A method for predicting subgrade settlement based on a long short-term memory network

Abstract

The application discloses a roadbed settlement prediction method based on a long short-term memory network and relates to the technical field of prediction systems. The method comprises collecting environment data difference values of a target roadbed in a time sequence, wherein the environment data difference values are used for constructing an input feature matrix of a long short-term memory neural network, and the environment data difference values are humidity difference values and temperature difference values. The application effectively captures the dynamic response characteristics of the roadbed under different times and depths by collecting the environment data difference values of the target roadbed in the time sequence and taking the environment data difference values as the input feature matrix of the long short-term memory network. Meanwhile, a correlation function model of the environment data and the axial distance of the roadbed is established, and the model is corrected in combination with the material type, the void ratio, the permeability coefficient, the elastic coefficient and the Poisson coefficient of the roadbed, so that the model not only has a data driving capability, but also integrates inherent physical characteristics and long-term working state constraints of the roadbed, thereby guaranteeing the precision and reliability of the prediction model.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of prediction systems, and in particular to a subgrade settlement prediction method based on a long short-term memory network. BACKGROUND

[0002] Subgrade settlement is a common geotechnical disease that can cause problems such as pavement cracking, subgrade instability and bridge pier tilting, and seriously affects the service life and driving safety of traffic engineering such as roads and railways, so a precise prediction method is urgently needed to guide engineering design and maintenance.

[0003] According to the search, the Chinese invention patent application with the publication number "CN120561455A" proposes a "building deviation correction settlement prediction method and system based on multi-source data-artificial intelligence fusion", which acquires multi-type sensor data related to building settlement through multi-source data acquisition, the multi-type sensors include a level gauge, a total station, a groundwater level sensor and a soil pressure sensor, forming multi-modal data input; and discloses a building deviation correction settlement prediction system based on multi-source data-artificial intelligence fusion, which comprises a data acquisition module, a data processing module, a deep learning prediction module, a model optimization module, an intelligent early warning module and a user interaction module. The application realizes accurate early warning and real-time decision support of settlement risk by dynamically adjusting the settlement risk threshold, introducing multi-level early warning signals and an intelligent decision support system, and significantly improves the adaptability, accuracy and response efficiency of the system.

[0004] In addition, the Chinese invention patent application with the publication number "CN120611569A" proposes a "prediction method for ground settlement caused by ground stress change in geothermal exploitation", which acquires geological survey data of a target geothermal exploitation area to construct a three-dimensional geomechanical model; based on the constructed three-dimensional geomechanical model, a numerical simulation method is used to simulate the ground stress evolution process at different exploitation stages to obtain ground stress tensor time series data; the ground stress tensor time series data are input into the three-dimensional geomechanical model as boundary conditions to calculate the deformation amount of each rock layer in the vertical direction to obtain a surface settlement estimate; a ground stress-settlement prediction model is constructed based on the nonlinear mapping relationship between the ground stress tensor time series data and the settlement estimate, and is trained; subsequent geothermal exploitation scheme parameters are input, and the trained ground stress-settlement prediction model is used to predict the future surface settlement trend. The application has a wide range of applications and can be applied to ground surface settlement prediction in geothermal exploitation areas under different geological conditions.

[0005] However, in actual use, the above disclosed method and the method existing in the prior art usually focus on environmental data or geological properties of the structure itself when considering influencing factors, and pay insufficient attention to the influence of human activities. Since road and railway subgrades are subjected to cumulative effects of human factors such as vehicle load, construction machinery vibration and construction activities on the stress state and settlement process of the soil body during long-term operation, especially in high-intensity traffic or frequent construction areas, the influence cannot be ignored. However, the prior art usually does not include human activity data in the model input, which makes it difficult for the prediction model to accurately reflect the settlement change rule of the subgrade under the actual operation environment, thereby affecting the prediction accuracy and the reliability of engineering decision-making. SUMMARY

[0006] The present application aims to provide a subgrade settlement prediction method based on a long short-term memory network to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: a subgrade settlement prediction method based on a long short-term memory network, comprising:

[0008] Collecting the difference value of environmental data of the target subgrade in the time series, the difference value of environmental data being used to construct an input feature matrix of the long short-term memory neural network, and the difference value of environmental data being a humidity difference value and a temperature difference value;

[0009] Correlating the axial distance of the target subgrade based on the time axis and the input feature matrix, and constructing a correlation function model;

[0010] Correcting the correlation function model according to the attribute data of the target subgrade, which is used to integrate the inherent physical properties and long-term working state constraints of the target subgrade;

[0011] Defining the mechanical response range of the target subgrade according to the human activity data corresponding to the difference value of environmental data in the time series;

[0012] Integrating the human activity data, the mechanical response range, the correlation function model and establishing a subgrade settlement prediction model of the long short-term memory network;

[0013] Inputting the real-time collected environmental data and human activity data into the trained long short-term memory network, and outputting the settlement prediction value of the target subgrade.

