Method for monitoring and early warning of salt-affected roadbed in seasonal frozen region based on neural network

Abstract

The application discloses a neural network-based monitoring and early warning method for subgrade deformation in a seasonal frozen region with salinization, which comprises static attribute data acquisition, dynamic monitoring data collection, action history data collection, data grid fusion, construction and training of a first neural network model, construction and training of a second neural network model, regular monitoring and intelligent diagnosis, and intelligent prediction and early warning; the vertical deformation field of the subgrade and the spatial continuous data of the salt concentration field, the survey data, the traffic load, and the action history data such as the freeze-thaw cycle are fused, and then the series-connected neural network model is used to predict the disease type and severity of the subgrade under the current and future working conditions, realizing the full-chain intelligent analysis from multi-source data fusion, deformation prediction to disease identification, solving the problems of fragmented monitoring and lagging early warning of the traditional method, and significantly improving the perception ability and early warning level of the health status of the subgrade in the seasonal frozen region with salinization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road engineering health monitoring and disaster prevention and mitigation, and particularly relates to an intelligent monitoring and early warning method for salt-affected roadbed deformation in the seasonal freezing region of northwest China. BACKGROUND

[0002] The northwest region of China has a dry and rainy climate, strong evaporation, and widely distributed soil salinization, with the additional feature of significant seasonal freeze-thaw cycles. In such a complex environment, the roadbed is subjected to repeated traffic loads, freeze-thaw cycles, and salt erosion, which can cause uneven settlement, salt heaving, mud boiling, and other deformation diseases, seriously affecting the safety and service life of the road, and becoming a major problem restricting the stable operation of the regional transportation network.

[0003] Currently, the health monitoring and early warning of salt-affected roadbeds in seasonal freezing regions mainly rely on traditional methods, which still have the following shortcomings: 1) fragmented monitoring data: conventional methods often use manual periodic inspections, discrete settlement markers, or single-point sensors, which can only obtain local and isolated deformation data, and cannot fully reflect the spatial continuous distribution characteristics of the overall vertical deformation field of the roadbed; 2) incomplete monitoring content: existing monitoring work does not fully consider key disaster-causing factors such as salt distribution, load history, and freeze-thaw frequency, and their coupling effects, and cannot establish an internal relationship between multiple factors and deformation responses from a mechanistic perspective; 3) lagging early warning mechanism: existing early warning is based on historical statistics to set fixed thresholds, and does not fully consider the spatial variability of soil properties, as well as the nonlinear and time-varying characteristics of load, freeze-thaw, and salt coupling, resulting in strong subjectivity in judgment, and often only alarms after the appearance of obvious diseases, missing the opportunity for proactive intervention.

[0004] The existing patent CN109653185B discloses a multi-data source roadbed monitoring system in a saline soil area and a monitoring and evaluation method thereof. The system collects roadbed multi-source data from a micro weather station installed in the roadbed land and performs correlation analysis, achieving monitoring and evaluation of saline soil roadbed diseases. However, this monitoring and evaluation method relies on fixed point monitoring for data acquisition, making it difficult to achieve full-area continuous coverage; in the analysis of roadbed multi-source data, the coupling characteristics of load-freeze-salt are not fully considered; the early warning mechanism is based on historical statistics to set fixed thresholds, which does not fully consider the spatial variability of soil properties, and cannot achieve early warning of roadbed diseases.

[0005] In recent years, the vehicle-mounted rapid detection technology provides a new way for efficient acquisition of spatial data such as roadbed deformation field and salt concentration field. However, the existing technology has not broken through the bottleneck of deep fusion and intelligent analysis of multi-source heterogeneous data, and has not established a complete intelligent mapping relationship from disaster-causing factors to deformation response and then to disease types, resulting in that the data value is not fully tapped and it is difficult to realize the leap from regular monitoring to precise early warning.

[0006] Therefore, there is an urgent need for an intelligent method capable of deeply fusing multi-source information, quantifying disaster-causing mechanism, and having dynamic prediction ability, to realize precise monitoring and early warning of salt-affected roadbed deformation in the seasonal frozen region. SUMMARY

[0007] To achieve the above-mentioned purpose, the present application provides a neural network-based monitoring and early warning method for salt-affected roadbed deformation in the seasonal frozen region. The method acquires multi-source dynamic data of the roadbed by vehicle-mounted rapid detection technology, and uses two serial neural network models to realize prediction from disaster-causing factors to deformation response and identification from deformation response to disease types, respectively, to finally form an intelligent analysis chain of disaster-causing factors, deformation prediction, and disease early warning.

