Pavement settlement real-time monitoring and predicting model training method based on multiple physical fields

Abstract

The invention provides a pavement settlement real-time monitoring and prediction model training method based on multiple physical fields, relates to the field of traffic safety, and solves the technical problems that the whole monitoring period is long, rapid monitoring cannot be performed, and dynamic changes of pavement settlement cannot be reflected in real time in the prior art. The method comprises the following steps: acquiring a Rayleigh wave time sequence signal, a time sequence signal of electric field and magnetic field data and a time sequence signal of natural potential data acquired by a micro-motion, magnetotelluric field and natural potential detection system; feature extraction and multi-source data fusion are carried out on the Rayleigh wave time sequence signals, the time sequence signals of the electric field and magnetic field data and the time sequence signals of the natural potential data, weighted feature vectors are constructed, division is carried out according to time windows, a structured time sequence matrix is formed, joint training is carried out, and a trained prediction model is obtained. The method is used in the process of predicting the specific condition of pavement settlement.

Description

Technical Field

[0001] The present application relates to the field of traffic safety, and in particular to a method for training a real-time monitoring and prediction model for road subsidence based on multi-physical fields. Background Art

[0002] Pavement subsidence is a common problem in the construction and operation of transportation infrastructure, particularly on highways, urban roads, and railway subgrades. Subsidence can lead to uneven surfaces, cracks, and even collapses, seriously impacting driving safety and road life. Furthermore, urban road subsidence can damage underground pipelines, impacting the normal operation of urban infrastructure.

[0003] Traditional methods for monitoring pavement subsidence rely primarily on technologies such as leveling and GPS. While these methods offer high accuracy, they often require manual monitoring of several monitoring points and calculations based on the data to determine pavement subsidence. However, these methods suffer from the disadvantages of long monitoring cycles and the inability to rapidly monitor the dynamic changes in pavement subsidence. Therefore, a multi-physics-based real-time monitoring and prediction model training method for pavement subsidence is urgently needed. This trained model can be used to monitor and predict pavement subsidence in real time. Summary of the Invention

[0004] The present application provides a method for training a real-time monitoring and prediction model for road subsidence based on multi-physical fields, which solves the technical problems of the prior art in that the entire monitoring cycle is long, rapid monitoring cannot be performed, and the dynamic changes of road subsidence cannot be reflected in real time.

[0005] To achieve the above objectives, this application adopts the following technical solutions:

[0006] In a first aspect, a method for training a road subsidence real-time monitoring prediction model based on multi-physics fields is provided, comprising:

[0007] Acquire the Rayleigh wave time series signals, electric and magnetic field data time series signals, and self-potential data time series signals collected by the micro-motion, magnetotelluric field and self-potential detection systems;

[0008] Perform feature extraction and multi-source data fusion on the time series signals of Rayleigh waves, electric and magnetic field data, and natural potential data, construct weighted feature vectors, and divide them into time windows to form a structured time series matrix.

[0009] The structured time series matrix and the road settlement values ​​are divided into a training set, a validation set, and a test set according to a preset ratio;

[0010] Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters;

[0011] Perform joint training based on all parameters of the pre-trained prediction model;

[0012] After the joint training is completed, the model fine-tuning is triggered based on the newly added time series signal to obtain the trained prediction model.

[0013] Based on the above technical solution, in the prediction model training method provided in the present application, real-time data is collected by the micro-motion, magnetotelluric field and natural potential detection system, and the model is trained by the real-time data. Subsequently, the Rayleigh wave time series signals, the time series signals of the electric field and magnetic field data, and the time series signals of the natural potential data collected by the micro-motion, magnetotelluric field and natural potential detection system are directly input into the trained prediction model, which can monitor the road surface settlement in real time. In addition, the physical branch of the initial model is pre-trained before training, which can enable the model to better understand and embed physical laws. When facing the monitoring and prediction of road surface settlement in multiple physical fields, predictions that violate physical common sense can be reduced as much as possible. In addition, the newly collected Rayleigh wave time series signals, the time series signals of the electric field and magnetic field data, and the time series signals of the natural potential data are used as data sources for model fine-tuning, which can make the model more in line with the current situation to further improve the accuracy of the prediction.

