A highway subgrade deep water content distributed monitoring system

Abstract

The application relates to the technical field of highway engineering roadbed safety monitoring, and particularly discloses a highway roadbed deep water content distributed monitoring system, which comprises a heterogeneous prior exploration layer module, which is used for actively acquiring the dielectric constant continuous profile of the target highway roadbed surface to deep layer by using a ground penetrating radar, and generating a three-dimensional heterogeneity weight atlas of a soil-rock mixed interface through inversion and spatial interpretation; the dielectric constant continuous profile of the roadbed surface to deep layer is actively acquired by the ground penetrating radar, the three-dimensional heterogeneity weight atlas is generated through inversion and spatial interpretation, and then the improved fruit fly algorithm is driven to perform sensor coordinate optimization, so that the embedding position of each intelligent sensor and the physical properties of the local soil layer where the intelligent sensor is located are best matched in space, the sampling deviation problem caused by the traditional equidistant arrangement mode falling in a non-representative soil area is avoided, and the representativeness of the monitoring data is ensured.

Description

Technical Field

[0001] This invention relates to the field of roadbed safety monitoring technology in highway engineering, and in particular to a distributed monitoring system for deep moisture content of highway roadbed. Background Technology

[0002] With the continuous growth of transportation demand, highway engineering is facing increasingly severe challenges. Especially under climate change and extreme weather conditions, changes in soil moisture content will directly affect the physical properties and bearing capacity of the roadbed. Traditional monitoring methods can often only provide moisture data for local areas, which is difficult to reflect the overall roadbed condition and cannot meet the needs of modern highway engineering for real-time and comprehensive monitoring.

[0003] In actual roadbed construction, layers of gravel and clay are often mixed. The traditional monitoring method of deploying sensors at equal intervals tends to result in non-representative soil areas, causing a mismatch between the initial spatial sampling points and the variability of deep soil. This directly leads to a fundamental deviation in the representativeness of subsequent data collection. Furthermore, the measurement process is susceptible to diurnal temperature variations, resulting in significant differences in the thermal response of adjacent sensors. This causes the time-series water content data to exhibit inverted jumps, thereby disrupting the original continuity of deep longitudinal migration patterns and leading to a break in the spatiotemporal correlation information. In addition, when a sudden surge in water content occurs at a certain depth, it is difficult to distinguish the early warning modes of upper layer infiltration recharge, lateral undercurrent intrusion, or internal piping initiation at the moment the signal appears. This results in ambiguous early warning attribution and limits the effective conversion of monitoring data into monitoring and early warning signals. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a distributed monitoring system for deep moisture content of highway subgrade, which can effectively solve the problems involved in the prior art.

[0005] The objective of this invention can be achieved through the following technical solution: This invention provides a distributed monitoring system for deep water content of highway subgrade, comprising the following modules: The heterogeneous prior exploration module is used to actively acquire the continuous dielectric constant profile of the target highway subgrade from shallow to deep layers using ground penetrating radar. After inversion and spatial interpretation, a three-dimensional heterogeneity weight map of the soil-rock interface is generated, providing a prior physical field for the layout of measurement points and ensuring that the sampling points can actively adapt to the spatial variability of the soil. The optimal deployment module, based on the three-dimensional heterogeneity weight map, utilizes the improved fruit fly algorithm's odor concentration determination mechanism and search step size strategy to search for the optimal burial coordinates of each smart sensor in the three-dimensional roadbed space that best match the local soil physical properties. This enables the sampling point community to actively adapt to the spatial variation of soil and ensures that the initial measurement point deployment has longitudinal representativeness. The temperature disturbance time series correction module is used to synchronously input the original moisture content time series and temperature data collected by each smart sensor into the evolution correction framework that integrates LSTM and knowledge graph. The former learns the local fluctuation mode under temperature and humidity coupling, while the latter performs topological logic verification and interpolation on abnormal jump data according to the heat conduction and moisture transport constraint rules between measuring points, restores the continuous evolution path of longitudinal time series data, and outputs corrected continuous time series data, eliminating the time series inversion jump caused by diurnal temperature difference and restoring the true evolution law of longitudinal migration of roadbed moisture. The probability inversion and deduction module is used to feed the corrected continuous water content sequence into a dynamic Bayesian network to construct a probability transfer chain and time slice transition matrix between multiple depth measurement points. Using the temporal sequence and amplitude gradient of water content surges at each node as observation evidence, it reversely calculates the source region probability field and propagation direction vector of the surge signal source region, and reversely traces the source location and diffusion path of the surge signal, providing a probabilistic basis for quickly locating the abnormal source region. The pattern discrimination and early warning module is used to perform probability matching between the source region probability field and propagation direction vector and the typical topological features of three early warning modes, output the maximum likelihood attribution label, and generate corresponding early warning signals and spatial pointing signals. The early warning modes include upper layer infiltration recharge, lateral undercurrent intrusion and internal piping initiation, so as to realize accurate attribution classification and targeted early warning of water anomaly events and improve the pertinence of emergency response.

[0006] Preferably, in the heterogeneous prior exploration module, the process of generating a three-dimensional heterogeneity weight map of the soil-rock interface is as follows: Based on the ground penetrating radar, multi-band radar waves are emitted to the shallow to deep layers of the target highway subgrade. Full waveform reflection echo data are collected, and after offset superposition and dielectric constant inversion, an initial dielectric constant profile continuously distributed along the longitudinal direction is generated, realizing the continuous characterization of the dielectric properties inside the subgrade and providing a high-resolution data foundation for subsequent soil stratification. Automatic layer picking and soil-rock interface identification are performed on the initial dielectric constant profile. The profile data is discretized along the horizontal direction at a preset interval and layered at equal intervals in the vertical depth to construct a three-dimensional regular grid framework covering the entire subgrade space. Then, sequential Gaussian simulation is used to interpolate and fill the dielectric constant values ​​of each profile sampling point into each grid cell of the three-dimensional regular grid framework, forming a three-dimensional continuous field of dielectric constant covering the entire subgrade space, so that each spatial location has a quantifiable physical property value. Based on the filled three-dimensional regular grid framework, the mean, variance and coefficient of variation of the dielectric constant in each grid cell are calculated. Weighting factors reflecting the degree of soil-rock mixing and local heterogeneity are assigned according to the three factors, and a three-dimensional heterogeneity weight map covering the entire subgrade space is output. This quantifies the spatial distribution of soil heterogeneity in the entire space and provides clear search guidance and evaluation basis for intelligent measurement point layout.

[0007] Preferably, the determination and optimization deployment module includes a fruit fly optimization search unit and a spatial matching verification unit; The fruit fly optimization search unit is used to encode the soil representative radius and coefficient of variation of each grid in the three-dimensional heterogeneity weight map as odor concentration determination factors, and to iteratively solve the sensor coordinate set in the three-dimensional roadbed space by using the improved fruit fly olfactory search step size and visual aggregation rules to obtain the three-dimensional layout coordinates of each smart sensor, so as to realize the active tracking and adaptive matching of the measurement point layout to the spatial variation of soil. The spatial matching verification unit is used to verify the consistency of the neighboring soil quality of the searched three-dimensional layout coordinates, calculate the deviation of the dielectric constant variance from the average value in the area surrounding the coordinates, and eliminate invalid candidate points. The invalid candidate points are coordinate points that fall on the boundary of soil abrupt change or the edge of heterogeneous interlayer. This ensures that the physical properties of the final buried location match the preset sensor range, eliminates the risk of data inaccuracy caused by the measuring point falling on the soil interface, and improves the confidence of the monitoring data.

[0008] Preferably, in the fruit fly optimization search unit, the process of obtaining the three-dimensional layout coordinates of each intelligent sensor is as follows: The three-dimensional heterogeneity weight map is read, and the soil representative radius and coefficient of variation of each grid cell are mapped to the two-component encoding of the odor concentration determination factor. The random distribution position of the fruit fly population in the three-dimensional roadbed space is initialized, and the soil properties are transformed into a searchable concentration field, thus establishing a unified optimization benchmark for swarm intelligence search. Based on the current position of each fruit fly individual, the odor concentration determination factor of the corresponding grid unit is extracted. The olfactory search step size with adaptive decay with the iteration number is adopted to drive the fruit fly group to perform multi-directional local exploration in three-dimensional space to approach the high concentration area. It takes into account the breadth of the global search in the early stage and the depth of the local fine search in the later stage to ensure that no key areas with high heterogeneity are missed. When multiple fruit fly individuals gather in the same high-concentration neighborhood, the visual aggregation rule is activated and the repulsive potential energy constraint is introduced to iteratively update the spatial distribution of the population until the convergence condition is met. Then, the three-dimensional layout coordinates of each smart sensor are output to prevent the population from gathering in a local suboptimal area too early and to ensure that each measuring point is evenly distributed in the most representative soil area.

