rail temperature distribution data to extract the local features of the temperature field. The number of convolution layers is set to 3, and the output channel numbers of each layer are 32, 64 and 128 respectively.

[0064] Secondly, the stress distribution features of the deformation-temperature correlation matrix are extracted through the three-dimensional convolution layer. Similarly, a 3x3x3 convolution kernel is used to convolve the correlation matrix to extract the local features of the stress distribution. The number of convolution layers is also set to 3, and the output channel numbers of each layer are 32, 64 and 128 respectively.

[0065] Next, the spatial temperature field features and the stress distribution features are concatenated in the channel to generate a fused feature map. The specific operation is to concatenate the two feature maps in the channel dimension to obtain a 256-channel fused feature map.

[0066] Then, the long short-term memory network is used to analyze the time evolution law of the fused feature map. An LSTM layer containing 128 hidden units is used to perform time series analysis on the fused feature map to capture the law of feature change over time.

[0067] Further, the spatio-temporal features are weighted and fused based on the attention mechanism to generate a state feature vector. The self-attention mechanism is used to calculate the correlation of features at different spatio-temporal positions and perform weighted fusion to obtain a 512-dimensional state feature vector.

[0068] Finally, the five track states and corresponding confidence are output by the fully connected layer classifier. Two fully connected layers are used, with 256 neurons in the first layer and 5 neurons in the second layer corresponding to the five track states. The probability of each state is calculated by the softmax function as the final prediction result and confidence.

[0069] The application further proposes a data expansion process before inputting the time series multi-source fusion data into the deep learning track state prediction module, including: monitoring the number of consecutive periods in which the track health evaluation index continuously exceeds the warning threshold; when the number of consecutive over-limit periods reaches a set value, automatically triggering the data expansion mechanism; retrieving a complete data set of a set number of historical monitoring periods before the current time point from the edge storage database, which includes track geometric deformation data and corresponding track temperature distribution data; reading track temperature distribution data of a set number of real-time monitoring periods after the current time point from the real-time data buffer; performing time dimension alignment operation on the historical data set and the real-time data set, and using timestamp matching method to unify the data sampling frequency; performing spatial grid filling on the aligned data, and using the temperature gradient change rate of adjacent grids to interpolate and complete the missing area data; constructing an expanded time series data set containing historical and real-time data, with a time span greater than the original monitoring period. The data expansion process ensures the timeliness of the expansion operation through the continuous period threshold triggering mechanism, avoiding the accumulation of invalid data. The historical data calling range is controlled by a set number of parameters, for example, calling the historical data of the previous three periods, and the real-time data calling range can be set to the data of the next two periods. Time dimension alignment uses a timestamp matching algorithm to unify data with different sampling frequencies to the same time reference, for example, aligning every minute sampling data with every five minute sampling data to five minute intervals. Spatial grid filling uses adjacent grid temperature gradient interpolation, for example, when a grid data is missing, the missing value is calculated according to the temperature change rate of the four adjacent grids before and after it. The time span of the expanded time series data set is realized by superimposing historical and real-time data, for example, the original monitoring period is seven days, and the expanded data set can cover ten days. Specifically, when the track health evaluation index exceeds the early warning threshold for three consecutive periods, the data expansion mechanism is automatically activated. The edge storage database retrieves the historical data of the previous two periods, including complete geometric deformation and track temperature distribution records. The real-time data buffer reads the track temperature data of the next three periods to form the data set to be processed. The timestamp matching operation unifies the hourly sampling of historical data and the half-hourly sampling of real-time data to an hourly frequency, and fills in the intermediate time point data through linear interpolation. The spatial grid filling fills in the missing areas in the track temperature distribution data, and calculates the interpolation weight according to the temperature gradient change rate of the adjacent grid, for example, the filling value of a missing grid is calculated by weighting the temperature rise rate of the left grid and the temperature drop rate of the right grid. The finally constructed expanded data set covers five consecutive periods, with a time span of twice the original monitoring period, providing more comprehensive time series features for the deep learning model and enhancing the accuracy of track state prediction.