[0008] A neural network-based monitoring and early warning method for salt-affected roadbed deformation in the seasonal frozen region, comprising the following steps:

[0009] S1, static attribute data acquisition:

[0010] Extracting rock-soil attribute parameters of the roadbed layer from the road survey and design file, including compression modulus, cohesion, internal friction angle, water content, porosity, and density;

[0011] S2, dynamic monitoring data collection:

[0012] Using a vehicle-mounted laser scanning device and an electromagnetic induction device to synchronously acquire laser point cloud data and conductivity data of the roadbed at a sampling frequency of no less than 10 Hz, and synchronously recording the corresponding time and coordinate position;

[0013] S3, action history data collection:

[0014] According to the time and coordinate position of each monitoring, collecting action history data for quantifying the cumulative influence of traffic load and freeze-thaw environment on the roadbed, including standard load cumulative action times and freeze-thaw cycle cumulative action times;

[0015] S4, grid fusion of data:

[0016] A unified geographic reference system and a regular plane grid are established; a linear interpolation method is used to map the rock-soil attribute parameters and action history data to each plane grid unit; the roadbed surface point cloud data is registered and difference calculated to extract the vertical deformation value of each ground surface point, and a spline interpolation algorithm is used to generate a vertical deformation field grid matrix corresponding to the plane grid unit covering the entire road section; the roadbed conductivity data is registered, the salt concentration is calculated according to the pre-established calibration model of the correlation between conductivity and salt concentration, and a spline interpolation algorithm is used to generate a salt concentration field grid matrix corresponding to the plane grid unit covering the entire road section;

[0017] S5, construction and training of the first neural network model:

[0018] The first neural network model is constructed and trained to establish a nonlinear mapping relationship between the rock-soil attribute parameters, action history data, salt concentration field grid matrix and vertical deformation field grid matrix, and to realize the prediction of the vertical deformation of the roadbed; the model input is the rock-soil attribute parameters, action history data and salt concentration field grid matrix of each grid unit, and the output is the vertical deformation prediction value of the corresponding grid unit; a combination architecture of graph neural network module and deep neural network is used to explicitly model the spatial correlation between grid units;

[0019] S6, construction and training of the second neural network model:

[0020] The second neural network model is constructed and trained to establish a nonlinear mapping relationship between the complex spatial distribution characteristics of the roadbed vertical deformation field and the disease type and its severity, and to realize the prediction of the roadbed disease; the model input is the vertical deformation field grid matrix representing the overall deformation space of the roadbed, and the output is the disease type and severity level; the second neural network model uses a double-path convolutional neural network based on an attention mechanism to extract multi-scale spatial features of the deformation field; wherein the disease type includes uneven settlement, salt heaving, dissolution, and frost boiling, and the severity level is divided into none, mild, moderate, and severe; the labels of the disease type and its severity are generated based on historical inspection records and expert judgment;

[0021] S7, regular monitoring and intelligent diagnosis:

[0022] During the road operation, the laser point cloud data and the electromagnetic induction data of the roadbed surface are synchronously collected by using the vehicle-mounted rapid detection technology with a preset period; through the data preprocessing process in step S4, the real-time data are converted into a current time vertical deformation field grid matrix and a salt concentration field grid matrix consistent with the model input format; the vertical deformation field grid matrix is input into the trained second neural network model, the spatial distribution features of the deformation field are automatically extracted by using the convolutional neural network and the attention mechanism of the model, the disease type is identified and the severity is judged, and the disease diagnosis results of the whole road section are output; this step realizes the normalization and automation of the precise diagnosis of the roadbed health status, and provides direct data support for the maintenance decision.

[0023] S8, intelligent prediction and early warning:

[0024] Based on the aforementioned trained neural network model, a series of prediction and early warning processes are constructed; first, the current monitoring data, the traffic development plan and the medium and long term climate prediction information are comprehensively considered to determine the future working condition parameter set of each spatial grid unit at a specific evaluation time in the future, including the predicted standard load cumulative action times, the freeze-thaw cycle cumulative action times and the salt concentration field grid matrix data; then, the future working condition parameter set is input into the trained first neural network model, and the predicted vertical deformation field grid matrix of the whole road section at the future time is calculated by the nonlinear mapping and spatial correlation analysis ability of the model; then, the vertical deformation field grid matrix is input into the trained second neural network model for spatial pattern recognition and disease mapping, and the disease type, spatial position and estimated severity level that may occur at the future time are output; finally, the system automatically generates spatially targeted early warning information before the disease actually appears, actively prompts the maintenance and management department to take proactive intervention measures, thereby realizing the fundamental change from post-treatment to pre-judgment, and significantly improving the timeliness and accuracy of roadbed health management.