[0009] Preferably, in the spatial matching verification unit, the process of eliminating invalid candidate points is as follows: The three-dimensional deployment coordinates of each smart sensor are obtained, a spherical neighborhood is delineated with each coordinate as the center, the original numerical sequence of dielectric constants of all grid cells in the spherical neighborhood is extracted, and the dielectric information of the local soil around the candidate point is extracted to provide data support for consistency determination. Calculate the variance and arithmetic mean of the original numerical sequence of dielectric constants, and obtain the deviation ratio between the two. Compare the deviation ratio with a preset threshold. When the deviation exceeds the preset threshold, determine that the current coordinates fall on the boundary of soil abrupt change or the edge region of heterogeneous interlayer, and quantify the degree of heterogeneity of local soil at the candidate point. Coordinate points that are determined to fall on the boundary of soil abrupt change or the edge of heterogeneous interlayer are marked as invalid candidate points and removed. The fruit fly optimization search unit is then triggered to search for replacement coordinates again until the deviation ratio of all retained coordinates is within the preset threshold range. This eliminates the hidden danger of data distortion caused by the measuring points falling on the soil interface and ensures that all final buried locations are located in representative soil areas.

[0010] Preferably, the temperature disturbance timing correction module includes a temperature and humidity mode learning unit and a knowledge graph topology interpolation unit; The temperature and humidity mode learning unit is used to jointly encode the time sequence of moisture content and temperature collected by the smart sensor at each measuring point. It captures the sensitive period and amplitude response law of moisture content fluctuation under the influence of diurnal temperature difference through the gating mechanism of long short-term memory network, separates the false anomaly caused by thermal disturbance from the real moisture change mode, effectively peels off the false fluctuation caused by thermal effect, and separates the real moisture change signal from the thermal disturbance. The knowledge graph topology interpolation unit is used to construct a knowledge graph of water content evolution with the spatial location of the measurement point as nodes and the thermal conductivity coefficient and water transport rate as edge weights. It performs topological logic verification on outliers in the time series, performs physical consistency interpolation on the broken periods based on the energy and mass conservation constraints between neighboring nodes, outputs a corrected continuous water content sequence, repairs the breaks and outliers in the time series data, and reconstructs a continuous water content evolution process that conforms to physical laws.

[0011] Preferably, in the temperature and humidity mode learning unit, the process of separating the pseudo-anomaly caused by thermal disturbance from the real moisture change mode is as follows: The moisture content time series and temperature time series of each measuring point are collected from the smart sensor. The dual-channel signals are synchronously sliced ​​and normalized and encoded using a sliding time window to construct a time series sample set containing joint features of temperature and moisture content. This ensures that the temperature and moisture content signals are on the same scale, thus establishing a standardized input sample space for subsequent coupled learning. The time series sample set is input into the Long Short-Term Memory Network. The forget gate is used to screen the sensitive period of moisture content fluctuation in the diurnal temperature range cycle. The amplitude response features are extracted by the input gate, and the local fluctuation mode parameters that characterize the temperature-humidity coupling relationship are output. The key period of the most intense thermal disturbance is locked, and the instantaneous gain relationship between temperature change and moisture content fluctuation is accurately quantified. Based on the local fluctuation mode parameters, the original moisture content time series is separated by thermal perturbation components. The residual signal after removing the thermal pseudo-anomaly is marked as the real moisture change mode. The false fluctuations caused by the diurnal temperature difference are eliminated, and the time-varying signal that truly reflects the water replenishment event is retained.

[0012] Preferably, in the knowledge graph topology interpolation unit, the process of outputting the corrected continuous moisture content sequence is as follows: Using the spatial location of each measuring point as nodes and the thermal conductivity coefficient and water transport rate as edge weights, a directed weighted knowledge graph of water content evolution is constructed by combining the subgrade soil stratification parameters. This is the water content evolution knowledge graph. Energy and mass conservation constraint equations are established between neighboring nodes. The subgrade soil stratification parameters in each depth interval can be obtained by calling the stratification identification results in the three-dimensional heterogeneity weighted graph. The thermal conductivity coefficient is obtained by back-calculation based on the phase lag of the temperature time series of each measuring point, and the water transport rate is calculated based on the hydraulic conductivity coefficient and hydraulic gradient. A structured knowledge system integrating the physical mechanism of heat-water coupling is constructed to provide a physical constraint benchmark for time series anomaly detection. Outlier detection is performed on the real moisture change modal time series. Jump points exceeding three standard deviations of the sliding window are marked as break periods. The heat conduction coefficient and moisture transport rate edge weights of the corresponding neighboring nodes in the moisture evolution knowledge graph are called to perform topological logic verification on the moisture change trend of the break periods. The physical coupling information of the neighboring nodes is used to verify the direction of abnormal changes in order to identify unreliable jumps that violate the conservation law. Based on the topological logic verification results and calling the subgrade soil layering parameter constraint interpolation boundary, the physical consistency of the fracture period is filled by the weighted interpolation of the neighboring nodes, so that the water content sequence after filling satisfies the energy conservation and mass conservation constraint equations at the same time. The corrected continuous water content sequence is output to ensure that the repaired time series data strictly conforms to the physical law of heat-water coupling and reconstructs the longitudinal continuous and conserved water content evolution process.

[0013] Preferably, in the probability inversion deduction module, the process of inversely calculating the source region probability field and propagation direction vector of the sudden increase signal source region is as follows: The corrected continuous water content sequence is organized into layers according to the depth of the measuring points. A dynamic Bayesian network topology is constructed with the depth layer as the state node and the time slice as the evolution step. The conditional probability transition matrix between each node is initialized to form a depth-time joint probabilistic reasoning framework, which provides topological support for locating the source area of ​​sudden increase in water content. The water content increment is calculated point by point on each time slice. When the water content increment exceeds the preset increment threshold, it is marked as a sudden increase event. The temporal sequence and magnitude gradient of the sudden increase event are recorded as the observation evidence sequence for the back reasoning of the dynamic Bayesian network. The spatiotemporal information of the sudden increase in water content is transformed into observational evidence that can be reasoned, and the temporal sequence and intensity gradient of each event are clarified. The observed evidence sequence is propagated backward along the time slice, and the posterior probability of each grid node as a signal source is inverted through particle filtering sequential Monte Carlo inference. The three-dimensional source region probability field and the propagation direction vector from the source region to the response region are output, and the spatial distribution of the signal source region is located in reverse, and the propagation direction and path of the moisture anomaly are clarified.

[0014] Preferably, in the pattern discrimination and early warning module, the process of generating the corresponding early warning signal and spatial pointing signal is as follows: Extract the three-dimensional spatial distribution shape of the source region probability field and the horizontal and vertical components of the propagation direction vector to form the feature vector of the current abnormal event, which serves as the input parameter for pattern recognition. Quantify the spatial shape of the source region and the propagation direction into structured features to establish a unified input form for attribution judgment. The feature vector of the current abnormal event is compared with the standard topological features of the three early warning modes by Bayes likelihood ratio calculation. The standard topological features of the three early warning modes are preset as follows: the high probability area of ​​the source region corresponding to the upper layer infiltration and recharge is biased towards the upper part and the downward direction of propagation is dominant; the source region corresponding to the lateral undercurrent intrusion is horizontally striped and the horizontal direction of propagation is dominant; the source region corresponding to the internal piping initiation is locally clustered and the propagation direction is radial or has no dominant direction. The three types of early warning modes are transformed into standard topological templates that can be quantitatively calculated to achieve rapid probabilistic matching and classification. The warning mode corresponding to the maximum likelihood value is selected as the attribution label output. At the same time, a warning signal containing the attribution type and a spatial direction signal indicating the source area location and propagation path are generated. The output includes differentiated warnings containing attribution type and spatial direction, providing clear signal support and location guidance for emergency response.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This distributed monitoring system for deep water content of highway subgrade actively detects and obtains a continuous profile of dielectric constant from the shallow to deep layers of the subgrade using ground-penetrating radar. After inversion and spatial interpretation, a three-dimensional heterogeneity weighted map is generated, which then drives an improved fruit fly algorithm to optimize sensor coordinates. This ensures that the placement of each smart sensor is optimally matched with the physical properties of the local soil layer, avoiding the sampling deviation problem caused by the traditional equidistant deployment method falling in non-representative soil areas, and ensuring the representativeness of the monitoring data.