[0070] As a preferred embodiment, the scheme of the application is implemented as follows: Before inputting the time series multi-source fusion data into the deep learning track state prediction module, the data expansion process is performed. First, the number of consecutive periods in which the track health evaluation index exceeds the early warning threshold is monitored. When the number of consecutive over-limit periods reaches 10, the data expansion mechanism is automatically triggered. The edge storage database retrieves the complete data set of the previous 30 historical monitoring periods at the current time point, which includes track geometric deformation data and corresponding track temperature distribution data. The real-time data buffer reads the track temperature distribution data of the next 5 real-time monitoring periods at the current time point. The historical data set and the real-time data set are aligned in the time dimension, and the timestamp matching method is used to unify the data sampling frequency to every 10 minutes. The spatial grid filling is performed on the aligned data, and the missing area data is interpolated and completed using the temperature gradient change rate of the adjacent grid. Finally, the expanded time series data set containing historical and real-time data is constructed, with a time span of 35 monitoring periods, which is greater than the original monitoring period.

[0071] The application further proposes to dynamically configure the preset degradation condition according to the operation scenario: in the summer high temperature period, when the single period decline of the track health evaluation index exceeds the high temperature scenario set threshold, the degradation determination is triggered; in the heavy load freight line monitoring period, when the cumulative decline of the track health evaluation index in three consecutive periods exceeds the heavy load scenario set threshold, the degradation determination is triggered; for the busy main line with annual total weight exceeding the standard, when the track health evaluation index continuously falls below the safety threshold for a set number of consecutive periods, the degradation determination is triggered; in the curve section with a radius less than a set value, when the range value of the track health evaluation index in a set length range exceeds the curve scenario set threshold, the degradation determination is triggered. The determination condition of dynamic configuration contains four types of typical operation scene difference threshold logic: single period descending amplitude threshold for high temperature scene, multi-period cumulative descending threshold for heavy load scene, continuous period number threshold for busy trunk line, and range threshold for curve section. The threshold parameters of different scenes are independently stored in the configuration database, and the corresponding determination rule is automatically matched and activated according to the real-time operation data at the beginning of the monitoring period. The trigger condition is dynamically compared with the calculation result of the track health evaluation index, and the deterioration determination signal is generated immediately when the threshold logic of any scene is met. Specifically, during the summer high temperature period, the track material thermal expansion effect is significant, and the single period mutation of the health index may reflect the potential stress concentration, and the high temperature scene setting threshold is set to realize rapid response; in the heavy load freight line, the multi-period cumulative descending threshold can effectively identify the progressive damage caused by continuous load; in the busy trunk line scene, the continuous period number threshold is used to capture the material fatigue caused by long-term overload operation; and the range threshold of the curve section is for the local stress distribution anomaly caused by the change of track curvature. The determination logic works cooperatively with the calculation process of the track health evaluation index, when the system detects the current operation scene type, the corresponding threshold parameter is automatically called to compare with the real-time index, the deterioration determination is triggered, and the trigger scene type code is recorded to provide classification basis for subsequent disposal decision.

[0072] As a preferred embodiment, the scheme of the application is implemented as follows: The preset deterioration condition is dynamically configured according to the operation scene: During the summer high temperature period monitoring period, when the single period descending amplitude of the track health evaluation index exceeds the high temperature scene setting threshold, the deterioration determination is triggered. For example, the high temperature scene threshold is set to 0.1, when the track health evaluation index decreases by more than 0.1 in a single monitoring period, the system automatically triggers the deterioration determination process.

[0073] During the heavy load freight line monitoring period, when the cumulative descending value of the track health evaluation index in three consecutive periods exceeds the heavy load scene setting threshold, the deterioration determination is triggered. Specifically, the heavy load scene threshold can be set to 0.15, the system continuously monitors the change of the track health evaluation index in three consecutive periods, and when the cumulative descending value exceeds 0.15, the deterioration determination is triggered.

[0074] For the busy trunk line with excessive annual total weight, when the track health evaluation index is continuously lower than the safety threshold for a set number of consecutive periods, the deterioration determination is triggered. For example, the safety threshold is set to 0.7, and the number of consecutive periods is 5, when the track health evaluation index is lower than 0.7 for 5 consecutive periods, the system determines that the track enters the deterioration state.