[0025] Preferably, in step S3, the standard load cumulative action times are calculated by the vehicle type, axle load and traffic quantity data of the traffic flow monitoring station provided by the road management department, the axle load of different vehicle types is uniformly converted into equivalent standard axle load action times, and then the standard load cumulative action times of each road section position are obtained by adding up according to the monitoring period; the freeze-thaw cycle cumulative action times are determined according to the hourly temperature data of the road weather station, and when the temperature of the monitoring point appears from below 0℃ to above 0℃ and then drops below 0℃ within 24 hours, it is counted as one effective freeze-thaw cycle, so that the freeze-thaw cycle cumulative action times of each road section position are obtained.

[0026] Preferably, in step S4, the size of the plane grid is not less than 0.2x0.2m and not more than 1.0x1.0m, which is set according to the monitoring accuracy requirement.

[0027] Preferably, in step S4, the calibration model of the relationship between the electrical conductivity and the salt concentration is calibrated by drilling not less than 30 representative soil samples along the line and in the salinization variation area of the field roadbed, and then determining the true salt concentration of the soil samples by conductivity titration in the laboratory S , and the electrical conductivity measured by the vehicle-mounted electromagnetic induction device at the same position T Regression analysis is performed; the regression analysis preferably uses a linear regression model, as shown in the following formula:

[0028] S = k 1 ·T + k 2

[0029] wherein, k 1 and k 2 are calibration coefficients; if the determination coefficient of the linear regression model is lower than a preset threshold value, a higher-order polynomial or a machine learning method is used for fitting to improve the adaptability of the calibration model; wherein the determination coefficient R 2 is calculated by the formula R 2 for evaluating the goodness of fit of the linear regression model, R 2 =1- S res / S tot ; the preset threshold value is determined according to historical data experiments to ensure the adaptability of the model. S res is the sum of squares of residuals, S tot is the total sum of squares; the preset threshold value is determined according to historical data experiments to ensure the adaptability of the model.

[0030] Preferably, in step S5, the input of the first neural network model includes the geotechnical attribute parameters, action history data, and salt concentration field grid matrix of each grid cell, and the output is the vertical deformation value of each grid cell; the first neural network model uses a combined architecture of a graph neural network module and a deep neural network, including an input layer, a graph neural network module, a fully connected network hidden layer, and an output layer connected in sequence; the input features of the input layer are 9-dimensional, including the compression modulus, cohesion, internal friction angle, water content, porosity, density, standard load cumulative action times, freeze-thaw cycle cumulative action times, and salt concentration field grid matrix; the grid node X , YThe coordinates are used as structure information of the graph neural network module to define the spatial adjacency relationship between nodes; the graph neural network module adopts a graph sampling aggregation network architecture, and the number of layers is 2; the full connection network has 3 hidden layers, and all use ReLU activation functions; the output of the output layer is the vertical deformation value of each grid node; during the running of the first neural network model, the spatial grid data is first modeled by the graph neural network module, wherein each grid unit is defined as a graph node, and the connection with the adjacent 8 grid units is defined as an edge; the neighborhood information is aggregated by two layers of graph convolution network to explicitly capture the deformation spatial dependence caused by freeze-thaw cycle and salt migration; then, the feature vector output by the graph neural network module, which has contained spatial context information, is input into a full connection network including three hidden layers for nonlinear mapping, and finally the deformation prediction value of each grid unit is output; the first neural network model is supervised trained using the data obtained in steps S1 to S4, and the training target is to minimize the mean square error loss between the model predicted deformation and the measured deformation; the Adam optimizer is used for parameter optimization, and the validation set is used to monitor the training process to prevent overfitting; the training set, the validation set and the test set are randomly divided in the ratio of 7:2:1, and the total number of training samples is not less than 10,000, wherein each training sample corresponds to all input features of a grid unit at a single monitoring time point and the corresponding measured vertical deformation value, and the samples should cover different spatial positions and different monitoring time points.