[0014] As a further preferred embodiment of the technical solution, the collection of the environmental data difference is performed by obtaining original environmental data of the target roadbed at each monitoring time point in a preset time sequence, the original environmental data including original temperature values and original humidity values of the target roadbed at the corresponding monitoring positions, the preset time sequence being determined according to the time span requirement of the roadbed settlement prediction, the time interval of each monitoring time point being a fixed value, and the original temperature values and the original humidity values being used to set the environmental sensors at different depths of the target roadbed to collect in real time, and the collection frequency of each sensor being consistent with the time interval of the monitoring time point.

[0015] As a further preferred embodiment of the technical solution, the attribute data includes the material type, the void ratio, the permeability coefficient, the elastic coefficient and the Poisson coefficient of the roadbed, and the correction of the correlation function is performed by introducing a material correction factor and a nonlinear cumulative response function and an event enhancement term into the axial direction environmental influence diffusion coefficient.

[0016] As a further preferred embodiment of the technical solution, the human activity data includes the vehicle load frequency, the construction machinery vibration intensity and the activity duration, the human activity data is obtained through a field monitoring device or a historical database, and the mechanical response range is defined as a maximum allowable strain threshold of the roadbed in the horizontal and vertical axial directions, and a calculation formula of the maximum allowable strain threshold is derived based on the linear elasticity theory and the dynamic load superposition principle.

[0017] As a further preferred embodiment of the technical solution, the construction method of the roadbed settlement prediction model of the long short-term memory network includes:

[0018] Data fusion processing is performed to splice the output values of the corrected correlation function model, the environmental data difference, the time sequence data, the human activity data and the boundary values of the mechanical response range, so as to construct an input feature matrix of the long short-term memory network;

[0019] Input feature normalization processing is performed to map the fused input feature matrix to the interval [0, 1];

[0020] A deep time sequence network structure of an input layer-2-layer long short-term memory layer-dropout layer-full connection layer-output layer is adopted to design the long short-term memory network structure;

[0021] The long short-term memory network is trained in an end-to-end manner.

[0022] As a further preferred embodiment of the technical solution, the correlation function is ;

[0023] wherein, is the "space-time-environment" correlation value of the axial direction position x at time t; is a time decay factor, wherein is an environmental impact feature time constant; , respectively temperature difference, humidity difference in the horizontal direction weight coefficient, subgrade deep monitoring point takes = 0.35, = 0.65, for embodying the engineering characteristics that deep subgrade is more significantly affected by humidity change.

[0024] As further preferred of the technical solution, the calculation formula of the maximum allowable strain threshold value is: , wherein is the total amount of vehicle load per unit time (KN / h), is a vehicle load influence coefficient, reflecting the difference in the contribution of different vehicle types and speeds to the subgrade strain, is a construction machinery disturbance intensity comprehensive value, is a construction disturbance influence coefficient, is the equivalent elastic modulus of the subgrade soil body, is the effective stress area of the subgrade, is a time attenuation coefficient, describing the attenuation effect of human activities on strain in different time periods, is an empirical correction coefficient.

[0025] As further preferred of the technical solution, wherein is the number of vehicle passes per unit time, is the average single vehicle load, which is obtained by laying several weighing sensors in the subgrade area, is determined by regression analysis according to the on-site vehicle load and strain monitoring data, is the vibration speed amplitude calculated by measuring the vibration intensity through the vibration monitoring sensor, and then converted to acceleration according to the frequency, and multiplied by the effective mass of the disturbed soil body to determine, is a human subjective coefficient, which is determined comprehensively according to engineering experience and on-site test data, is the modulus value at the corresponding strain level determined by triaxial test in the prior art, is estimated by wheel track width and load distribution depth, is determined by fitting according to the monitoring data, is determined according to the feature distribution of the training data set by the Bayesian optimization method.

[0026] As further preferred of the technical solution, for the fusion of each input feature matrix, the maximum value and the minimum value of each feature dimension in the historical data set are calculated, and the formula normalized value = (original value-minimum value) / (maximum value-minimum value) is used for conversion and fusion.