[0014] In combination with the first aspect above, in one possible implementation, the loss function of the N rounds of pre-training process is: Among them, except for the loss mixing coefficient of the physical parameter mapping layer which is non-zero, S phys is the collected physical value; y true is the pre-trained prediction value, and MSE is the mean square error.

[0015] In combination with the first aspect above, in one possible implementation, the loss function of the joint training based on all parameters is: in,, is the predicted settlement sequence, Y true is the true settlement sequence, R is the residual vector, ‖‖2 is the L2 norm, which is used to quantify the amplitude of the residual or vector. is the first-order difference of the predicted value, is the first-order difference of historical data.

[0016] In combination with the first aspect above, in one possible implementation, in the process of triggering model fine-tuning based on the newly added Rayleigh wave time series signal, the time series signal of electric and magnetic field data, and the time series signal of natural potential data, elastic weight integration (EWC) is used to prevent catastrophic forgetting; wherein, in, is the total loss function, F i is the diagonal element of the ith parameter in the Fisher information matrix, θ i is the current value of the i-th parameter of the model, θ i,old is the historical value of the i-th parameter of the model.

[0017] In conjunction with the first aspect above, in one possible implementation, after fine-tuning the model based on the newly added Rayleigh wave time series signal, the time series signal of electric and magnetic field data, and the time series signal of spontaneous potential data, the method further includes: performing a lightweight design on the prediction model;

[0018] The lightweight design method is to add a channel compression module to the last layer of the backbone network: H compressed =ConvlD 1×1 (H TCN , channels = 16); the backbone network uses the time convolution network TCN; Conv1D 1×1 To apply 1×1 convolution to the output of layer 1, H TCN is the output of the temporal convolutional network, channels=16 means the number of output channels is set to 16.

[0019] In combination with the first aspect above, in a possible implementation, the architecture of the prediction model includes an input layer, a backbone network TCN, a physical constraint branch, a residual fusion mechanism, a loss function, and an output layer;

[0020] The output content of the output layer includes predicted values, real-time risk scores and early warning signals;

[0021] It should be noted that the predicted value is a real-time predicted value of road subsidence.

[0022] Method for obtaining the real-time risk score: in, Where p t For real-time risk scoring, is the predicted settlement at the current time step, is the predicted settlement at the i-th time step in the past 30 days;

[0023] The warning signal is divided into warning levels according to real-time scores.

[0024] In combination with the first aspect above, in a possible implementation, the input layer is a high-dimensional time series matrix obtained by aligning and fusion of structured time series matrices according to time windows.

[0025] In combination with the first aspect above, in a possible implementation, the backbone network TCN is composed of 4 layers of dilated causal convolution, residual connection and skip connection;

[0026] The calculation formula of the dilated causal convolution is: H l =ReLU(W l * d H l-1 +b l ); where H l is the output of the lth layer, W l is the convolution kernel weight matrix of layer i, * d represents the convolution operation with a dilation factor of d, b l is the bias term of the lth layer;

[0027] The calculation formula of the residual connection is: in, is the final output of the lth layer;

[0028] The calculation formula of the skip connection is: Among them, H skip The final output of the backbone network.

[0029] It should be noted that the above H TCN =ReLU(H skip ).

[0030] In combination with the first aspect above, in a possible implementation, the physical constraint branch is an embedded geomechanical equation;

[0031] The embedded geomechanical equation is: Where τ is the one-dimensional consolidation equation fitted by the historical settlement curve: E is the elastic modulus, E = ρ·v 2 , ρ is soil density, v is electromagnetic wave velocity; ΔP is pore pressure change, g is the acceleration due to gravity, is the natural potential gradient, k is the permeability coefficient, σ is the electrical conductivity of the soil, h is the thickness of the soil layer, α is the regional experience coefficient, S max is the maximum allowable compression of the soil or the design settlement limit, t is the time variable (unit: s), which indicates the duration of the consolidation process;

[0032] In a second aspect, an electronic device is provided, comprising: a communication unit and a processing unit; the communication unit is used to obtain Rayleigh wave time series signals, time series signals of electric and magnetic field data, and time series signals of self-potential data collected by micro-motion, magnetotelluric field, and self-potential detection systems; the processing unit is used to perform feature extraction and multi-source data fusion on the Rayleigh wave time series signals, the time series signals of electric and magnetic field data, and the time series signals of self-potential data to obtain a structured time series matrix;

[0033] The structured time series matrix and the road settlement values ​​are divided into a training set, a validation set, and a test set according to a preset ratio;

[0034] Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters;

[0035] It should be noted that the physical parameters here are those in the geomechanical equations, including elastic modulus, pore pressure change, and soil thickness;

[0036] Perform joint training based on all parameters of the pre-trained prediction model;

[0037] All parameters here include physical parameters and model parameters; model parameters include weights, biases, and residual connection parameters in the model.