[0016] 2. This distributed monitoring system for deep moisture content of highway subgrade utilizes a long short-term memory network to jointly encode and learn the time series data of moisture content and temperature. Through a gating mechanism, it captures the sensitive periods and amplitude response patterns of moisture content fluctuations under the influence of diurnal temperature differences, effectively separating pseudo-anomaly components caused by thermal disturbances from the original time series. Simultaneously, it combines the heat conduction and moisture transport constraint rules between measuring points stored in the knowledge graph to perform topological logic verification on residual anomalous jump data, ensuring that the corrected time series data truly reflects the continuous evolution process of the moisture field inside the subgrade and eliminates the inverted jump phenomenon caused by thermal response differences.

[0017] 3. This distributed monitoring system for deep moisture content of highway subgrade constructs a knowledge graph of moisture content evolution using the spatial location of measuring points as nodes and the thermal conductivity coefficient and water transport rate as edge weights. It establishes energy and mass conservation constraints between neighboring nodes and performs physical consistency interpolation on the broken time periods caused by abnormal jumps in the time series. This ensures that the filled moisture content sequence simultaneously satisfies the basic laws of thermodynamics and seepage mechanics, thereby restoring the original continuity of the deep longitudinal migration law and avoiding the problem of data unavailability caused by the breakage of spatiotemporal correlation information.

[0018] 4. This distributed monitoring system for deep water content of highway subgrade feeds corrected continuous time-series data into a dynamic Bayesian network to construct a probability transfer chain and time slice transition matrix between multiple depth measuring points. Using the temporal sequence and amplitude gradient of water content surges at each node as observation evidence, it reverse-engineers the spatial distribution and propagation direction vector of the surge signal source area. It performs probability matching between the two types of features, namely source area morphology and propagation direction, and the typical topological features of three early warning modes. This system can distinguish different causal mechanisms at the moment the signal appears, effectively solving the problem of ambiguous early warning attribution. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the working process of a distributed monitoring system for deep moisture content of highway subgrade according to the present invention; Figure 2 This is a schematic diagram of the system architecture of a distributed monitoring system for deep moisture content of highway subgrade according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a distributed monitoring system for deep moisture content of highway subgrade, comprising the following modules: The heterogeneous prior exploration module is used to actively acquire a continuous dielectric constant profile of the target highway subgrade from shallow to deep layers using ground-penetrating radar. After inversion and spatial interpretation, a three-dimensional heterogeneity weighted map of the soil-rock interface is generated, providing a prior physical field for the layout of measurement points and ensuring that sampling points can actively adapt to the spatial variability of the soil. Based on the ground-penetrating radar transmitting multi-band radar waves to the target highway subgrade from shallow to deep layers, full-waveform reflected echo data is collected. After migration superposition and dielectric constant inversion, an initial dielectric constant profile continuously distributed along the longitudinal direction is generated, realizing a continuous characterization of the dielectric properties inside the subgrade and providing a high-resolution data foundation for subsequent soil stratification. Automatic layer picking and soil-rock interface identification are performed on the initial dielectric constant profile, and the profile data is discretized along the horizontal direction at preset intervals. The process involves constructing a three-dimensional regular grid framework covering the entire roadbed space by dividing it into equal-interval layers at vertical depth. Then, sequential Gaussian simulation is used to interpolate and fill the dielectric constant values ​​of each profile sampling point into each grid cell of the three-dimensional regular grid framework, forming a three-dimensional continuous field of dielectric constant covering the entire roadbed space. This ensures that each spatial location has a quantifiable physical attribute value. Based on the filled three-dimensional regular grid framework, the mean, variance, and coefficient of variation of the dielectric constant in each grid cell are calculated. Weighting factors reflecting the degree of soil-rock mixing and local heterogeneity are assigned according to these three factors, and a three-dimensional heterogeneity weight map covering the entire roadbed space is output. This quantifies the spatial distribution of soil heterogeneity throughout the space, providing clear search guidance and evaluation criteria for intelligent measurement point deployment. As a preferred method, the ground-penetrating radar uses three shielded antennas with center frequencies of 100MHz, 400MHz, and 900MHz, respectively. It performs continuous drag-type detection along the longitudinal survey line of the highway subgrade at equal intervals of 0.5m, covering a depth range from 0m to 3m below the top surface of the subgrade. Each frequency antenna synchronously transmits a Ricker wavelet pulse with a duration of 20ns and receives the full-waveform reflected echo signal. The sampling interval is set to 0.1ns, and each recording window is 200ns. Zero-time correction is performed on the collected full-waveform reflected echo data to eliminate... In addition to system delay, a superposition algorithm based on finite-difference reverse-time migration is then applied to realign the reflected energy of each channel to its true spatial location, eliminating diffraction interference. Simultaneously, a full-waveform inversion strategy is employed, using an initial dielectric constant model as the starting guess and iteratively updating the model parameters using a quasi-Newton method. This reduces the L2 norm residual between the synthesized and measured waveforms to within 1%, generating an initial dielectric constant profile continuously distributed longitudinally along the survey line with a vertical sampling interval of 0.05 m. Automatic stratigraphic picking is then performed on the initial dielectric constant profile, and the instantaneous phase gradient is extracted based on complex signal analysis, combined with Ca... The NNY edge detection operator identifies dielectric constant abrupt change zones to calibrate the spatial distribution trajectory of the soil-rock interface and the boundaries of each soil layer. The profile data is discretized horizontally at 0.5m intervals, and horizontally divided according to the roadbed width by lanes at 1.0m intervals. Vertical depth is layered at 0.1m intervals, thus constructing an I×J×K three-dimensional regular grid framework covering the entire roadbed space, where I, J, and K correspond to the total number of grid nodes in the longitudinal, transverse, and vertical directions, respectively. Then, a sequential Gaussian simulation algorithm is used to calculate the dielectric constant values ​​of known sampling points on each profile. For hard data, a variogram model (spherical model, nugget value 0.5, sill value 6.0, range set to 2.5m longitudinally, 1.5m laterally, and 0.3m vertically) is used to describe spatial autocorrelation. Sequential simulation is performed on each grid node, assigning each grid cell a value representing the equivalent dielectric constant at that location. After completing the full-space dielectric constant filling, for each grid cell in the three-dimensional regular grid framework, a 3×3×3 neighborhood window is defined centered on that cell. The dielectric constant values ​​of all grids within the window are statistically analyzed, and their arithmetic mean is calculated. Unbiased estimation of variance and coefficient of variation Further set weighting factors , Take 0.6, A value of 0.4 is chosen to reflect both the local relative fluctuation amplitude and the absolute dispersion. The higher the value, the stronger the soil-rock mixture and the more prominent the heterogeneity in the area. This weight factor is stored grid by grid in the three-dimensional data field. The weight value recorded by each grid node is the comprehensive heterogeneity index of that location. Then, a three-dimensional heterogeneity weight map covering the entire roadbed space is output. The spatial resolution of the grid cells in the map is completely consistent with the three-dimensional grid frame, that is, 0.5m in the longitudinal direction, 1.0m in the transverse direction, and 0.1m in the vertical direction.