[0027] Compared with the prior art, the present application has the following advantages:

[0028] The long short-term memory network-based roadbed settlement prediction method effectively captures the dynamic response characteristics of the roadbed at different times and depths by collecting the environmental data difference value of the target roadbed in the time series, including the temperature difference value and the humidity difference value, and taking it as the input feature matrix of the long short-term memory network, at the same time, the correlation function model of the environmental data and the axial distance of the roadbed is established, and the material type, void ratio, permeability coefficient, elastic coefficient and Poisson coefficient of the roadbed are modified, so that the model not only has data driving ability, but also integrates the inherent physical characteristics and long-term working state constraints of the roadbed, thereby ensuring the accuracy and reliability of the prediction model;

[0029] In addition, by collecting vehicle load frequency, construction machinery vibration intensity and activity duration and other data, the mechanical response range of the roadbed is defined, and these information is fused with the environmental data difference value and the correlation function model to build a long short-term memory network settlement prediction model that can reflect the coupling influence of multiple source factors. Compared with the method of ignoring human factors or only using empirical correction in the prior art, the settlement change law of the roadbed in the actual running environment can be more accurately reflected, the interpretability and engineering applicability of the prediction result are realized, and the model parameters are optimized in an end-to-end training mode using normalization processing and a deep time sequence network structure, so that the roadbed settlement prediction value can be quickly output on the basis of real-time collection of environmental data and human activity data, providing a scientific basis for roadbed maintenance and risk assessment. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The method flowchart disclosed by the present application is shown in the figure;

[0031] Figure 2 The horizontal diffusion coefficient of the present application changes with time;

[0032] Figure 3 The time attenuation factor of the present application changes with time;

[0033] Figure 4 The present application The graph changes with time at x=2.0m;

[0034] Figure 5 The saturated response graph of the present application is shown in the figure;

[0035] Figure 6 The modified stress and its component graph of the present application is shown in the figure;

[0036] Figure 7 The maximum strain prediction graph of the present application within 24 hours is shown in the figure. DETAILED DESCRIPTION

[0037] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0038] Before the technical solutions in the present application are understood, it is necessary to make it clear that the actual use scene of the technical solutions is the field of settlement monitoring and prediction in roadbed engineering such as highways and railways, and the technical solutions are specifically applied to real-time evaluation of settlement trend, long-term stability analysis and disaster prevention in the construction process, so as to optimize engineering design parameters, improve engineering safety and prolong service life.

[0039] As shown in Figure 1 The present application provides a technical solution: a roadbed settlement prediction method based on a long short-term memory network, comprising steps S100-S600.

[0040] Step S100: collecting the difference value of the environmental data of the target roadbed in the time sequence, and the difference value of the environmental data is used to construct the input feature matrix of the long short-term memory neural network.

[0041] It should be noted that, in the technical solution, the difference value of the environmental data refers to the difference value of the humidity and the difference value of the temperature.

[0042] As a preferred embodiment, in the technical solution, since the long short-term memory neural network has strong time sequence dependence on the input features, it is difficult to effectively capture the trend characteristics of the data changing over time by directly inputting the original environmental data (such as temperature, humidity), therefore, the present solution pre-processes the original data by calculating the difference value of the environmental data.

[0043] It is worth noting that the collection of the difference value of the environmental data is by obtaining the original environmental data of the target roadbed at each monitoring time point in a preset time sequence, wherein the original environmental data includes the original temperature value and the original humidity value of the corresponding monitoring position of the target roadbed; it is worth noting that the preset time sequence is determined according to the time span requirement of the roadbed settlement prediction (the time sequence of the past 12 months is selected for predicting the settlement of the next 3 months), the time interval of each monitoring time point is a fixed value (1 hour, 1 day or 1 week, which can be adjusted artificially according to the monitoring accuracy), the original temperature value and the original humidity value are obtained by real-time collection of the environmental sensors arranged at different depths (roadbed surface, middle of roadbed filler layer, roadbed bottom) of the target roadbed, the collection frequency of each sensor is consistent with the time interval of the monitoring time point to ensure the time sequence integrity of the original data, wherein the environmental sensor is specifically a humidity sensor and a temperature sensor.

[0044] Step S200: correlate the axial distance of the target roadbed based on the time axis and the input feature matrix, and construct a correlation function model.

[0045] It is worth noting that in the present technical solution, the axial distance of the target roadbed includes horizontal axial distance, vertical distance and inclined axial distance. In addition, since the settlement of the roadbed does not occur uniformly, the settlement amount at different axial positions (roadbed center and slope, surface layer and deep layer) is significantly different in terms of the degree of influence of environmental factors (temperature fluctuation, humidity change). Therefore, in order to accurately capture the "space-time-environment" coupling characteristics of the settlement, the spatial position information needs to be fused with the time series and environmental data through the correlation function model. Therefore, in the present technical solution, the correlation function model quantifies the dynamic law of the roadbed response to environmental factors at different axial positions by establishing a nonlinear mapping relationship between the axial distance and the time series and the environmental data difference.