[0038] After the joint training is completed, the model is fine-tuned based on the newly added Rayleigh wave time series signals, the time series signals of electric and magnetic field data, and the time series signals of natural potential data to obtain the trained prediction model.

[0039] It should be noted that the newly added time series signals have the characteristics of multi-physics fields, real-time, spatiotemporal alignment, and sensitivity to dynamic changes, which are used to continuously improve the model's prediction capabilities.

[0040] In a third aspect, the present application provides an electronic device comprising: a processor and a storage medium; the storage medium comprises instructions, and the processor is used to execute the instructions to implement the method described in the first aspect and any possible implementation manner of the first aspect.

[0041] In the fourth aspect, the present application provides a real-time monitoring and prediction model training system for road settlement based on multiple physical fields, including: communication equipment, edge computing nodes and training platforms; wherein, the communication equipment 101 is used to obtain the Rayleigh wave time series signals, the time series signals of electric field and magnetic field data, and the time series signals of natural potential data collected by the micro-motion, magnetotelluric field and natural potential detection systems; the edge computing node 102 is used to perform feature extraction and multi-source data fusion on the Rayleigh wave time series signals, the time series signals of electric field and magnetic field data, and the time series signals of natural potential data to obtain a structured time series matrix; the training platform 103 is used to divide the structured time series matrix and the settlement value of the road surface into a training set, a verification set and a test set according to a preset ratio, construct an initial prediction model, and perform N rounds of pre-training based on the physical branches of the physical parameters of the road surface, and perform joint training on the pre-trained prediction model based on all parameters. After the joint training is completed, the model fine-tuning is triggered based on the newly added Rayleigh wave time series signals, the time series signals of electric field and magnetic field data, and the time series signals of natural potential data to obtain a trained prediction model.

[0042] In a fifth aspect, the present application provides a computer-readable storage medium, which stores instructions. When the instructions are executed on an electronic device, the electronic device executes the method described in the first aspect and any possible implementation of the first aspect.

[0043] In a sixth aspect, the present application provides a computer program product comprising instructions, which, when executed on an electronic device, enables the electronic device to execute the method as described in the first aspect and any possible implementation manner of the first aspect.

[0044] The present application provides a real-time monitoring and prediction model training method and electronic equipment for road subsidence based on multiple physical fields. The method takes Rayleigh wave time series signals, time series signals of electric field and magnetic field data, and time series signals of natural potential data, and performs multi-source data fusion on the Rayleigh wave time series signals, the time series signals of electric field and magnetic field data, and the time series signals of natural potential data to obtain a structured time series matrix. The prediction model is trained by the structured time series matrix to obtain the collected data, which can be directly input into the prediction model after simple feature processing. The prediction model can directly and quickly obtain the specific situation of road subsidence. In this way, the monitoring cycle length is greatly shortened to achieve real-time monitoring of road subsidence. In addition, by pre-training the physical branch when training the prediction model to be trained, it is possible to avoid the output of the trained prediction model that violates physical common sense. When using the prediction model, the prediction model is fine-tuned by the newly added data, thereby further ensuring that the model can be applied to scene changes and special scenes.