[0075] In the curve section with a curve radius less than a set value, the deterioration determination is triggered when the range of the track health evaluation index in the set length range exceeds the set threshold value of the curve scene. Further, the curve radius set value can be set to 300 meters, the set length range is 100 meters, and the curve scene threshold value is 0.2. The system calculates the difference between the maximum value and the minimum value of the track health evaluation index in the range of 100 meters, and when the range exceeds 0.2, the deterioration determination is triggered.

[0076] The application further proposes that when the risk level is determined as level four or level five, the acoustic sensor array installed on the rail waist is automatically awakened; the original signal of rail vibration is collected and signal preprocessing is performed, and the wavelet transform algorithm is used to filter out the interference of locomotive running noise; three types of feature parameters are extracted from the denoised vibration signal: base frequency amplitude decay rate, third harmonic distortion, and spectral energy entropy value; the feature parameter combination is input into the pre-trained rail crack recognition model, which is constructed based on deep neural network; the output is the implicit crack propagation probability value and the fracture risk index of the rail, wherein the fracture risk index is a continuous value of 0-100%; the corresponding disposal level instruction is generated according to the preset interval of the fracture risk index.

[0077] Among them, the acoustic sensor array is installed along the rail waist at equal intervals, and the sensor spacing is set as an integer multiple of the sleeper spacing. The vibration signal sampling frequency is set to 40 kHz, covering the base frequency to the third harmonic frequency band of the rail vibration. The wavelet transform uses db4 wavelet basis function for five-layer decomposition, and sets the threshold to filter out low-frequency locomotive running noise below 200 Hz. The base frequency amplitude decay rate is obtained by calculating the ratio of the current signal base frequency amplitude to the historical reference value, the third harmonic distortion is calculated by the total harmonic distortion formula, and the spectral energy entropy value is obtained by calculating the Shannon entropy of the signal power spectrum in the 1-10 kHz frequency band. The rail crack recognition model adopts a three-layer convolutional neural network and a two-layer fully connected layer structure, the input layer receives a vector composed of three types of feature parameters, and the output layer generates a fracture risk index through a sigmoid function.

[0078] Specifically, when the deep learning model determines that the track failure risk level reaches level four or five, the acoustic verification process is triggered. An acoustic sensor array installed on the rail waist collects rail vibration signals at a sampling rate of 40 kHz. The original signals are decomposed by a db4 wavelet basis function for five layers to filter out low-frequency noise generated by the operation of the locomotive. The denoised signals are transformed into frequency spectra by fast Fourier transform, the base frequency amplitude attenuation rate is calculated to reflect the change of rail structure stiffness, the third harmonic distortion degree represents the degree of material nonlinear distortion, and the spectral energy entropy quantifies the complexity of vibration energy distribution. The three characteristic parameters are combined to input a pre-trained deep neural network model, which extracts local feature correlations through convolution layers and fuses global features through fully connected layers, and finally outputs crack propagation probability and fracture risk index. The fracture risk index is divided into three levels according to the preset interval, when the index exceeds 70%, a first-level speed limit command is triggered, when it exceeds 85%, a maintenance work order is generated, and when it exceeds 95%, the driving permission is forcibly blocked. The verification mechanism effectively reduces the misjudgment probability of a single data source through cross-validation of physical vibration characteristics and deep learning prediction results.

[0079] As a preferred embodiment, the scheme of the application is implemented as follows: when the track failure risk level is determined to be level four or five, an acoustic sensor array installed on the rail waist is automatically awakened, which includes piezoelectric vibration sensors arranged every 30 meters along the longitudinal direction of the rail. The amplitude of the original rail vibration signal is normalized by a charge amplifier, and then decomposed and reconstructed by a Daubechies wavelet basis function for five layers to filter out the frequency band components of wheel-rail impact noise. The denoised vibration signal is segmented into 500ms time domain segments, and the frequency spectrum distribution is calculated by fast Fourier transform. The base frequency amplitude attenuation rate is obtained by comparing the ratio of the current segment base frequency amplitude to the historical normal state reference value, the third harmonic distortion degree is calculated using the total harmonic distortion formula, and the spectral energy entropy is quantified by the Shannon entropy formula for the normalized frequency spectrum energy distribution. The characteristic parameter combination is input into a rail crack identification model based on the ResNet-50 architecture, which uses an acoustic database containing one hundred thousand crack samples for transfer learning in the training stage. The model output includes the probability value of rail internal transverse crack propagation and the fracture risk index of longitudinal crack, and the fracture risk index is represented as a continuous value of 0-100%. When the fracture risk index is in the interval of 20%-40%, a yellow warning command is generated; when it is in the interval of 40%-60%, a second-level speed limit command is triggered; and when it exceeds 60%, the line blocking process is started.