[0022] The measurement and optimization deployment module, based on a three-dimensional heterogeneity weighted map, utilizes an improved fruit fly algorithm's odor concentration determination mechanism and search step size strategy to search for the optimal burial coordinates of each smart sensor within the three-dimensional roadbed space that best match the local soil layer's physical properties. This allows the sampling point community to actively adapt to spatial variations in soil, ensuring the initial measurement point deployment has longitudinal representativeness. It also ensures that each sensor is buried in the spatial location that best represents its target soil layer, guaranteeing the representativeness and reliability of data collection from the source. The measurement and optimization deployment module includes a fruit fly optimization search unit and a spatial matching verification unit. The fruit fly optimization search unit is used to encode the soil representative radius and coefficient of variation of each grid in the three-dimensional heterogeneity weight map into odor concentration determination factors. It then uses an improved fruit fly olfactory search step size and visual clustering rules to iteratively solve the sensor coordinate set in the three-dimensional roadbed space to obtain the three-dimensional deployment coordinates of each intelligent sensor. This enables the active tracking and adaptive matching of the measurement point layout to soil spatial variations. The unit reads the three-dimensional heterogeneity weight map, maps the soil representative radius and coefficient of variation of each grid unit into a two-component encoding of the odor concentration determination factor, and initializes the random distribution position of the fruit fly population in the three-dimensional roadbed space. This transforms soil properties into a searchable concentration field, facilitating swarm intelligence search. A unified optimization benchmark is established, and the odor concentration determination factor of the corresponding grid unit is extracted based on the current position of each fruit fly individual. An olfactory search step size that decreases adaptively with the iteration number is adopted to drive the fruit fly population to perform multi-directional local exploration in three-dimensional space to approach high-concentration areas. The breadth of the global search in the early stage and the depth of the local fine search in the later stage are taken into account to ensure that no key areas with high heterogeneity are missed. When multiple fruit fly individuals gather in the same high-concentration neighborhood, the visual aggregation rule is activated and the repulsive potential energy constraint is introduced to iteratively update the spatial distribution of the population until the convergence condition is met. Then, the three-dimensional layout coordinates of each smart sensor are output to prevent the population from gathering in the local suboptimal area too early and to ensure that each measuring point is evenly distributed in the most representative soil area. As a preferred approach, after reading the three-dimensional heterogeneity weighted map, the coefficient of variation Cv and the soil representative radius R (determined by the autocorrelation length of the dielectric constant in the neighborhood of that grid, reflecting the influence range of that point on the surrounding soil) stored in each grid cell in the map are used as two-component basic data. The odor concentration determination factor is encoded in the form of a two-dimensional vector. The first component is the normalized coefficient of variation value, and the second component is the normalized soil representative radius value. The two components are weighted and summed to form the comprehensive odor concentration scalar of that grid cell. The fruit fly population size is dynamically set according to the total roadbed space volume and the preset number of sensors, set to 15 to 20 times the preset number of sensors. During the iteration search, each fruit fly individual follows a uniform random distribution within the three-dimensional spatial boundary of the roadbed. The initial position coordinates (X0, Y0, Z0) are limited to the longitudinal survey line range, the transverse width of the roadbed, and the depth range of 0 to 3m, respectively. During the iterative search, each fruit fly individual is indexed to the corresponding grid cell in the three-dimensional heterogeneity weight map using its current three-dimensional coordinates. The weight factor W and the coefficient of variation Cv of that grid cell are read, and after encoding and mapping, they are used as the instantaneous odor concentration value of the current individual position. The olfactory search step size adaptively decreases with the number of iterations. The initial step size is set to 0.5m longitudinally, 0.5m laterally, and 0.15m vertically. After every 5 iterations, the step size is reduced to 85% of the current value. The percentage decays to 30% of the initial step size and remains constant, balancing the breadth of global exploration in the early stage with the depth of local refinement in the later stage. In addition, each fruit fly individual generates 6 orthogonal directions and 4 diagonal directions as candidate movement positions in 3D space. The odor concentration value of each candidate position is calculated, and the direction with the highest concentration is selected for movement, forming a multi-directional local exploration mechanism, which makes the fruit fly population gradually gather towards highly heterogeneous areas. When the Euclidean distance between the current positions of multiple fruit fly individuals is less than the judgment threshold, i.e., 0.3m vertically, 0.5m horizontally, and 0.08m vertically, the fruit fly individuals are determined to be in the same high-concentration neighborhood, and then the visual clustering rule is activated, selecting the area with the highest odor concentration in the neighborhood. The individual's location serves as a local aggregation center, guiding other individuals in the same neighborhood to move towards this center. Simultaneously, a repulsive potential constraint is introduced, establishing a repulsion zone around the aggregation center, defined as 0.2m longitudinally, 0.3m laterally, and 0.06m vertically. Individuals within this repulsion zone experience a repulsive force away from the center, forcing them to diffuse towards the less explored areas of the neighborhood. The iteration termination condition is set to the rate of change of the optimal odor concentration value of the population being less than 0.5% in 30 consecutive iterations, or reaching the preset maximum number of iterations of 300. After convergence, the locations of the top N fruit fly individuals in terms of fitness are selected, where N equals the preset number of sensors, and their three-dimensional coordinates are output as the optimal deployment coordinates for each intelligent sensor. The spatial matching verification unit is used to verify the consistency of the neighboring soil conditions of the searched three-dimensional layout coordinates. It calculates the deviation of the dielectric constant variance from the average value within the area surrounding the coordinate, eliminating invalid candidate points. Invalid candidate points are coordinate points located at soil abrupt boundary or heterogeneous interlayer edge. This ensures that the physical properties of the final burial location match the preset sensor range, eliminating the risk of data inaccuracy caused by measuring points falling on soil interfaces, improving the confidence of monitoring data. It acquires the three-dimensional layout coordinates of each smart sensor, delineates a spherical neighborhood centered on each coordinate, extracts the original numerical sequence of dielectric constants for all grid cells within the spherical neighborhood, and extracts the dielectric information of the local soil surrounding the candidate points for consistency judgment. The system provides data support, calculates the variance and arithmetic mean of the original numerical sequence of the dielectric constant, and obtains the deviation ratio between the two. The deviation ratio is compared with a preset threshold. When the deviation exceeds the preset threshold, it is determined that the current coordinate falls on the boundary of soil abrupt change or the edge of heterogeneous interlayer. The heterogeneity of the local soil at the candidate point is quantified. The coordinate points determined to fall on the boundary of soil abrupt change or the edge of heterogeneous interlayer are marked as invalid candidate points and removed. The fruit fly optimization search unit is triggered to search for replacement coordinates again until the deviation ratio of all retained coordinates is within the preset threshold range. This eliminates the hidden danger of data distortion caused by the measuring point falling on the soil interface and ensures that all final buried positions are located in representative soil areas. As a preferred approach, after obtaining the three-dimensional deployment coordinates of each smart sensor through iterative solving using the improved fruit fly algorithm, a neighborhood soil consistency check is performed for each candidate coordinate point. Specifically, a spherical neighborhood with a radius of 0.25m is delineated in the three-dimensional heterogeneity weight map, centered on the candidate coordinate point. The radius of the spherical neighborhood is determined based on the vertical resolution of the ground-penetrating radar (0.05m) and the horizontal grid spacing (0.5m vertically and 1.0m horizontally), ensuring that the spherical neighborhood contains at least 45 grid cells, including at least 3 vertical, 3 horizontal, and 5 vertically cells. Then, the dielectric constant values ​​stored in all grid cells covered by the spherical neighborhood are extracted, read out one by one according to spatial location, forming a raw dielectric constant value sequence. After extraction, the raw dielectric constant value sequence is temporarily stored in a buffer. The arithmetic mean of the extracted raw dielectric constant value sequence is calculated. With unbiased estimation of variance Then calculate the deviation ratio. This comprehensively reflects the relative dispersion of the dielectric constant within the neighborhood. A larger deviation ratio indicates more significant differences in soil properties within the neighborhood. With preset threshold The comparison is performed, and the preset threshold is determined based on the statistical distribution of the dielectric constant fluctuation range of the soil-rock mixture layer in the roadbed filling project. That is, when This indicates that the soil quality within the adjacent area is relatively uniform, the measuring points are located in a representative soil area, and the verification is successful; when This indicates that the candidate coordinate point is near a soil-rock abrupt interface or the edge of a heterogeneous interlayer, and the verification fails. Candidate coordinates are directly marked as invalid and eliminated. It should be noted that when the spherical neighborhoods of multiple candidate coordinates overlap, the judgment of each coordinate is independent and does not interfere with each other, using only the dielectric constant sequence within its respective neighborhood as the criterion. For the coordinates of the marked invalid candidate points, their spatial coordinates are recorded and fed back to the fruit fly optimization search unit, triggering a replacement coordinate re-search process. During the re-search, the fruit fly algorithm, based on the current iteration convergence state, regenerates initial candidate individuals within a local spatial range of ±0.5m longitudinally, ±0.5m laterally, and ±0.15m vertically around the eliminated point, inheriting the original odor concentration encoding parameters and step size decay strategy, and performing a local refined search. After each round of replacement search, the newly generated candidate coordinates re-enter the spatial matching verification unit to undergo the same spherical neighborhood extraction and deviation ratio judgment, until the deviation ratios of all retained coordinates meet the requirements. Meanwhile, during the roadbed laying operation, all the verified coordinate points are numbered in the order of the longitudinal survey lines, and the actual burial depth corresponding to each coordinate is marked to the centimeter level accuracy, forming the final layout plan and handing it over to the site for construction.

[0023] The temperature disturbance time series correction module is used to synchronously input the original moisture content time series and temperature data collected by each smart sensor into the evolution correction framework that integrates LSTM and knowledge graph. The former learns the local fluctuation mode under temperature and humidity coupling, while the latter performs topological logic verification and interpolation on abnormal jump data according to the heat conduction and moisture transport constraint rules between measuring points, restores the continuous evolution path of longitudinal time series data, and outputs corrected continuous time series data, eliminating the time series inversion jump caused by diurnal temperature difference and restoring the true evolution law of longitudinal migration of roadbed moisture.