[0046] Specifically, the construction of the correlation function model is to define the axial distance of a monitoring point of the target roadbed as x (taking the center of the roadbed cross section as the origin, the positive direction is along the horizontal, vertical or inclined direction towards the slope side, unit: meter), the time series as the discretized monitoring time t (t=1, 2,..., n, corresponding to the continuous time points from the initial monitoring time t0 to the current time t n , the environmental data difference as the temperature difference ΔT(t) (unit: ℃) and the humidity difference ΔH(t) (unit: %RH) at the time point.

[0047] Specifically, the correlation function model is used to quantify the response sensitivity difference of different positions to environmental factors. For each time point t, the environmental influence weight coefficient is constructed by using Gaussian radial basis function (RBF), and the formula is as follows:

[0048] ;

[0049] Wherein, is the axial direction environmental influence diffusion coefficient varying with time, and the physical meaning is the influence range of environmental factors in the horizontal direction, The larger the environmental factors at the time point, the more widely the influence of the environmental factors on the axial direction of the roadbed, and vice versa, which is concentrated in the local area, The calculation of needs to combine the cumulative effect of historical environmental data, and the specific expression is as follows:

[0050] ;

[0051] In the formula, is the initial diffusion coefficient in the axial direction (determined by the physical parameters of the roadbed filler such as permeability coefficient and particle size distribution, which is a parameter in the prior art); is the humidity accumulation influence coefficient (fitted by field monitoring data); is the time window for cumulative humidity calculation (determined according to the hysteresis of subgrade settlement response, such as 7 days); is the absolute value of the humidity difference in the past time (unit: %RH), which is used to reflect the expansion effect of the environmental influence range in the axial direction due to recent humidity changes (in the rainy season , the cumulative amount is large, which means that the influence range of environmental factors on the slope area is wider).

[0052] Further, in the technical solution, the correlation function of the axial direction and the environmental data difference is:

[0053] ;

[0054] wherein, is the "space-time-environment" correlation value of the axial direction position x at time t; is the time decay factor, in the technical solution, wherein is the environmental influence characteristic time constant (such as 30 days, reflecting the influence of old environmental data on the current settlement prediction decaying over time); , and are the weight coefficients of the temperature difference and the humidity difference in the horizontal direction (in the technical solution, the surface monitoring points of the subgrade take =0.55, =0.45), and the deep monitoring points of the subgrade take =0.35, =0.65, to reflect the engineering characteristics that the deep subgrade is more significantly affected by humidity changes).

[0055] It is worth noting that the larger the axial distance x in the correlation function model of the axial direction, exponentially decays with x, which means that the monitoring points near the slope are less affected by environmental factors than the center area of the subgrade, and increases with the increase of the humidity accumulation.

[0056] It is worth noting that, as shown in Figure 2 , at the initial stage (t is less than 20 days), the value is small, indicating that the humidity change is not obvious, and the environmental influence range is limited to the local area, and with the passage of time, the cumulative humidity change increases, showing an increasing trend, reflecting the characteristics of the expansion of the environmental influence range in the rainy season or high humidity period, and in addition, as shown in Figure 3 , in the technical solution, at the initial stage (t is between 1 and 10 days), ≈1.0, which means that the recent environmental information has a higher impact weight, and the impact weight decreases to about 0.14 as time t increases to t ≈ 60 days, which means that the environmental information in the past month has an impact on the current state, and when t is greater than 100 days, ≈1.0, which means that the recent environmental information has a higher impact weight, and the impact weight decreases to about 0.14 as time t increases to t ≈ 60 days, which means that the environmental information in the past month has an impact on the current state, and when t is greater than 100 days, ≈1.0, which means that the recent environmental information has a higher impact weight, and the impact weight decreases to about 0.14 as time t increases to t ≈ 60 days, which means that the environmental information in the past month has an impact on the current state, and when t is greater than 100 days, ≈1.0, which means that the recent environmental information has a higher impact weight, and the impact weight decreases to about 0.14 as time t increases to t ≈ 60 days, which means that the environmental information in the past month has an impact on the current state, and when t is greater than 100 days, Figure 4 It is worth noting that, as shown in , at a fixed horizontal position x = 2.0 m, a clear periodic fluctuation is presented, with a wave peak corresponding to a humidity rise or temperature rise stage, and a wave trough corresponding to a dry or temperature drop stage, and the amplitude slightly decays over time, which is consistent with the time decay characteristics of .

[0057] Step S300: modifying the correlation function model according to the attribute data of the target roadbed, for integrating the inherent physical characteristics and long-term working state constraints of the target roadbed.

[0058] Specifically, in step S300, the attribute data includes the material type, porosity, permeability coefficient, elastic coefficient, and Poisson's ratio of the roadbed.

[0059] Specifically, in the present technical solution, the modification of the correlation model is performed by introducing a material correction factor and a nonlinear cumulative response function and an event enhancement term into the axial direction environmental influence diffusion coefficient.