[0080] The application further proposes an execution process of hierarchical treatment instructions, which includes: when the fracture risk index is in a first preset interval, sending a first-level speed limit instruction to a train dispatching control system, requiring the speed of a train passing through the section to be reduced to a set value; when the fracture risk index is in a second preset interval, automatically generating a red warning maintenance work order with geographic coordinates, pushing it to a maintenance section emergency terminal and starting a countdown treatment process; when the fracture risk index exceeds a preset fracture threshold, triggering a railway signal blocking system control protocol to forcibly block the driving right of the affected section; simultaneously activating a high-definition trackside camera group to shoot steel rail surface images at a set frame rate per second; running a steel rail surface defect detection algorithm to identify crack, fracture, and chunk physical damage features in the images; and cross-verifying the image recognition results with the acoustic detection conclusions to confirm the final treatment instruction when they are consistent.

[0081] The first preset interval corresponds to a fracture risk index in the range of 20%-40%, and the set speed limit value is 60km / h; the second preset interval corresponds to a fracture risk index in the range of 40%-70%, and the countdown treatment process is set to a 2-hour response time limit; the preset fracture threshold is set to 70%, and the interlocking control instruction of the signal blocking system is triggered. The trackside high-definition camera group adopts a multi-angle arrangement, and the frame rate is set to 30fps to ensure dynamic capture accuracy. The steel rail surface defect detection algorithm is based on the YOLOv5 model, and the training set contains 100,000 labeled samples, with an identification accuracy of 98.6%. The cross-verification uses logical AND operation, and when the fracture risk index of acoustic detection and the physical damage features identified by image recognition both meet the preset conditions, a treatment instruction confirmation signal is generated.

[0082] Specifically, after the fracture risk index is divided by the preset interval, the hierarchical triggering mechanism is used to realize the optimal allocation of treatment resources. The first-level speed limit instruction is transmitted to the train-mounted control system through the wireless communication module to dynamically adjust the running speed of the train. The red warning maintenance work order is positioned by the geographic information system of the maintenance management platform, and is automatically assigned to the nearest maintenance team. The control protocol of the signal blocking system needs to complete the interlocking state verification with the adjacent section before execution to ensure the safety of the driving right change. The activation instruction of the trackside high-definition camera group is triggered synchronously with the risk level determination result, and the image acquisition time window is set to the five-minute period after risk determination. The surface defect detection algorithm uses a sliding window scanning mechanism to identify the image areas on the top and side surfaces of the steel rail. In the cross-verification link, the data fusion server compares the spectral features of acoustic detection and the morphological features of image recognition, and when there is a matching relationship in spatial position and damage degree, a treatment instruction execution confirmation signal is generated to avoid the risk of misjudgment by a single detection method.

[0083] As a preferred embodiment, the scheme of the present application is implemented as follows: when the fracture risk index is in the interval of 20%-40%, a speed limit instruction is sent to the train dispatching control system, requiring the train running speed through the corresponding section to be reduced to 45 kilometers per hour. When the fracture risk index is in the interval of 40%-60%, a red warning maintenance work order containing geographic coordinates is automatically generated, which is pushed to the emergency terminal of the track maintenance section through the special communication network of the railway, and at the same time, a 72-hour countdown disposal process is started. When the fracture risk index exceeds 85%, the interlocking control protocol of the railway signal blocking system is triggered, the signal of the affected section is forcibly closed and the switch state is locked, and the train is prohibited from entering the section. The line array high-definition camera group deployed on the track side is activated synchronously to collect a sequence of rail surface images at a rate of 25 frames per second, and a surface defect detection algorithm based on convolutional neural network is run on an edge computing device to identify transverse cracks, longitudinal fractures and rail head block features in the images. The image recognition result and the crack propagation probability detected by the acoustic sensor array are logically ANDed, and when both are determined to exist crack propagation, the final disposal instruction is confirmed to take effect. The core operations performed by the edge computing gateway further include: receiving the original deformation data stream uploaded by the track detector array in real time, using a sliding window filtering algorithm to eliminate outliers and smooth data fluctuations; synchronously processing the temperature data uploaded by the track temperature sensor array, and performing real-time compensation for temperature drift based on the sensor calibration parameters; executing a deformation anomaly section extraction algorithm on local computing resources, dynamically adjusting the time window length to adapt to changes in train operation density; constructing a deformation-temperature matrix associated with space and time, and using distributed memory computing technology to accelerate the matrix operation process; calculating the track health evaluation index according to the monitoring period, and compressing and storing the historical data of the index to the local solid state storage; loading a lightweight version of the deep learning track state prediction model to directly perform risk level prediction calculation on the edge; according to the prediction result, it is decided whether to trigger the acoustic verification and graded disposal process, and the delay of cloud data transmission is reduced.