[0024] The probability inversion and deduction module is used to feed the corrected continuous water content sequence into a dynamic Bayesian network to construct a probability transfer chain and time slice transition matrix between multiple depth measurement points. Using the temporal sequence and amplitude gradient of water content surges at each node as observation evidence, it reversely calculates the source region probability field and propagation direction vector of the surge signal source region, and traces the source location and diffusion path of the surge signal in reverse, providing a probabilistic basis for quickly locating the abnormal source region.

[0025] The pattern discrimination and early warning module is used to perform probability matching between the source region probability field and propagation direction vector and the typical topological features of three early warning modes, output the maximum likelihood attribution label, and generate corresponding early warning signals and spatial pointing signals. The early warning modes include upper layer infiltration recharge, lateral undercurrent intrusion and internal piping initiation, so as to realize accurate attribution classification and targeted early warning of water anomaly events and improve the pertinence of emergency response.

[0026] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the temperature disturbance timing correction module includes a temperature and humidity mode learning unit and a knowledge graph topology interpolation unit; The temperature and humidity modal learning unit is used to jointly encode the moisture content and temperature dual-channel time series collected by the intelligent sensor at each measuring point. Through the gating mechanism of a long short-term memory network, it captures the sensitive periods and amplitude response patterns of moisture content fluctuations under the influence of diurnal temperature differences, separating pseudo-anomalies caused by thermal disturbances from the true moisture change modes. This effectively removes false fluctuations caused by thermal effects, separating the true moisture change signal from thermally induced disturbances. It collects the moisture content and temperature time series from the intelligent sensor at each measuring point, and uses a sliding time window to synchronously slice and normalize the dual-channel signals, constructing a time series sample set containing joint temperature-moisture content features, thus ensuring that the temperature and moisture content signals are in a unified state. A standardized input sample space is established for subsequent coupled learning. The time series sample set is input into the Long Short-Term Memory network. The forget gate is used to screen the sensitive periods of water content fluctuation in the diurnal temperature range cycle. The amplitude response features are extracted using the input gate, and the local fluctuation mode parameters that characterize the temperature-humidity coupling relationship are output. The key period of the most intense thermal disturbance is locked, and the instantaneous gain relationship between temperature change and water content fluctuation is accurately quantified. Based on the local fluctuation mode parameters, the thermal disturbance component of the original water content time series is separated. The residual signal after removing thermal pseudo-anomalies is marked as the real water content change mode. The false fluctuations caused by diurnal temperature range are eliminated, and the time-varying signal that truly reflects the water replenishment event is retained. As a preferred method, at the roadbed monitoring site, after each intelligent sensor is buried at a designated depth according to the deployment plan, it continuously collects raw time-series data of volumetric moisture content and temperature in two channels at a sampling interval of 10 minutes. The dual-channel signals are timestamped to ensure that the moisture content and temperature readings at each moment are synchronized. A sliding time window is used to slice the synchronized dual-channel signals. The window width is set to 24 hours, i.e., 144 continuous sampling points, and the sliding step size is set to 1 hour, i.e., 6 sampling points, thus completely covering a full cycle of diurnal temperature variation. The moisture content and temperature sequences within the window are Z-score normalized to ensure that the dual-channel data are on the same dimensional scale. The normalized temperature and moisture content sequences are then concatenated side-by-side along the time axis to form a multidimensional time series sample set containing joint temperature-moisture content features. This multidimensional time series sample set is then input into a pre-trained Long Short-Term Memory (LSTM) network. The network structure employs a two-layer LSTM stacked architecture with 128 hidden layer neurons. The time unfolding step size and sliding window width are both set to 24. The LSTM network's forget gate dynamically assigns weights to the joint temperature-moisture content features at each time step based on the input gating signal. This identifies sensitive time subsequences where the moisture content fluctuation amplitude is significantly amplified during the afternoon warming period and the early morning cooling period. These sensitive time subsequences correspond to… During the period of most intense thermal disturbance at the alternation of day and night temperature peaks and troughs, the input gate extracts amplitude response features. The instantaneous gain relationship between the rate of temperature change and the rate of moisture content change is quantified through the output activation value of the gating unit. The gain coefficient and phase lag at each time step are encapsulated as local fluctuation mode parameters. The network output layer uses a fully connected mapping to output the learned temperature-humidity coupling law as a mode parameter vector. Based on the local fluctuation mode parameters output by the Long Short-Term Memory network, a thermal disturbance component estimation sequence of the same length as the original moisture content time series is constructed. Specifically, the measured temperature value at each time step is multiplied by its corresponding gain coefficient, and a phase lag compensation term is added to calculate the thermal disturbance component estimation sequence at that time step. The estimated value of the pseudo-anomaly in moisture content caused by thermal effects is extracted. The estimated value of this thermal disturbance component is subtracted point by point from the original moisture content time series. The resulting residual signal is the true moisture change mode after removing the pseudo-anomaly. Based on this, when the absolute value of the residual signal amplitude is less than ±0.2% and the duration exceeds 30 minutes, it is determined to be a false fluctuation caused entirely by thermal effects and is smoothed out (using local weighted regression or moving average smoothing). Conversely, if the residual signal shows a monotonically increasing trend and the continuous change amplitude exceeds 0.5% / h, it is determined to be a real moisture replenishment event. After separation, the time series of the true moisture change mode is output to the knowledge graph topology interpolation unit at the original sampling interval (10 minutes). The knowledge graph topology interpolation unit is used to construct a moisture content evolution knowledge graph with spatial locations of measuring points as nodes and thermal conductivity coefficient and water transport rate as edge weights. It performs topological logic verification on outliers in the time series, performs physical consistency interpolation on broken periods based on energy and mass conservation constraints between neighboring nodes, outputs a corrected continuous moisture content sequence, repairs breaks and outliers in the time series data, and reconstructs a continuous moisture content evolution process that conforms to physical laws. Using spatial locations of each measuring point as nodes and thermal conductivity coefficient and water transport rate as edge weights, it constructs a directed weighted knowledge graph of moisture content evolution, i.e., a moisture content evolution knowledge graph, and establishes energy and mass conservation constraint equations between neighboring nodes. The subgrade soil stratification parameters within each depth interval can be obtained by calling the stratification identification results in the three-dimensional heterogeneity weighted graph. The thermal conductivity coefficient is calculated based on the phase lag of the temperature time series at each measuring point, and the water transport rate is calculated based on hydraulic... The conductivity coefficient and hydraulic gradient are calculated to construct a structured knowledge system that integrates the physical mechanism of heat-water coupling, providing a physical constraint benchmark for time series anomaly detection. Outlier detection is performed on the real moisture change modal time series, and jump points exceeding three standard deviations of the sliding window are marked as fracture periods. The heat conduction coefficient and water transport rate edge weights of the corresponding neighboring nodes in the moisture content evolution knowledge graph are called to perform topological logic verification on the moisture content change trend of the fracture period. The physical coupling information of the neighboring nodes is used to verify the direction of abnormal changes to identify unreliable jumps that violate the conservation law. Based on the topological logic verification results and the roadbed soil layering parameters constrain the interpolation boundary, the weighted interpolation of the neighboring nodes is used to fill the physical consistency of the fracture period, so that the filled moisture content sequence simultaneously satisfies the energy conservation and mass conservation constraint equations. The corrected continuous moisture content sequence is output to ensure that the repaired time series data strictly conforms to the physical law of heat-water coupling and reconstructs the longitudinal continuous and conserved moisture content evolution process. As a preferred approach, the knowledge graph of moisture content evolution uses the locations of each smart sensor as graph nodes. The total number of nodes equals the actual number of sensors buried. The coordinates of each node in three-dimensional space are determined by the deployment scheme. The edge connections of the graph are established based on the principle of spatial proximity. That is, when the spatial Euclidean distance between two nodes is less than 1.5 times the diagonal length of the grid cell, a directed connection edge is established between them. The weight of each directed edge consists of two physical quantities: thermal conductivity coefficient and water transport rate. The thermal conductivity coefficient is obtained by extracting 72 consecutive hours of temperature time-series data from each measuring point, using cross-correlation analysis to calculate the phase lag of the temperature response of each node relative to its neighboring nodes, and then, based on the phase dispersion relationship of the one-dimensional heat diffusion equation, calculating the equivalent heat diffusion coefficient of the node in the corresponding depth range, and then multiplying it by the soil moisture content in that depth range. The volumetric heat capacity of the body (obtained by converting the dielectric constant in the three-dimensional heterogeneity weighted graph using an empirical formula) is used to obtain the thermal conductivity coefficient. The water transport rate is inferred in the following way: During natural periods without precipitation and artificial replenishment, the hydraulic gradient between two adjacent measuring points is selected (converted by the difference in burial depth and the matrix potential corresponding to the water content). Combined with the hydraulic conductivity coefficient independently determined by field permeability tests or soil-water characteristic curves and saturated permeability coefficient, the water transport rate in the connection direction is obtained. The soil stratification parameters are directly retrieved from the layer automatic picking results in the three-dimensional heterogeneity weighted graph. The average dielectric constant of each depth interval is read at a vertical interval of 0.1m and classified into three categories: gravel layer, clay layer, or soil-rock transition layer. Each category corresponds to a set of preset physical parameter value ranges. The edge weights and node attributes together constitute the basic data layer of the directed weighted knowledge graph. The expression for thermal conductivity is as follows: ; ; In the formula: The thermal conductivity coefficient characterizes the ability of soil to conduct heat. The equivalent thermal diffusivity is used to characterize the speed at which heat is transferred in the soil. The larger the value, the faster the thermal disturbance propagates. Volumetric heat capacity represents the amount of heat required to raise the temperature of a unit volume of soil by one unit. , The dry density of the soil. Both specific heat capacity and dielectric constant exhibit statistical regression relationships. The angular frequency of temperature fluctuations is determined by the period of diurnal temperature variation. ,in (24-hour cycle); The spatial distance between two adjacent measuring points is the absolute value of the difference in burial depth between the two nodes; The temperature response phase lag is calculated by cross-correlation analysis and reflects the time delay required for heat to propagate from a neighboring node to the current node. The expression for the rate of water transport is as follows: ; ; In the formula: Water transport rate is the velocity at which water moves along a certain direction in the soil. The hydraulic conductivity coefficient; For hydraulic gradient; The total head difference between two adjacent measuring points is determined by the difference in burial depth (positional head) and the matrix potential (obtained by conversion from water content); In the knowledge graph, each node stores not only its own attributes but also the conservation constraints associated with all its incoming and outgoing edges. The energy conservation constraint equation is expressed as follows: the change in heat content per unit volume of soil at the node per unit time is equal to the negative divergence of heat flux density introduced and derived through each connecting edge, where heat flux density is given by the product of the thermal conductivity coefficient and the temperature gradient on that edge. The mass conservation constraint equation is expressed as follows: the change in volumetric water content at the node per unit time is equal to the negative divergence of water flux infiltrating and escaping through each connecting edge, where water flux is given by the product of the water transport rate and the water content gradient on that edge. Specifically, each node must satisfy the discretized form of the above two equations at each time step, with the residual threshold set to an energy term not exceeding ±0.5 J / (m³). 3 •s), the mass item shall not exceed ±0.005m 3 / (m 3•h) When performing outlier detection on the real moisture change modal time series output by the temperature and humidity modal learning unit, the sliding window width is fixed at 24 hours, i.e., 144 consecutive sampling points. Based on the moving average and moving standard deviation of the data within the window, sampling points exceeding the window mean ± 3 times the moving standard deviation are marked as candidate outliers. If candidate outliers appear consecutively and the duration exceeds 30 minutes, the continuous segment is marked as a break in time. The start and end timestamps of the break in time and the moisture content readings of each intermediate point are recorded. After detection, the edge weight data of all neighboring nodes of the node involved in the break in time are extracted. The topology logic verification is performed in the following order: Based on the start and end times of the break in time, the heat conduction coefficient and moisture transport rate edge weights of all neighboring nodes of the node are read from the moisture content evolution knowledge graph. The theoretical values ​​of the total heat conducted from each neighboring node to the node and the total amount of water transported during the break in time are calculated. The theoretical value of the total amount of water must simultaneously satisfy the discrete difference form of the energy conservation and mass conservation equations. Secondly, the theoretically calculated moisture content change trend is compared with the measured value of each sampling point during the break in time. Actual readings are compared point by point. If the direction of deviation is inconsistent with the theoretical trend, i.e., the measured value shows a jump direction that violates the law of conservation of energy or mass, then the measured value is determined to be an unreliable topological anomaly. For sampling points determined to be topological anomalies, a physical consistency interpolation process is initiated: the interpolation boundary conditions consist of the average moisture content of the 30 minutes before the start point of the fault period and the average moisture content of the 30 minutes after the end point. The boundary constraints are the upper and lower limits of the moisture saturation corresponding to the soil stratification parameters of that node. The interpolation algorithm uses neighborhood node weighted interpolation. The weight coefficient of each neighboring node is proportional to the product of its thermal conductivity and water transport rate, and the sum of all neighboring weights is normalized to 1. For each anomalous sampling point within the fault period, the water content readings of the corresponding time of the neighboring nodes are weighted and averaged according to the above weights, and the residuals are checked by substituting them into the discrete form of the energy conservation and mass conservation equations. If the residual exceeds the aforementioned threshold, the weight allocation is iteratively adjusted until the residual converges to within the threshold. After the filling is completed, the interpolated value replaces the original anomalous reading to form a complete corrected continuous water content sequence.