[0060] ;

[0061] wherein , is the porosity, is the permeability coefficient, and is the elastic coefficient and Poisson's ratio, which is artificially valued between 0 and 0.5, the nonlinear cumulative response g is a saturation function, specifically, , is a quantity after the saturation function is substituted into , and γ represents the control of the degree of nonlinearity, which is artificially valued between 1 and 2 in the present technical solution, and the event enhancement term is used for rainfall events, which is the percentage of rainfall in the total month.

[0062] It is worth noting that, as shown in Figure 5 , as the value of increases, the output of gradually tends to saturation, and in the case of a small , the output of gradually tends to saturation, and in the case of a small , the output of gradually tends to saturation, and in the case of a small The model is sensitive to the changes of ; when exceeds a certain threshold, the response curve tends to be flat, which shows that the model proposed in the technical solution has a strong response to external disturbances in the early stage of cumulative impact, and the response of the system gradually weakens in the long-term cumulative or high saturation state. In addition, it can be known from Figure 6 that in the technical solution, the change trend of the total stress is basically the same as that of , but a small disturbance of the material constant baseline and event enhancement is superimposed. When the rainfall is frequent and is large, increases; and in the drought or low accumulation stage, decreases and tends to be stable, which shows that the model can reasonably reflect the axial stress correction behavior of the material caused by environmental changes (especially the penetration and rainfall conditions).

[0063] Step S400: According to the difference of environmental data corresponding to the time sequence of human activity data, the mechanical response range of the target roadbed is defined.

[0064] It should be noted that in the technical solution, step S400 is used to quantify the influence of human activities on the mechanical response of the roadbed. By integrating traffic flow data, construction disturbance intensity and time distribution information, the additional stress field and strain boundary conditions are calculated to define the dynamic mechanical response range.

[0065] Specifically, the human activity data includes vehicle load frequency (unit: times / hour), construction machinery vibration intensity (unit: dB), and activity duration (unit: hour). These data are obtained through on-site monitoring equipment or historical database, and the mechanical response range is defined as the maximum allowable strain threshold of the roadbed in the horizontal and vertical axial directions .

[0066] It should be noted that in the technical solution, the calculation formula of the maximum allowable strain threshold is: , wherein is the total amount of vehicle load per unit time (KN / h), is the vehicle load influence coefficient, which reflects the difference in strain contribution of different vehicle types and speeds to the roadbed, is the construction machinery disturbance intensity comprehensive value (kN), is the construction disturbance influence coefficient, is the equivalent elastic modulus of the roadbed soil (MPa), is the effective stress area of the roadbed (m²), is the time attenuation coefficient, which describes the attenuation effect of human activities in different time periods on strain, is the empirical correction coefficient.

[0067] It should be noted that the calculation formula of the maximum allowable strain threshold in the technical solution is derived based on the linear elasticity theory and the dynamic load superposition principle. First, the vehicle load and the construction disturbance are regarded as two main human activity sources, and the equivalent force acting on the subgrade can be represented as The equivalent force acts on the effective stress area A of the subgrade, and the additional stress generated is Under the small strain assumption, the linear elastic relationship between the subgrade strain and the additional stress is As can be seen above, by introducing the empirical correction coefficient and the decay coefficient reflecting the time action law , the maximum allowable strain threshold can be obtained. The calculation formula of the maximum allowable strain threshold can represent the equivalent mechanical response amplitude of the subgrade soil under the superimposed load of traffic and construction, and is used to evaluate the stress safety margin and dynamic response boundary of the subgrade.

[0068] It should be further supplemented that in the technical solution, is the number of vehicle passes per unit time (times / h), is the average single vehicle load (kN), which is obtained by laying a number of weighing sensors in the subgrade area, is determined by regression analysis based on field vehicle load and strain monitoring data, The vibration intensity (dB) measured by setting a vibration monitoring sensor is used to calculate the vibration velocity amplitude, and then the frequency is converted to acceleration, and the effective mass of the disturbed soil is multiplied to determine the effective mass of the disturbed soil. The effective mass of the disturbed soil is determined by field geological exploration and density testing method. Specifically, first, the average density of the soil is measured by core sampling or in-situ density meter, then the effective disturbance volume is defined based on the construction disturbance influence range, and finally the effective mass is calculated by multiplication, is a subjective coefficient, which is determined comprehensively according to engineering experience and field test data, and in the technical solution, the value range is 0.8 to 1.2, to reflect the uncertainty under different construction conditions, The modulus value under the corresponding strain level is determined by triaxial test in the prior art, It is estimated by the wheel track width and load distribution depth, and in the technical solution, the value range is 0.8 to 1.5 m², and the value is determined according to the actual road fault plane size and load penetration depth, is determined by fitting the monitoring data, and in the technical solution, the value range is 0.6 to 1.0, and the value range is obtained by fitting the exponential decay model of the field monitoring data. The specific fitting process involves analyzing the correlation between the historical human activity time series and the strain response, and optimizing the decay function parameters by the least square method to accurately reflect the influence fluctuation of traffic flow peak and construction intermittent period in different time periods, In this technical solution, the value range is from 0.5 to 1.0. The value is determined by Bayesian optimization based on the feature distribution of the training dataset. The specific optimization process involves iteratively adjusting the parameter combination and evaluating the convergence of the loss function on the validation set to balance model complexity and prediction accuracy. At the same time, the influence of different soil types and load history is considered, and the optimal value range is screened through sensitivity analysis.