[0084] Among them, the sliding window filtering algorithm sets a fixed length data window, performs moving average calculation on the continuously incoming deformation data, and eliminates transient noise interference. The temperature drift compensation uses a preset sensor calibration curve to linearly correct the original temperature data. The time window length is automatically adjusted according to the time interval of adjacent trains, and the window length is shortened when the running density is high to improve the processing frequency. The distributed memory computing technology decomposes the deformation-temperature matrix into multiple sub-matrices, and completes the matrix operation through parallel threads. The lightweight prediction model compresses the neural network size through model pruning and quantization technology to adapt to the edge computing resources.

[0085] Specifically, the deformation data stream generated by the track detector array is transmitted to the edge gateway through the wireless communication module, and the sliding window filtering algorithm is used to smooth the data stream at a fixed sampling frequency to eliminate abnormal fluctuations caused by track vibration. The temperature sensor data is compensated for drift at the time of reception, and the stored calibration parameters are used to correct the influence of environmental temperature on sensor accuracy. The deformation anomaly section extraction algorithm dynamically adjusts the analysis window according to the real-time train operation density, using a short window for fast processing during peak hours and a long window for improved analysis accuracy during low peak hours. When constructing the deformation-temperature matrix, the distributed memory computing engine divides the sensor spatial grid into multiple computing units and performs matrix filling and operation operations in parallel. After calculating the health assessment index, the differential compression algorithm is used to store the historical data to the local solid state disk. The lightweight prediction model is loaded into the dedicated computing chip of the edge gateway, which directly processes the preprocessed data in real time. When the prediction result reaches the preset risk threshold, the edge gateway immediately starts the acoustic sensor verification process to avoid cloud decision delay.

[0086] As a preferred embodiment, the scheme of the application is implemented as follows: the edge computing gateway receives the original deformation data stream uploaded by the track detector array in real time, and uses the sliding window filtering algorithm to preprocess the data stream, wherein the width of the sliding window is automatically adjusted according to the train running speed, and the data segment of a certain length is continuously intercepted in the data stream for mean filtering processing, and the outlier points deviating from the three times standard deviation of the median value are simultaneously removed. The original temperature data collected by the track temperature sensor array is input into the edge computing gateway in real time, and combined with the temperature-resistance curve parameters calibrated at the factory for each sensor, zero drift compensation and gain calibration are performed in the data preprocessing stage to eliminate the influence of environmental temperature on the measurement accuracy of the sensor.

[0087] The deformation anomaly section extraction algorithm runs in the embedded processor of the edge computing gateway, and the length of the time window of the algorithm is dynamically adjusted according to the time interval between the adjacent trains passing through the section, specifically by analyzing the train operation diagram data in the last twenty-four hours to calculate the average train tracking interval time as the window reference length. In the construction process of the deformation-temperature matrix related to space and time, the distributed memory database is used to store the matrix data in blocks, and each computing node is responsible for processing the data block of the corresponding spatial grid, and the data synchronization and merging operations between nodes are realized through the message queue. The calculation results of the track health assessment index are updated at a minute level, and the historical index data is losslessly compressed using the LZ4 compression algorithm and stored in the local solid state disk.