[0027] In the probability inversion and deduction module, the process of inversely calculating the source region probability field and propagation direction vector of the sudden increase signal source region is as follows: the corrected continuous water content sequence is organized into layers according to the depth of the measuring points, a dynamic Bayesian network topology structure with the depth layer as the state node and the time slice as the evolution step size is constructed, and the conditional probability transition matrix between each node is initialized to form a depth-time joint probabilistic inference framework, which provides topological support for locating the source region of the sudden increase in water content signal. The water content increment is calculated point by point on each time slice, and when the water content increment exceeds the preset increment threshold, it is marked as a sudden increase. The event records the temporal sequence and amplitude gradient of the sudden increase in water content, which serves as the observation evidence sequence for the back-inference of the dynamic Bayesian network. This transforms the spatiotemporal information of the sudden increase in water content into inferable observation evidence, clarifies the temporal order and intensity gradient of each event, and propagates the observation evidence sequence backward along the time slice. Through particle filtering and sequential Monte Carlo inference, the posterior probability of each grid node as a signal source is inverted, and the three-dimensional source region probability field and the propagation direction vector from the source region to the response region are output. This reverses the spatial distribution of the signal source region and clarifies the propagation direction and path of the water anomaly. As a preferred approach, the corrected continuous moisture content sequence is organized into layers according to the burial depth of the measuring points, with a depth interval of 0.5m. The average moisture content of all measuring points within each layer is used as the state observation value of that depth layer. A dynamic Bayesian network topology is constructed with depth layers as state nodes and 10-minute time slices as evolution steps. The total number of nodes in the depth direction of the network is determined based on the actual burial depth range, set to 6 to 10 layers, covering the depth range of 0m to 3m below the top surface of the roadbed. Spatial dependency edges within time slices and state transition edges between time slices are established between nodes of adjacent depth layers. Spatial dependency edges within time slices represent the moisture exchange relationship between different depth layers, and state transition edges between time slices represent the state exchange relationship between different depth layers. To characterize the Markov dependence of water cut evolution over time at the same depth, a Gaussian distribution is used to describe the state transition probabilities when initializing the conditional probability transition matrix between nodes. The mean is set by the direction of water flux determined by Darcy's law, and the variance is calibrated based on the fluctuation statistics of historical time series data. At each time slice, the increment of water cut at the current moment relative to the previous moment is calculated for each node, i.e., the water cut reading at the next sampling point is subtracted from the reading at the previous sampling point. When this increment exceeds a preset increment threshold of 0.008, the time-space point is marked as a water cut surge event. At the same time, the time sequence number and amplitude gradient of the surge event are recorded. The amplitude gradient is defined as the peak value of the surge. The difference between the water content and the average water content 30 minutes before the event's onset, divided by the response time, is used to organize all sudden events into an observational evidence sequence in chronological order. Each evidence item in the sequence includes the event's time slice index, depth layer number, and amplitude gradient value. This observational evidence sequence serves as the input for subsequent inverse inference, retaining only valid events within one hour before and after the onset of the sudden event, filtering out isolated single-point noise events to ensure the spatiotemporal continuity of the evidence sequence meets the input requirements of the dynamic Bayesian network inference. The observational evidence sequence is then input into the dynamic Bayesian network sequentially along the time slices, and the particle filtering sequential Monte Carlo inference method is used to invert the signal source region. The total number of particles is set to 1000. The initial particles are uniformly distributed in the roadbed space. Each time slice performs two steps: prediction and update. The prediction step deduces the particle state based on the conditional probability transition matrix. The update step calculates the particle weights based on the observation evidence of the current time slice and performs resampling. The threshold for the effective number of particles in resampling is set to one-third of the total number of particles. After the iteration is completed, the cumulative weight ratio of particles in each spatial grid is calculated. This ratio is the posterior probability that the grid node is the signal source. Then, a three-dimensional source region probability field is generated. The spatial resolution is consistent with the heterogeneity weight map. The propagation direction vector is determined by the line connecting the maximum probability point in the source region probability field to the probability center of each response measurement point. It is output in the form of horizontal azimuth and vertical tilt.