[0069] In addition to the above, traffic flow Data is collected using a high-precision dynamic weighing system; construction disturbance is strong. The time decay coefficient is calculated from the acceleration signal measured by the vibration detector. Based on the time weighting of traffic flow or the vibration attenuation law, by inputting the above data into the calculation model, the following can be obtained: Figure 7 The typical result curve shown is in Figure 7 As can be seen from the content, there are obvious strain peaks during the morning and evening traffic peaks (approximately 8h and 18h), and secondary strain peaks during the construction disturbance period (10h to 16h). The overall strain curve shows a three-stage structure of "morning peak - construction peak - evening peak". Therefore, it can be seen that the calculation model proposed in this invention can effectively capture the superposition effect of traffic and construction loads and accurately reflect the contribution of human activities to the roadbed strain in the time dimension.

[0070] Step S500: Integrate human activity data, mechanical response range, correlation function model, and establish a roadbed settlement prediction model using a long short-term memory network.

[0071] Specifically, in this technical solution, the method for constructing the roadbed settlement prediction model of the long short-term memory network includes steps S501-S504.

[0072] Step S501: Data fusion processing, which involves concatenating the corrected correlation function model output value, environmental data difference, time series data, human activity data, and mechanical response range boundary values ​​to construct the input feature matrix of the long short-term memory network.

[0073] It should be noted that the fusion of step S501 is not simply data stacking, but through feature cross mining the nonlinear correlation between each feature, for example, for a certain city main road base, the "settlement difference value (2.1 mm) output by the horizontal axis correlation function" is crossed with the "surrounding building construction intensity coefficient (0.85, the value range is 0-1, the larger the value, the greater the construction impact)", to get the "2.1x0.85=1.785" fusion feature, quantifying the amplification effect of construction activities on the horizontal direction settlement difference. Cross "temperature difference value (5℃)" with "temperature sensitive weight of mechanical response range (0.7, determined by step S400)", to get "5x0.7=3.5" fusion feature, highlighting the contribution of temperature change within the allowed range of mechanical response. In this way, the input feature matrix is upgraded from "multi-source data juxtaposition" to "multi-source feature correlation", which is more in line with the learning needs of long short-term memory network for time series correlation features.

[0074] Step S502: Input feature normalization processing, using the Min-Max normalization method in the prior art to map the fused input feature matrix to the [0, 1] interval.

[0075] Specifically, for each feature dimension (correlation function output value, environmental data difference cross feature, and cultural activity intensity coefficient), the maximum and minimum values of each feature dimension in the historical data set are calculated, and the conversion is performed by the formula "normalized value=(original value-minimum value) / (maximum value-minimum value)".

[0076] It should be emphasized that due to the large differences in the dimensions of each feature (such as the unit of the correlation function output value is mm, and the cultural activity intensity coefficient has no unit) and the numerical range (such as the temperature difference value is -10℃ to 30℃, and the mechanical response range weight is 0 to 1), the unnormalized data will cause the network to be overly sensitive to features with large numerical values (such as the numerical range of the temperature difference value is much larger than the cultural activity intensity coefficient, the network pays more attention to the temperature feature and ignores the cultural activity feature), affecting the convergence speed and prediction accuracy of the model, for example, the "correlation function output value" of a certain roadbed has a historical range of 0.5mm to 12mm, and the normalized value is (x-0.5) / 11.5; the "cultural activity intensity coefficient" has a historical range of 0.1 to 0.9, and the normalized value is (y-0.1) / 0.8, ensuring that all features are on the same scale to participate in network training.

[0077] Step S503: A deep time series network structure of input layer-2 layer long short-term memory layer-dropout layer-full connection layer-output layer is used to design the long short-term memory network structure.