[0088] A lightweight version of the deep learning track state prediction model is achieved through model pruning techniques, in which the number of channels of the three-dimensional convolution layer is reduced to thirty percent of the original model, and the hidden layer dimension of the long short-term memory network is reduced to fifty percent of the original size. The prediction model runs in the GPU coprocessor of the edge computing gateway. When the prediction result reaches the fourth or fifth risk level, the wake-up instruction is immediately sent to the trackside acoustic sensor array, and the preloaded acoustic feature extraction algorithm in the local storage cache area is started.

[0089] Through the above technical solutions, the present application realizes real-time local processing and intelligent decision-making of track monitoring data, effectively reducing the delay problem caused by cloud data transmission. Through the sliding window filtering and dynamic window adjustment mechanism, the data preprocessing quality under different train operation densities is guaranteed, and the problem of insufficient adaptability of traditional fixed parameter algorithms in complex operation scenarios is solved. Distributed memory computing technology is used to accelerate matrix operations, which significantly improves the processing efficiency of large-scale spatio-temporal data, and shortens the calculation period of the health degree evaluation index to seconds. The edge deployment of the lightweight prediction model realizes real-time online evaluation of track failure risk, providing technical support for the rapid start of the subsequent grading disposal process.

[0090] The above only describes the embodiments of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for periodic intelligent monitoring of railway track based on Internet of Things, characterized in that, The method comprises the following steps: acquiring track monitoring data collected by sensing nodes deployed on the trackside and track, including track geometry deformation data collected by a track detector array and track temperature distribution data collected by a track temperature sensor array; extracting abnormal sections from the track geometry deformation data based on time window segmentation, screening out sections with deformation values exceeding a set threshold within a continuous monitoring period, and generating deformation abnormal section data; mapping the deformation abnormal section data to the spatial grid corresponding to the track temperature distribution data to generate a deformation-temperature correlation matrix; calculating the track stress change gradient based on the deformation-temperature correlation matrix and historical track temperature distribution data to generate a track health degree evaluation index; when the track health degree evaluation index meets the preset degradation condition, inputting the constructed time series multi-source fusion data into a deep learning track state prediction module to output a track failure risk level; wherein the time series multi-source fusion data includes track temperature distribution data and deformation-temperature correlation matrix in a continuous monitoring period.

2. The method of claim 1, wherein, The method further comprises the following steps for generating the deformation-temperature correlation matrix: performing Gaussian-Krueger projection conversion on the deformation abnormal section data to calculate the geodetic coordinates of the abnormal sections; performing bilinear interpolation based on the grid resolution of the track temperature sensor array to determine the mapping unit of the deformation section in the track temperature grid; constructing an empty matrix with the same dimension as the track temperature distribution data to initialize the deformation-temperature correlation matrix; filling the temperature gradient values at the mapping unit coordinates to generate a weighted correlation matrix; setting the un-mapped units in the correlation matrix to invalid values; wherein the track detector array and the track temperature sensor array are distributedly deployed at a fixed interval.

3. The method of claim 2, wherein, After generating the deformation-temperature correlation matrix, the method further comprises the following steps: extracting areas with temperature gradient changes exceeding a set rate from the correlation matrix to generate a track temperature abnormality mask; analyzing the second derivative of the track geometry deformation data to screen out sections with positive deformation acceleration; performing a morphological dilation operation on the deformation acceleration section to generate a deformation trend mask; calculating the spatial overlap area ratio of the track temperature abnormality mask and the deformation trend mask; when the overlap area ratio exceeds a set proportion, generating a multi-source alarm trigger factor.

4. The method of claim 1, wherein, The method further comprises the following steps for generating the track health degree evaluation index: extracting the maximum stress concentration coefficient from the deformation-temperature correlation matrix; calling the stress reference value in the historical track temperature distribution data at the same location; dynamically correcting the stress reference value using an exponential weighted moving average algorithm; calculating the relative deviation of the current stress concentration coefficient and the corrected reference value; mapping the health degree evaluation index based on the deviation value, with an index value interval of 0-1.

5. The method of claim 1, wherein, The deep learning track state prediction module performs the following operations: extracting the spatial temperature field features of the track temperature distribution data through a three-dimensional convolution layer; extracting the stress distribution features of the deformation-temperature correlation matrix through a three-dimensional convolution layer; concatenating the spatial temperature field features and the stress distribution features at the channel level to generate a fusion feature map; analyzing the time evolution law of the fusion feature map using a long short-term memory network; generating a state feature vector based on attention mechanism weighted fusion of spatio-temporal features; outputting five types of track states and corresponding confidence levels through a fully connected layer classifier.