[0031] Preferably, the second neural network model adopts a double-path convolutional neural network based on an attention mechanism, the convolutional neural network of the first path uses a small convolution kernel of 3x3 to focus on extracting local subtle deformation features, including but not limited to deformation gradients corresponding to micro cracks; the convolutional neural network of the second path uses a large convolution kernel of 7x7 to focus on capturing macro deformation patterns, including but not limited to the morphology of large-scale subsidence basins; then, a spatial attention module is used to fuse and weight the features of the two paths, so that the model can focus on deformation anomalies in disease-prone areas such as shoulders and crack zones, and finally output the disease type and severity through a full connection layer;

[0032] The second neural network model adopts a combined loss function combining disease type identification and severity classification during the training process L as shown in the following formula:

[0033] L = c 1 ·L t + c 2 ·L s

[0034] wherein the disease classification loss L t and the severity loss L s are both cross-entropy loss functions; c 1 and c 2 are both weight coefficients for balancing the importance of the two tasks of disease classification and severity assessment, and their specific values are determined by hyperparameter optimization using a validation set during model training, so that the combined loss function L performs optimally on the validation set.

[0035] In summary, compared with the prior art, the present application has the following beneficial effects:

[0036] 1) Intelligent, full-chain evaluation and early warning of the health status of salt-affected roadbeds in the seasonal frozen region are achieved: by constructing a series neural network analysis framework of disaster-causing factors, continuous deformation field, and spatial disease mode, the drawbacks of the prior art that the monitoring, evaluation, and early warning links are disconnected are overcome; by the first neural network model, deformation prediction is achieved, and by the second neural network model, disease diagnosis is achieved, forming a complete closed loop from monitoring data access to intelligent early warning output, significantly improving the systematicness and automation level of roadbed health management;

[0037] 2) The accuracy and spatial continuity of roadbed deformation prediction are significantly improved: the first neural network model introduces a graph neural network module to explicitly model the spatial physical correlation between roadbed units, such as load transfer, water and salt migration; this structure can effectively learn the complex mechanism of the spatial coupling of salinization, freeze-thaw cycles, and traffic load, thereby generating a physically more reasonable and spatially continuous and smooth predicted deformation field, overcoming the defect of traditional point prediction models that cannot reflect the overall deformation distribution pattern;

[0038] 3) Precise and automatic identification from deformation data to disease type is achieved: the second neural network model uses the powerful spatial feature extraction capability of convolutional neural networks to automatically learn the unique spatial distribution patterns of different diseases such as uneven settlement, salt heaving, and frost boiling corresponding to the vertical deformation field; this method reduces the dependence on human experience, achieving the objectivity and precision of disease identification, and even early-stage disease patterns that have not yet fully emerged can be effectively identified;

[0039] 4) Roadbed disease prediction is achieved: by combining real-time monitoring data with future traffic and climate prediction information, the trained neural network model chain is used to prospectively deduce roadbed deformation and disease development, so that the early warning signal can be issued before visible damage occurs, providing a key decision window for proactive prevention and control;

[0040] 5) The fusion utilization value and decision support capability of massive multi-source heterogeneous data are improved: through a unified space-time gridding framework, multi-source heterogeneous information such as survey design data, high-frequency vehicle-mounted monitoring data, traffic load history and environmental freeze-thaw history is deeply fused; the powerful nonlinear mapping capability of the neural network model can deeply mine the internal correlation between these data and the roadbed response, and convert fragmented information into deep knowledge that can be used for spatial precise early warning, thereby providing direct and intuitive data support for scientific maintenance decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0041] Fig. 1 A flowchart of the neural network-based seasonal frozen region salinization roadbed deformation monitoring and early warning method shown in the embodiments of the present application;

[0042] Fig. 2 A data flow diagram of periodic monitoring and intelligent diagnosis shown in the embodiments of the present application;

[0043] Fig. 3 A data flow diagram of intelligent prediction and early warning shown in the embodiments of the present application. DETAILED DESCRIPTION

[0044] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given here are only for the purpose of illustrating and explaining the present application, and cannot be used to limit the present application.

[0045] The first aspect of the present application discloses a neural network-based seasonal frozen region salinization roadbed deformation monitoring and early warning method as shown in Figs. 1-3 The neural network-based seasonal frozen region salinization roadbed deformation monitoring and early warning method comprises the following steps:

[0046] S1, static attribute data acquisition: extracting the geotechnical attribute parameters of the roadbed layer from the road survey design file, including compression modulus, cohesion, internal friction angle, water content, porosity and density.