[0078] Specifically, the number of neurons in the input layer is equal to the dimension of the input feature matrix (e.g., if the fused feature matrix is 12-dimensional, the input layer has 12 neurons); the first long short-term memory layer is set to have 64 hidden units, and the activation function is tanh (hyperbolic tangent function, used to capture the non-linear relationship between features), and the returned sequence is used to preserve the timing information for the second long short-term memory layer; the second long short-term memory layer is set to have 32 hidden units, and the activation function is also tanh, and no sequence is returned, and a one-dimensional timing feature vector is output; the dropout rate of the dropout layer is set to 0.2 (20% of the neuron connections are randomly discarded), which is used to prevent network overfitting (i.e., to avoid the network from learning too much noise in the historical data, affecting the prediction ability for unknown data), and the fully connected layer is set to have 16 neurons, and the activation function is relu (rectified linear unit, used to introduce nonlinearity, to solve the problem that linear models cannot capture complex settlement rules); the output layer is set to have one neuron, and the activation function is linear, and the output is the settlement prediction value of the target roadbed (unit: mm).

[0079] It should be noted that the parameters of the network structure (number of long short-term memory layers, number of hidden units, dropout rate) are obtained by grid search optimization, for example, for the historical data of a certain highway roadbed (2xx8-2x22 years, a total of 5 years, 1 monitoring per month, a total of 60 samples), the combinations of "1 layer long short-term memory layer (32 / 64 / 128 hidden units)" "2 layer long short-term memory layer (32+16 / 64+32 / 128+64 hidden units)" and "dropout rate 0.1 / 0.2 / 0.3" are tested, and finally the structure of "2 layer long short-term memory layer (64+32 hidden units) + dropout rate 0.2" is selected, at this time the validation set mean square error (MSE) of the model is the smallest (0.89 mm²).

[0080] Step S504: Using the Adam optimizer (adaptive moment estimation optimizer, suitable for training of time series data) and mean square error (MSE) as the loss function, the long short-term memory network is trained end-to-end.

[0081] Specifically, the historical data set is divided into a training set (70%, such as 42 samples), a validation set (20%, such as 12 samples), and a test set (10%, such as 6 samples) in a ratio of 7:2:1; during training, the training set data is input, the predicted settlement value is output, the difference between the predicted value and the actual settlement value (such as the settlement amount in the monitoring data) is calculated by the loss function, and then the network parameters (weight matrix and bias vector of the long short-term memory layer) are updated by the back propagation algorithm in the prior art; every time an epoch (all training samples are traversed) is trained, the model performance is evaluated (MSE is calculated) using the validation set; if the validation set MSE does not decrease for 5 consecutive epochs, training is stopped (early stopping strategy to prevent overfitting); for example, during the training process of a certain roadbed long short-term memory model, the training set MSE gradually decreased from the initial 12.3 mm² to 1.1 mm², and the validation set MSE decreased from the initial 11.8 mm² to 0.89 mm²; when training to the 30th epoch, the validation set MSE remained between 0.89-0.91 mm² for 5 consecutive epochs, at which point training was stopped, and the final long short-term memory network roadbed settlement prediction model was obtained.

[0082] Step S600: input the real-time collected environmental data and cultural activity data into the trained long short-term memory network, and output the settlement prediction value of the target roadbed.

[0083] Specifically, the output process of the settlement prediction value includes the following steps:

[0084] Data collection: real-time collection of environmental data (temperature 25°C, humidity 60%, roadbed top load 0.15 MPa) and cultural activity data at the current time point through the roadbed monitoring system (buried sensor);

[0085] Data preprocessing: calculate the environmental data difference according to step S100, calculate the mechanics response range boundary value according to step S400, perform feature fusion according to step S501, and perform normalization processing according to step S502;

[0086] Model prediction: input the preprocessed input feature matrix into the long short-term memory network, and output the settlement prediction value at the current time point;

[0087] Result post-processing: compare the predicted value with the preset settlement warning threshold (set according to the settlement warning threshold in the prior art “Highway Roadbed Design Specification”); if the predicted value exceeds the threshold, output “normal” warning; if it exceeds the threshold, output “warning”.

[0088] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the embodiments disclosed except insofar as recited in the claims.