6. The method of claim 5, wherein, Before inputting the time series multi-source fusion data into the deep learning track state prediction module, the method further comprises the following steps: The number of periods in which the track health evaluation index continuously exceeds the warning threshold is monitored. When the number of consecutive over-limit periods reaches a set value, the data expansion mechanism is automatically triggered. The complete data set of a set number of historical monitoring periods before the current time point is called from the edge storage database. The data set includes track geometric deformation data and corresponding rail temperature distribution data. The rail temperature distribution data of a set number of real-time monitoring periods after the current time point is read from the real-time data buffer. The historical data set and the real-time data set are time-dimensionally aligned. The data sampling frequency is unified using timestamp matching. Spatial grid interpolation is performed on the aligned data. The temperature gradient change rate of adjacent grids is used for interpolation to complete the missing area data. An expanded time series data set containing historical and real-time data is constructed. The time span is greater than the original monitoring period.

7. The method of claim 1, wherein, The preset degradation conditions are dynamically configured according to the operating scenario: In the summer high-temperature period, if the single-period decline of the track health evaluation index exceeds the high-temperature scenario set threshold, the degradation is determined. In the heavy-load freight line monitoring period, if the cumulative decline of the track health evaluation index in three consecutive periods exceeds the heavy-load scenario set threshold, the degradation is determined. For busy mainlines with annual total weight exceeding the standard, if the track health evaluation index continuously falls below the safety threshold for a set number of consecutive periods, the degradation is determined. In curve sections with a radius less than a set value, if the range of the track health evaluation index in a set length exceeds the curve scenario set threshold, the degradation is determined.

8. The method of claim 5, wherein, After outputting the track failure risk level, acoustic verification is also included: When the risk level is determined to be level four or five, the acoustic sensor array installed on the rail waist is automatically awakened. The original vibration signal of the steel rail is collected and preprocessed. The wavelet transform algorithm is used to filter out the interference of locomotive running noise. Three types of feature parameters are extracted from the denoised vibration signal: fundamental frequency amplitude decay rate, third harmonic distortion, and spectral energy entropy value. The feature parameters are combined and input into a pre-trained steel rail crack identification model based on a deep neural network. The internal implicit crack propagation probability value and the fracture risk index are output, with the fracture risk index being a continuous value from 0 to 100%. According to the preset interval of the fracture risk index, the corresponding disposal level instruction is generated.

9. The method of claim 8, wherein, The disposal level instruction includes: When the fracture risk index is in the first preset interval, a first-level speed limit instruction is sent to the train dispatching control system, requiring the train speed through the section to be reduced to a set value. When the fracture risk index is in the second preset interval, a red warning maintenance work order with geographic coordinates is automatically generated and pushed to the maintenance section emergency terminal and the countdown disposal process is started. When the fracture risk index exceeds the preset fracture threshold, the railway signal blocking system control protocol is triggered, forcibly blocking the affected section. The high-definition camera group on the side of the track is activated to take images of the steel rail surface at a set frame rate per second. The steel rail surface defect detection algorithm is run to identify crack, fracture, and chunk physical damage features in the images. The image recognition result is cross-verified with the acoustic detection conclusion. When they are consistent, the final disposal instruction is confirmed.

10. The method of claim 1, wherein, The method also includes: Real-time receive the original deformation data stream uploaded on the track detector array, use the sliding window filtering algorithm to eliminate outliers and smooth the data fluctuations; Synchronously process the temperature data uploaded on the track temperature sensor array, and perform real-time compensation for temperature drift based on the sensor calibration parameters; Perform the deformation anomaly segment extraction algorithm on the local computing resources, and dynamically adjust the time window length to adapt to the train operation density changes; Construct the deformation-temperature matrix associated with space and time, and use the distributed memory computing technology to accelerate the matrix operation process; Calculate the track health degree evaluation index according to the monitoring period, and compress and store the historical data of the index to the local solid-state storage; Load the lightweight version of the deep learning track state prediction model, and directly perform the risk level prediction calculation on the edge side; According to the prediction result, decide whether to trigger the acoustic verification and hierarchical disposal process, and reduce the cloud data transmission delay.

Images