[0047] S2, dynamic monitoring data collection: using a vehicle-mounted laser scanning device and an electromagnetic induction device to synchronously acquire laser point cloud data and conductivity data of the roadbed at a sampling frequency of not less than 10Hz, and synchronously recording the corresponding time and coordinate position; in a specific implementation, the vehicle-mounted laser scanning device uses a phase laser radar, the single-point ranging accuracy is not less than ±1mm, and the scanning angle resolution is not less than 0.05°; the electromagnetic induction device uses a ground conductivity meter based on the electromagnetic induction principle, and the conductivity measurement accuracy is not less than ±0.5 mS / m; the vehicle-mounted laser scanning device and the electromagnetic induction device are calibrated on site before use to ensure data accuracy.

[0048] S3, action history data collection: according to the time and coordinate position of each monitoring, the action history data for quantifying the cumulative influence of traffic load and freeze-thaw environment on the roadbed is collected, including the standard load cumulative action times and the freeze-thaw cycle cumulative action times.

[0049] In specific implementation, the standard load cumulative action times are converted by the vehicle type, axle load and traffic quantity data of the traffic flow monitoring station provided by the road management department, that is, the axle load action of different vehicle types is uniformly converted into equivalent standard axle load action times by using the equivalent axle load conversion formula specified in the "Highway Asphalt Pavement Design Specification" (JTG D50-2017), and then accumulated according to the monitoring period, so as to obtain the standard load cumulative action times of each road section position; the freeze-thaw cycle cumulative action times are determined according to the hourly temperature data of the weather station along the road, when the temperature of the monitoring point appears the process of rising from below 0℃ to above 0℃ and then falling below 0℃ within 24 hours, it is counted as an effective freeze-thaw cycle, so as to obtain the freeze-thaw cycle cumulative action times of each road section position.

[0050] S4, grid fusion of data: a unified geographic reference system and a regular plane grid are established; in specific implementation, the size of the plane grid is not less than 0.2x0.2m and not more than 1.0x1.0m, which is set according to the monitoring accuracy requirement; the geotechnical attribute parameters and the action history data are mapped to each plane grid unit by using linear interpolation method; the roadbed surface point cloud data is registered and difference calculated, the vertical deformation value of each ground surface point is extracted, and the spline interpolation algorithm is used to generate the vertical deformation field grid matrix corresponding to the plane grid unit covering the whole road section; the conductivity data of the roadbed is registered, the salt concentration is calculated according to the pre-established calibration model of the conductivity and salt concentration correlation, and the spline interpolation algorithm is used to generate the salt concentration field grid matrix corresponding to the plane grid unit covering the whole road section.

[0051] In specific implementation, the calibration model of the conductivity and salt concentration correlation is obtained by drilling not less than 30 representative soil samples along the field roadbed and the salinization variation area, ensuring that the samples cover the expected salt concentration, water content and soil quality variation range, then the real salt concentration of the soil samples is measured by using the conductivity titration method in the laboratory S , and the conductivity T measured at the same position by the vehicle-mounted electromagnetic induction equipment is subjected to regression analysis; the regression analysis preferably uses a linear regression model, as shown in the following formula:

[0052] S = k 1 ·T + k 2 (1)

[0053] wherein, k 1 and k 2 are the determination coefficients; if the determination coefficients of the one-dimensional linear regression model are lower than the preset threshold value 0.85, a higher-order polynomial or a machine learning method is used for fitting to improve the adaptability of the calibration model; wherein the determination coefficients R 2 are calculated for evaluating the goodness of fit of the one-dimensional linear regression model, R 2 are calculated by the formula R 2 =1- S res / S tot ; the residual sum of squares is S res ; the total sum of squares is S tot ; and the preset threshold value is set to 0.85, which is determined based on historical data experiments to ensure the adaptability of the model.

[0054] S5, construction and training of the first neural network model: a first neural network model is constructed and trained to establish a nonlinear mapping relationship between the geotechnical attribute parameters, action history data, salt concentration field grid matrix and vertical deformation field grid matrix, and to realize prediction of the vertical deformation of the roadbed;

[0055] In specific implementation, the input of the first neural network model includes the geotechnical attribute parameters, action history data and salt concentration field grid matrix of each grid cell, and the output is the vertical deformation value of each grid cell; to solve the specific technical problem of the spatial conductivity and correlation of the salinization roadbed deformation, the first neural network model adopts a combined architecture of a graph neural network module and a deep neural network, including an input layer, a graph neural network module, a fully connected network hidden layer and an output layer connected in sequence; the input features of the input layer are 9-dimensional, including the compression modulus, cohesion, internal friction angle, water content, porosity, density, standard load cumulative action frequency, freeze-thaw cycle cumulative action frequency and salt concentration field grid matrix; the coordinates of the grid nodes are used as the structural information of the graph neural network module to define the spatial adjacency relationship between the nodes; the graph neural network module adopts a graph sampling aggregation network architecture, the aggregation function is mean aggregation, the number of hidden units is 64, and the number of layers is 2; the fully connected network hidden layer has 3 layers, and the number of neurons is 128, 64 and 32 respectively, all using ReLU activation function; the output of the output layer is the vertical deformation value of each grid node. X 、 Y