Claims

1. A method for predicting roadbed settlement based on long short-term memory networks, characterized in that, include: The environmental data difference of the target roadbed within a time series is collected. The environmental data difference is used to construct the input feature matrix of a long short-term memory neural network. The environmental data difference is the humidity difference and the temperature difference. The axial distances of the target roadbeds are correlated based on the time axis and the input feature matrix, and a correlation function model is constructed. The correlation function model is modified based on the attribute data of the target roadbed to incorporate the inherent physical characteristics and long-term working state constraints of the target roadbed. Based on the human activity data within the time series corresponding to the environmental data differences, the mechanical response range of the target roadbed is defined; A roadbed settlement prediction model was established by integrating human activity data, mechanical response range, correlation function model, and long short-term memory network. The real-time collected environmental data and human activity data are input into the trained long short-term memory network, and the predicted settlement value of the target roadbed is output. The correlation function is: ; in, ; in, The axial diffusion coefficient represents the environmental impact over time, and its physical meaning is the extent of the environmental factor's influence in the horizontal direction. The larger the value, the wider the influence of environmental factors on the axial direction of the roadbed at the current point in time; conversely, the smaller the value, the more concentrated the influence is on a local area. The calculation needs to take into account the cumulative effect of historical environmental data, and the specific expression is as follows: ; In the formula, The initial diffusion coefficient is in the axial direction; This is the cumulative humidity effect coefficient; The time window for calculating cumulative humidity; For the past The cumulative absolute value of humidity difference over a period of time is used to reflect the expanding effect of recent humidity changes on the axial environmental influence range. The "space-time-environment" correlation value of the axial position x at time t; The time decay factor, ,in The time constant representing the environmental impact characteristics; , These are the weighting coefficients for temperature difference and humidity difference in the horizontal direction, respectively. The deep monitoring points of the roadbed are... =0.35、 =0.65, used to reflect the engineering characteristics of deep subgrades being more significantly affected by humidity changes; The attribute data includes: the material type, porosity, permeability coefficient, elastic coefficient and Poisson's coefficient of the roadbed. The correlation function is corrected by introducing a material correction factor, a nonlinear cumulative response function and an event enhancement term into the axial direction environmental influence diffusion coefficient. The data on human activities includes: vehicle load frequency, construction machinery vibration intensity, and activity duration. The data on human activities is obtained through on-site monitoring equipment or historical databases. The mechanical response range is defined as the maximum allowable strain threshold of the roadbed in the horizontal and vertical axes. The calculation formula for the maximum allowable strain threshold is derived based on the linear elastic theory and the principle of dynamic load superposition.

2. The method for predicting roadbed settlement based on long short-term memory networks according to claim 1, characterized in that: The collection of environmental data difference is achieved by acquiring the original environmental data of the target roadbed at each monitoring time point within a preset time series. The original environmental data includes the original temperature and original humidity values ​​at the corresponding monitoring locations of the target roadbed. The preset time series is determined based on the time span requirements of the roadbed settlement prediction. The time interval between each monitoring time point is a fixed value. The original temperature and original humidity values ​​are acquired in real time by environmental sensors set at different depths of the target roadbed. The acquisition frequency of each sensor is consistent with the time interval between the monitoring time points.

3. The method for predicting roadbed settlement based on long short-term memory networks according to claim 1, characterized in that: Methods for constructing roadbed settlement prediction models using long short-term memory networks include: Data fusion processing involves concatenating the corrected correlation function model output value, environmental data difference, time series data, human activity data, and mechanical response range boundary values ​​to construct the input feature matrix of the long short-term memory network. Input feature normalization process maps the fused input feature matrix to the [0,1] interval; A deep temporal network structure is used to design the long short-term memory network structure, which consists of an input layer, two long short-term memory layers, a dropout layer, a fully connected layer, and an output layer. End-to-end training of the Long Short-Term Memory Network.

4. The method for predicting roadbed settlement based on long short-term memory networks according to claim 1, characterized in that: The formula for calculating the maximum allowable strain threshold is: ,in This refers to the total vehicle load per unit time. The vehicle load influence factor reflects the differences in the contribution of different vehicle types and speeds to roadbed strain. This represents the comprehensive value of the disturbance intensity caused by construction machinery. This is the construction disturbance impact coefficient. The equivalent elastic modulus of the roadbed soil. This is the effective stress-bearing area of ​​the roadbed. This is the time decay coefficient, which describes the decay effect of human activities on strain over different time periods. This is an empirical correction factor.

5. The method for predicting roadbed settlement based on long short-term memory networks according to claim 4, characterized in that: The determination was made based on regression analysis of on-site vehicle load and strain monitoring data. The vibration intensity is measured by vibration monitoring sensors to calculate the vibration velocity amplitude, which is then converted into acceleration based on the frequency. This acceleration is then determined by multiplying it by the effective mass of the disturbed soil. This is a subjective coefficient determined based on engineering experience and on-site test data. The modulus value at the corresponding strain level is determined by triaxial testing using existing technology. Estimation is made using wheel track width and load distribution depth. Determined based on the fitting of monitoring data. The value of is determined by Bayesian optimization based on the feature distribution of the training dataset.

6. The method for predicting roadbed settlement based on long short-term memory networks according to claim 3, characterized in that: For the fusion of each input feature matrix, the maximum and minimum values ​​of each feature dimension in the historical dataset are calculated, and the normalized value is transformed and fused using the formula: normalized value = (original value - minimum value) / (maximum value - minimum value).

Images