[0028] In the pattern discrimination and early warning module, the process of generating corresponding early warning signals and spatial pointing signals is as follows: The three-dimensional spatial distribution morphology and the horizontal and vertical components of the propagation direction vector of the source region probability field are extracted to form the feature vector of the current abnormal event, which serves as the input parameter for pattern recognition. The spatial morphology and propagation direction of the source region are quantified into structured features to establish a unified input form for attribution discrimination. The feature vector of the current abnormal event is compared with the standard topological features of the three early warning modes using Bayes' likelihood ratio calculation. The standard topological features of the three early warning modes are preset as follows: the high-probability area of ​​the source region corresponding to upper-layer infiltration recharge is biased towards the upper part. Furthermore, the downward propagation direction is dominant, the lateral undercurrent intrusion corresponds to a horizontal band in the source area with a horizontally dominant propagation direction, and the internal piping initiation corresponds to a local cluster in the source area with a radial or no dominant propagation direction. The three types of early warning modes are transformed into standard topological templates that can be quantitatively calculated, enabling rapid probabilistic matching and classification. The early warning mode corresponding to the maximum likelihood value is selected as the attribution label output, and an early warning signal containing the attribution type and a spatial pointing signal indicating the source area location and propagation path are generated simultaneously. The output includes differentiated early warnings containing attribution type and spatial pointing, providing clear signal support and location guidance for emergency response. As a preferred approach, the three-dimensional source region probability field data and propagation direction vector output by the probability inversion and deduction module are received. The source region probability field is stored in a three-dimensional matrix with a spatial resolution of 0.5m longitudinally, 1.0m laterally, and 0.1m vertically. Each grid cell stores the posterior probability value of the location as a sudden signal source. Probability threshold segmentation is performed on the source region probability field, using 60% of the maximum probability value as the truncation threshold to extract the spatial envelope of high-probability connected domains. During the extraction process, the probability centroid coordinates of the high-probability domains in the vertical direction are statistically analyzed, and their normalized burial depth ratio relative to the top surface of the roadbed is calculated. Simultaneously, the probability centroid coordinates of the high-probability domains in the horizontal direction are statistically analyzed. The horizontal and lateral distribution scales are calculated, and the ratio of horizontal to vertical distribution is used as the first component of the morphological feature. Simultaneously, the horizontal azimuth and vertical inclination angles in the propagation direction vector are normalized to the interval [-1, 1], where the vertical inclination angle is positive downwards and negative upwards, and the horizontal azimuth angle is zero degrees relative to the roadbed longitudinal direction. The morphological and directional features are then sequentially concatenated to form a 6-dimensional feature vector for the current abnormal event, consisting of the vertical probability centroid burial depth ratio, horizontal longitudinal distribution scale, horizontal lateral distribution scale, horizontal-vertical distribution ratio, normalized horizontal azimuth angle, and normalized vertical inclination angle. Three early warning modes... Standard topological features are pre-set in the system knowledge base in the form of a multidimensional Gaussian distribution. Each mode corresponds to a set of mean vectors and covariance matrices, obtained statistically from a large number of numerical simulations and field calibration tests. Under the upper layer infiltration recharge early warning mode, the mean vertical probability centroid burial depth ratio is set to 0.25, the mean horizontal distribution to vertical distribution ratio is set to 1.8, and the mean vertical inclination angle is set to 65° downwards. Under the lateral undercurrent intrusion early warning mode, the mean horizontal distribution to vertical distribution ratio is set to 4.5, the mean vertical inclination angle approaches 0°, and the horizontal azimuth angle is concentrated towards the lateral azimuth of the roadbed. Under the internal piping initiation early warning mode, the mean horizontal distribution to vertical distribution ratio is set to 0.25, the mean horizontal distribution to vertical distribution ratio is set to 1.8, and the mean vertical inclination angle is set to 65° downwards. The mean of the Bubi is set to 1.2. There is no obvious dominant direction in the vertical inclination angle, and the probability distribution in each direction is uniform. The current feature vector is substituted into the Gaussian distribution likelihood function of the three Gaussian models to calculate the likelihood value, and then multiplied by the prior probability of each warning mode. The prior probability is set according to the statistics of roadbed historical defects as follows: upper layer infiltration 0.15, lateral undercurrent 0.10, internal piping 0.05, and the rest 0.70 are classified as no anomaly. The posterior probability values ​​of the four classes are compared, and the mode corresponding to the maximum value is taken as the discrimination result output. For cases where the posterior probability of the no anomaly class exceeds 0.85, it is directly judged as no warning event and no subsequent warning action is triggered. The expression for the Gaussian distribution likelihood function is as follows: ; In the formula: For the current feature vector Belonging to the The Gaussian likelihood value of a warning pattern is used to determine the most typical characteristics of the current event. For pattern index, the range of values ​​is... These correspond to three early warning modes: upper-level infiltration recharge, lateral undercurrent intrusion, and internal piping initiation, respectively. The dimension of the feature space; For the first The covariance matrix of the early warning model, which characterizes the fluctuation amplitude within each characteristic component of the model and the correlation between each pair, was obtained by statistical analysis of a large number of numerical simulations and field calibration experiments. The determinant of the covariance matrix (a scalar) reflects the overall dispersion of the pattern's feature distribution; It is the inverse of the covariance matrix, used to map the deviation between the eigenvector and the mean vector to the normalized squared Mahalanobis distance. This is the six-dimensional feature vector of the current abnormal event; For the first The mean vector of various early warning modes; This is the transpose of the deviation vector, used in matrix multiplication operations; When identified as one of the three warning modes, a differentiated warning signal is generated according to the preset warning level rules. The upper layer infiltration recharge mode triggers a blue warning, and the signal content includes the attribution label of "suspected upper layer infiltration recharge" and a suggestion to inspect the upstream drainage facilities. The lateral subsurface intrusion mode triggers a yellow warning, and the signal content includes the attribution label of "suspected lateral subsurface intrusion" and a suggestion to investigate the surface water accumulation and groundwater level changes on both sides of the roadbed. The internal piping initiation mode triggers a red warning, and the signal content includes the attribution label of "suspected internal piping initiation" and a suggestion to immediately organize special detection and drilling verification. The spatial pointing signal is output in the form of a combination of roadbed cross section and longitudinal section diagram. The spatial location of the maximum probability grid of the probability field of the source area is marked with a red dot on the cross section diagram, and the propagation direction from the center of the source area to the center of the response area is marked with an arrowed vector line. The length of the arrow line segment is proportional to the probability propagation intensity. Then, the warning signal and the spatial pointing signal are pushed to the warning release port through a standardized interface, and are simultaneously stored locally in the event log database, recording the complete timestamp and associated measurement point number.

[0029] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A distributed monitoring system for deep moisture content of highway subgrade, characterized in that, Includes the following modules: The heterogeneous prior exploration module is used to actively acquire the continuous dielectric constant profile of the target highway subgrade from shallow to deep layers using ground penetrating radar, and generate a three-dimensional heterogeneity weight map of the soil-rock interface through inversion and spatial interpretation. The optimal deployment module is determined based on a three-dimensional heterogeneity weight map. An improved fruit fly algorithm is used to search for the optimal burial coordinates of each smart sensor in the three-dimensional roadbed space that best match the local soil physical properties. The temperature disturbance time series correction module is used to synchronously input the original moisture content time series and temperature data into the evolution correction framework that integrates LSTM and knowledge graph, perform topological logic verification and interpolation on abnormal jump data, restore the continuous evolution path of the longitudinal time series data, and output the corrected continuous time series data. The probability inversion and deduction module is used to feed the corrected continuous water content sequence into a dynamic Bayesian network, construct the probability transfer chain and time slice transition matrix between multiple depth measurement points, and inversely calculate the source region probability field and propagation direction vector of the sudden signal source region. The pattern discrimination and early warning module is used to perform probability matching between the source region probability field and the propagation direction vector and the typical topological features of three early warning modes, output the maximum likelihood attribution label, and generate corresponding early warning signals and spatial pointing signals. The early warning modes include upper layer infiltration recharge, lateral undercurrent intrusion and internal piping initiation.