[0056] ​In a specific implementation, when the first neural network model is running, first, the spatial grid data is modeled by a graph neural network module, wherein each grid cell is defined as a graph node, and the connection with the adjacent 8 grid cells is defined as an edge, and the neighborhood information is aggregated by a two-layer graph convolution network to explicitly capture the spatial dependence of deformation caused by freeze-thaw cycles and salt migration; then, the feature vector containing spatial context information output by the graph neural network module is input into a fully connected network containing three hidden layers for nonlinear mapping, and finally the deformation prediction value of each grid cell is output; the combination of graph neural network and deep neural network is specially used to learn the spatial deformation law of subgrade under the coupling of load, freeze-thaw and salt.

[0057] In a specific implementation, the first neural network model is supervised trained using the data obtained in steps S1 to S4, and the training target is to minimize the mean square error loss between the predicted deformation and the measured deformation; the Adam optimizer is used for parameter optimization, the initial learning rate is 0.001, and the validation set is used to monitor the training process to prevent overfitting; the training set, the validation set and the test set are randomly divided in the ratio of 7:2:1, and the total number of training samples is not less than 10,000, wherein each training sample corresponds to all input features of a grid cell at a single monitoring time point and the corresponding measured vertical deformation value, and the samples should cover different spatial positions and different monitoring time points; the training round is 200 rounds; after training is completed, the model can accurately predict the vertical deformation of each position in the whole road section according to the given future working conditions, such as predicted traffic load and freeze-thaw cycle, combined with relatively stable subgrade properties and salt distribution, to generate the predicted vertical deformation field at the future time.

[0058] S6, construction and training of the second neural network model: the second neural network model is constructed and trained to establish the nonlinear mapping relationship between the complex spatial distribution characteristics of the vertical deformation field of the subgrade and the disease type and its severity, and to realize the prediction of the subgrade disease; wherein the disease type includes but is not limited to uneven settlement, salt heaving, dissolution, frost boiling, and the severity level is divided into no, mild, moderate and severe; the labels of disease type and its severity are generated based on historical inspection records and expert judgment.

[0059] In specific implementation, the input of the second neural network model is a roadbed vertical deformation field grid matrix, and the output is the probability of four types of diseases and the probability of the severity level thereof; in order to realize accurate mapping from continuous deformation field to discrete disease types, the second neural network model adopts a double-path convolutional neural network based on an attention mechanism, the convolutional neural network of the first path uses a small convolution kernel of 3*3 to focus on extracting local subtle deformation features, including but not limited to deformation gradients corresponding to micro cracks; the convolutional neural network of the second path uses a large convolution kernel of 7*7 to focus on capturing macro deformation patterns, including but not limited to the shape of large-scale settlement basins; then, a spatial attention module is used to fuse and weight the features of the two paths, so that the model can focus on deformation abnormalities in disease-prone areas such as road shoulders and crack zones, and finally output the disease type and severity through a fully connected layer; this design is aimed at the multi-scale characteristics of roadbed diseases in the deformation field, and is used to improve the recognition accuracy of complex diseases.

[0060] In specific implementation, the second neural network model adopts a combined loss function combining disease type recognition and severity classification during the training process L , and the minimization of cross-entropy loss is taken as the optimization objective; after the model training is completed, end-to-end mapping from vertical deformation field data to intelligent diagnosis of disease state can be realized, and accurate and reliable disease recognition and evaluation conclusions can be output regardless of whether the input deformation field is derived from measured data or the prediction result of the first neural network model; the combined loss function L is the weighted sum of disease classification loss L t and severity loss L s , as shown in the following formula:

[0061] L = c 1 ·L t + c 2 ·L s (2)

[0062] wherein, the disease classification loss L t and the severity loss L s are both cross-entropy loss functions; c 1 and c 2 are weight coefficients, used to balance the importance of disease classification and severity evaluation, and their specific values are determined by using a validation set to optimize hyperparameters during the model training process, so that the combined loss function LThe performance on the validation set is optimal.

[0063] S7, regular monitoring and intelligent diagnosis: during road operation, laser point cloud data and electromagnetic induction data of the roadbed surface are synchronously collected by using the vehicle-mounted rapid detection technology according to a preset period; through the data preprocessing process described in step S4, the real-time data are converted into a current time vertical deformation field grid matrix and a salt concentration field grid matrix consistent with the model input format; the vertical deformation field grid matrix is input into the trained second neural network model, the spatial distribution features of the deformation field are automatically extracted by using the convolutional neural network and the attention mechanism of the model, the disease type is identified and the severity is judged, and the disease diagnosis result of the whole road section is output; this step realizes the normalization and automation of the precise diagnosis of the roadbed health status, and provides direct data support for maintenance decision-making;

[0064] S8, intelligent prediction and early warning:

[0065] Based on the aforementioned trained neural network model, a series of prediction and early warning processes are constructed; first, the future working condition parameter set of each spatial grid unit at a specific evaluation time in the future is determined by comprehensively considering the current monitoring data, traffic development planning and medium and long term climate prediction information, including the predicted standard load cumulative action times, the freeze-thaw cycle cumulative action times, and the salt concentration field grid matrix data updated by trend analysis; then, the future working condition parameter set is input into the trained first neural network model, and the predicted vertical deformation field grid matrix of the whole road section at the future time is calculated and generated by using the nonlinear mapping and spatial correlation analysis capability of the model; then, the vertical deformation field grid matrix is input into the trained second neural network model for spatial pattern recognition and disease mapping, and the disease type, spatial position and estimated severity grade that may occur at the future time are output; finally, the system automatically generates spatially targeted early warning information before the disease actually appears, actively prompts the maintenance and management department to take proactive intervention measures, thereby realizing the fundamental change from post-treatment to pre-judgment, and significantly improving the timeliness and accuracy of roadbed health management.

[0066] The above is one or more embodiments of the present application, which is described in more detail and in more detail, but cannot be understood as limiting the scope of the present patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present patent should be subject to the appended claims.

Claims

1. A method for monitoring and early warning of deformation of saline roadbed in seasonally frozen areas based on neural networks, characterized in that, Includes the following steps: S1. Static attribute data acquisition: Extract the soil and rock property parameters of the roadbed from the road survey and design documents. The soil and rock property parameters include compression modulus, cohesion, internal friction angle, water content, porosity and density. S2. Dynamic monitoring data acquisition: Using vehicle-mounted laser scanning equipment and electromagnetic induction equipment, laser point cloud data and conductivity data of the roadbed are collected simultaneously, and the collection time and spatial location information are recorded simultaneously. S3. Action history data acquisition: Based on the monitoring time and spatial location information, action history data is acquired, including the cumulative number of standard loads and the cumulative number of freeze-thaw cycles. The cumulative number of standard load actions is calculated by converting the vehicle type, axle load, and traffic volume data from the traffic flow monitoring station provided by the road management department. The axle load of different vehicle types is uniformly converted into the equivalent number of standard axle load actions, and then accumulated according to the monitoring cycle to obtain the cumulative number of standard load actions at each road segment location. The cumulative number of freeze-thaw cycles is determined based on hourly temperature data from meteorological stations along the road. When the temperature at a monitoring point rises from below 0°C to above 0°C and then falls back below 0°C within 24 hours, it is counted as one effective freeze-thaw cycle. Thus, the cumulative number of freeze-thaw cycles at each road section location is statistically obtained. S4. Data gridding and fusion: Establishing a unified geographic reference system and a regular planar grid; The soil and rock property parameters and action history data are interpolated and mapped to grid cells; The laser point cloud data is registered and differentially calculated to extract vertical deformation values ​​and interpolate to generate a vertical deformation field raster matrix covering the entire road section. The conductivity data is registered, converted to salt concentration based on a pre-established calibration model, and interpolated to generate a salt concentration field raster matrix covering the entire road section. The calibration model is established as follows: at least 30 representative soil samples are drilled along the roadbed and in salinization variation areas, and their laboratory salt concentration and conductivity measured by on-board equipment are determined. Regression analysis is performed to obtain the mapping relationship between salt concentration and conductivity. At least 30 representative soil samples are systematically drilled along the roadbed and in salinization variation areas, and then the true salt concentration of the soil samples is determined in the laboratory using conductivity titration. S and conductivity measured at the same location as the vehicle-mounted electromagnetic induction device. T Perform regression analysis; the preferred regression analysis method is a univariate linear regression model, as shown in the following equation: S = k 1 ·T + k 2 (1) Among them, k 1 and k 2 All are calibration coefficients; if the coefficient of determination of the univariate linear regression model R 2 If the value is below a preset threshold, a higher-order polynomial or machine learning method is used for fitting to improve the adaptability of the calibration model; among which, the determination coefficient... R 2 Through formula R 2 =1- S res / S tot Calculations are performed to evaluate the goodness of fit of a univariate linear regression model. S res For the sum of squared residuals, S tot The sum of squares is the total; the preset threshold is determined experimentally based on historical data. S5. Construction and training of the first neural network model: Construct and train the first neural network model. Its input is the soil and rock property parameters, action history data and salt concentration field grid matrix of each grid cell. The output is the vertical deformation prediction value of the corresponding grid cell. A combination architecture of graph neural network module and deep neural network is adopted. S6. Construction and training of the second neural network model: Construct and train the second neural network model. Its input is the vertical deformation field grid matrix representing the overall deformation space of the roadbed, and its output is the type and severity level of the disease. A dual-path convolutional neural network based on the attention mechanism is adopted. S7. Regular monitoring and intelligent diagnosis: During the operation period, the acquired vertical deformation field grid matrix is ​​input into the trained second neural network model to obtain the type and severity of the disease. S8. Intelligent Prediction and Early Warning: Based on the cumulative number of standard loads, cumulative number of freeze-thaw cycles, and salt concentration field grid matrix data for future predictions, a set of parameters for future predicted working conditions is formed and input into the first trained neural network model to obtain the predicted vertical deformation field grid matrix. This matrix is ​​then input into the second trained neural network model to obtain the types and severity of defects under the future predicted working conditions, and early warning information is generated accordingly.