2. The distributed monitoring system for deep moisture content of highway subgrade according to claim 1, characterized in that: In the heterogeneous prior exploration layer module, the process of generating a three-dimensional heterogeneity weight map of the soil-rock interface is as follows: Based on the ground penetrating radar transmitting multi-band radar waves to the shallow to deep layers of the target highway subgrade, collecting full waveform reflected echo data, and generating an initial dielectric constant profile that is continuously distributed along the longitudinal direction through offset superposition and dielectric constant inversion; Automatic layer picking and soil-rock interface identification are performed on the initial dielectric constant profile. The profile data is discretized along the horizontal direction at a preset interval and layered at equal intervals in the vertical depth to construct a three-dimensional regular grid framework covering the entire roadbed space. Then, sequential Gaussian simulation is used to interpolate and fill the dielectric constant values ​​of each profile sampling point into each grid cell of the three-dimensional regular grid framework. Based on the filled three-dimensional regular grid framework, the mean, variance and coefficient of variation of the dielectric constant in each grid cell are calculated, and weight factors reflecting the degree of soil-rock mixing and local heterogeneity are assigned according to the three factors, and a three-dimensional heterogeneity weight map covering the entire subgrade space is output.

3. The distributed monitoring system for deep moisture content of highway subgrade according to claim 2, characterized in that: The determination and optimization deployment module includes a fruit fly optimization search unit and a spatial matching verification unit; The fruit fly optimization search unit is used to encode the soil representative radius and coefficient of variation of each grid in the three-dimensional heterogeneity weight map as odor concentration determination factors, and to iteratively solve the sensor coordinate set in the three-dimensional roadbed space by using the improved fruit fly olfactory search step size and visual aggregation rules to obtain the three-dimensional deployment coordinates of each smart sensor. The spatial matching verification unit is used to verify the consistency of the neighboring soil quality of the searched three-dimensional layout coordinates, calculate the deviation of the dielectric constant variance from the average value in the area surrounding the coordinates, and eliminate invalid candidate points. The invalid candidate points are coordinate points that fall on the boundary of soil abrupt change or the edge of heterogeneous interlayer.

4. The distributed monitoring system for deep moisture content of highway subgrade according to claim 3, characterized in that: In the fruit fly optimization search unit, the process of obtaining the three-dimensional layout coordinates of each intelligent sensor is as follows: Read the three-dimensional heterogeneity weight map, map the soil representative radius and coefficient of variation of each grid cell to the two-component encoding of the odor concentration determination factor, and initialize the random distribution position of the fruit fly population in the three-dimensional roadbed space. Based on the current position of each fruit fly individual, the odor concentration determination factor of the corresponding grid cell is extracted. An olfactory search step size that decreases adaptively with the iteration number is used to drive the fruit fly swarm to perform multi-directional local exploration in three-dimensional space to approach the high concentration area. When multiple fruit fly individuals gather in the same high-concentration neighborhood, the visual aggregation rule is activated and a repulsive potential energy constraint is introduced to iteratively update the population spatial distribution until the convergence condition is met, at which point the three-dimensional deployment coordinates of each smart sensor are output.

5. A distributed monitoring system for deep moisture content of highway subgrade according to claim 3, characterized in that: In the spatial matching verification unit, the process of eliminating invalid candidate points is as follows: Obtain the three-dimensional deployment coordinates of each smart sensor, delineate a spherical neighborhood centered on each coordinate, and extract the original numerical sequence of dielectric constants of all grid cells within the spherical neighborhood. Calculate the variance and arithmetic mean of the original numerical sequence of dielectric constants, and obtain the deviation ratio between the two. Compare the deviation ratio with a preset threshold. When the deviation exceeds the preset threshold, determine that the current coordinates fall on the boundary of soil abrupt change or the edge region of heterogeneous interlayer. Coordinate points that are determined to fall within the boundary of soil abrupt change or the edge of heterogeneous interlayer are marked as invalid candidate points and removed. The fruit fly optimization search unit is then triggered to search for replacement coordinates again until the deviation ratio of all retained coordinates is within the preset threshold range.

6. A distributed monitoring system for deep moisture content of highway subgrade according to claim 3, characterized in that: The temperature disturbance timing correction module includes a temperature and humidity modal learning unit and a knowledge graph topology interpolation unit; The temperature and humidity mode learning unit is used to jointly encode the time series of moisture content and temperature collected by the smart sensor at each measuring point. It captures the sensitive time period and amplitude response law of moisture content fluctuation under the influence of day and night temperature difference through the gating mechanism of long short time memory network, and separates the pseudo-anomaly caused by thermal disturbance from the real moisture change mode. The knowledge graph topology interpolation unit is used to construct a knowledge graph of water content evolution with the spatial location of the measurement point as nodes and the thermal conductivity coefficient and water transport rate as edge weights. It performs topological logic verification on outliers in the time series, performs physical consistency interpolation on the broken periods based on the energy and mass conservation constraints between neighboring nodes, and outputs a corrected continuous water content sequence.

7. A distributed monitoring system for deep moisture content of highway subgrade according to claim 6, characterized in that: In the temperature and humidity modal learning unit, the process of separating the pseudo-anomaly caused by thermal disturbance from the real moisture change mode is as follows: The moisture content and temperature time series of each measuring point are collected from the smart sensor. The dual-channel signals are synchronously sliced ​​and normalized by a sliding time window to construct a time series sample set containing joint features of temperature and moisture content. The time series sample set is input into the Long Short-Term Memory network. The forget gate is used to screen the sensitive periods of moisture content fluctuation in the diurnal temperature range cycle. The amplitude response features are extracted using the input gate, and the local fluctuation mode parameters that characterize the temperature-humidity coupling relationship are output. Based on the local fluctuation mode parameters, the original moisture content time series is separated by thermal perturbation components, and the residual signal after removing the thermal pseudo-anomaly is marked as the real moisture change mode.

8. A distributed monitoring system for deep moisture content of highway subgrade according to claim 6, characterized in that: The process of outputting the corrected continuous moisture content sequence in the knowledge graph topology interpolation unit is as follows: Using the spatial location of each measuring point as nodes and the thermal conductivity coefficient and water transport rate as edge weights, a directed weighted knowledge graph of water content evolution is constructed by combining the subgrade soil stratification parameters. This is the water content evolution knowledge graph. Energy and mass conservation constraint equations are also established between neighboring nodes. Outlier detection is performed on the real moisture change modal time series. Jump points that exceed three standard deviations of the sliding window are marked as break periods. The heat conduction coefficient and moisture transport rate edge weights of the corresponding neighboring nodes in the moisture content evolution knowledge graph are called to perform topological logic verification on the moisture content change trend during the break periods. Based on the topological logic verification results and calling the subgrade soil layering parameter constraint interpolation boundary, the physical consistency of the fracture period is filled by the weighted interpolation of the neighboring nodes, so that the water content sequence after filling satisfies both the energy conservation and mass conservation constraint equations, and the corrected continuous water content sequence is output.

9. A distributed monitoring system for deep moisture content of highway subgrade according to claim 6, characterized in that: In the probability inversion deduction module, the process of inversely calculating the source region probability field and propagation direction vector of the sudden increase signal source region is as follows: The corrected continuous moisture content sequence is organized into layers according to the depth of the measuring points. A dynamic Bayesian network topology is constructed with the depth layer as the state node and the time slice as the evolution step size. The conditional probability transition matrix between each node is initialized. The water content increment is calculated point by point on each time slice. When the water content increment exceeds the preset increment threshold, it is marked as a sudden increase event. The temporal sequence and magnitude gradient of the sudden increase event are recorded as the observation evidence sequence for the back reasoning of the dynamic Bayesian network. The observed evidence sequence is propagated backward along the time slice, and the posterior probability of each grid node as a signal source is inverted through particle filtering sequential Monte Carlo inference. The three-dimensional source region probability field and the propagation direction vector from the source region to the response region are output.

10. A distributed monitoring system for deep moisture content of highway subgrade according to claim 9, characterized in that: In the pattern discrimination and early warning module, the process of generating the corresponding early warning signal and spatial pointing signal is as follows: Extract the three-dimensional spatial distribution shape of the source region probability field and the horizontal and vertical components of the propagation direction vector to form the feature vector of the current abnormal event; The feature vector of the current abnormal event is compared with the standard topological features of the three early warning modes by Bayes likelihood ratio calculation. The standard topological features of the three early warning modes are preset as follows: the high probability area of ​​the source area corresponding to the upper layer infiltration and recharge is biased towards the upper part and the downward direction of propagation is dominant; the source area corresponding to the lateral undercurrent intrusion is horizontally striped and the horizontal direction of propagation is dominant; the source area corresponding to the internal piping initiation is locally clustered and the propagation direction is radial or has no dominant direction. The warning mode corresponding to the maximum likelihood value is selected as the attribution label output. Simultaneously, a warning signal containing the attribution type and a spatial pointing signal indicating the source area location and propagation path are generated, and a differentiated warning containing the attribution type and spatial pointing is output.

Images