2. The method for monitoring and early warning of deformation of saline roadbed in seasonally frozen areas based on neural networks according to claim 1, characterized in that, In step S4, the unit size of the planar grid is not less than 0.2×0.2m and not more than 1.0×1.0m, and is set according to the monitoring accuracy requirements.

3. The method for monitoring and early warning of deformation of saline roadbed in seasonally frozen areas based on neural networks according to claim 1, characterized in that, In step S5, the input of the first neural network model includes the soil and rock property parameters, action history data, and salt concentration field grid matrix of each grid cell, and the output is the vertical deformation value of each grid cell. The first neural network model adopts a combined architecture of graph neural network module and deep neural network, including an input layer, graph neural network module, fully connected network hidden layer, and output layer connected in sequence. The input features of the input layer are 9-dimensional, including compressive modulus, cohesion, internal friction angle, water content, porosity, density, cumulative number of standard loads, cumulative number of freeze-thaw cycles, and salt concentration field grid matrix. The X and Y coordinates of the grid nodes serve as the structural information of the graph neural network module, used to define the spatial adjacency relationship between nodes. The network module adopts a graph sampling aggregation network architecture with 2 layers; the fully connected network has 3 hidden layers, all using the ReLU activation function; the output layer outputs the vertical deformation value of each grid node; the first neural network model is trained under supervision using the data obtained in steps S1 to S4, and the training objective is to minimize the mean squared error loss between the model's predicted deformation and the measured deformation; the Adam optimizer is used for parameter optimization, and the validation set is used to monitor the training process to prevent overfitting; the training set, validation set, and test set are randomly divided in a ratio of 7:2:1, and the total number of training samples is not less than 10,000, where each training sample corresponds to all input features of a grid cell at a single monitoring time point and the corresponding measured vertical deformation value.

4. The method for monitoring and early warning of deformation of saline roadbed in seasonally frozen areas based on neural networks according to claim 1, characterized in that, In step S6, the second neural network model adopts a dual-path convolutional neural network based on an attention mechanism. The first path convolutional neural network uses a 3×3 small convolutional kernel to focus on extracting local subtle deformation features. The second path's convolutional neural network uses a large 7×7 convolutional kernel to focus on capturing macroscopic deformation patterns; subsequently, a spatial attention module fuses and weights the features from the two paths, and finally outputs the disease type and severity through a fully connected layer; during the training of the second neural network model, a combined loss function combining disease type recognition and severity classification is used. L As shown in the following formula: L = c 1 ·L t + c 2 ·L s (2) Among them, the loss due to disease classification L t and severity of loss L s Both are cross-entropy loss functions; c 1 and c 2 All are weighting coefficients.

Images