[0045] It should be understood that the description of technical features, technical solutions, beneficial effects or similar language in this application does not imply that all features and advantages can be realized in any single embodiment. On the contrary, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution or beneficial effect is included in at least one embodiment. Therefore, the description of a technical feature, technical solution or beneficial effect in this specification does not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions and beneficial effects described in the present embodiment can also be combined in any appropriate manner. Those skilled in the art will understand that the embodiment can be implemented without one or more specific technical features, technical solutions or beneficial effects of a specific embodiment. In other embodiments, additional technical features and beneficial effects can also be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A system architecture diagram of a road subsidence real-time monitoring and prediction model training system based on multi-physics fields provided in an embodiment of the present application;

[0047] Figure 2 A flowchart of a method for training a road surface settlement real-time monitoring and prediction model based on multi-physics fields provided in an embodiment of the present application;

[0048] Figure 3 A schematic structural diagram of an electronic device provided in an embodiment of the present application;

[0049] Figure 4 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In the description of this application, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, "at least one" means one or more, and "a plurality" means two or more. Words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not limit them to be necessarily different.

[0051] It should be noted that, in this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0052] The present invention provides a method for training a road subsidence real-time monitoring prediction model based on multi-physics fields, which can be applied to Figure 1 In the road settlement real-time monitoring prediction model training system 100 based on multi-physics fields, as shown in FIG. Figure 1 As shown, the communication system includes: communication equipment 101, edge computing node 102 and training platform 103.

[0053] The communication device 101 is used to obtain the Rayleigh wave time series signals, the electric field and magnetic field data time series signals, and the self-potential data time series signals collected by the micro-motion, magnetotelluric field and self-potential detection systems.

[0054] The edge computing node 102 is used to perform feature extraction and multi-source data fusion on the Rayleigh wave time series signal, the electric field and magnetic field data time series signal, and the natural potential data time series signal to obtain a structured time series matrix.

[0055] The training platform 103 is used to divide the structured time series matrix and the road surface settlement value into a training set, a validation set and a test set according to a preset ratio;

[0056] Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters;

[0057] Perform joint training based on all parameters of the pre-trained prediction model;

[0058] After the joint training is completed, the model is fine-tuned based on the newly added Rayleigh wave time series signals, the time series signals of electric and magnetic field data, and the time series signals of natural potential data to obtain the trained prediction model.

[0059] In order to solve the technical problem that the existing technology cannot monitor road subsidence in real time, the embodiment of the present application provides a method for training a real-time monitoring prediction model for road subsidence based on multi-physical fields, the method comprising: obtaining Rayleigh wave time series signals, time series signals of electric field and magnetic field data, and time series signals of natural potential data collected by micro-motion, magnetotelluric field and natural potential detection systems; performing feature extraction and multi-source data fusion on the Rayleigh wave time series signals, time series signals of electric field and magnetic field data, and time series signals of natural potential data to obtain a structured time series matrix; training a model based on the structured time series matrix, and thereby obtaining a trained prediction model. In subsequent use, the multi-source data collected in real time is directly input into the prediction model, which can quickly obtain real-time road subsidence conditions and issue early warnings according to specific conditions, thereby solving the current problem of being unable to reflect the dynamic changes of road subsidence in real time, and thus having the disadvantages of a long monitoring cycle and being unable to perform real-time dynamic monitoring.

[0060] like Figure 2 As shown, the embodiment of the present application provides a method for training a road surface settlement real-time monitoring prediction model based on multi-physics fields, including:

[0061] S201 , obtaining, through the communication device 101 , Rayleigh wave time series signals, electric field and magnetic field data time series signals, and self-potential data time series signals collected by the micro-motion, magnetotelluric field, and self-potential detection system.

[0062] The micro-motion detection system before step S201 obtains micro-motion data through micro-motion sensors, and the specific micro-motion data can be vibration waveform, vibration propagation direction, etc., and the data collected by the magnetotelluric field monitoring system can be electric field components, magnetic field components, etc. Different types of sensors can be used when collecting corresponding data, including but not limited to electric field sensors and magnetic field sensors. The data collected by the natural point monitoring system can be potential gradient, natural potential time series and current density distribution, etc.

[0063] In some implementations, the system acquisition consists of a central control console, a signal transmission system, a ground signal three-component electromagnetic sensor, a linear micro-motion array, electrodes, a ground data acquisition station, and a signal transmission line.

[0064] a.Electromagnetic sensor monitoring

[0065] Layout of measuring lines and measuring points: The measuring lines should be laid out along the direction of the highway to monitor the settlement changes of the entire road section; the spacing between measuring points should be designed according to the monitoring task, and the density of measuring points should be adjusted according to the geological conditions and settlement risks of the monitoring area. In areas with higher settlement risks, the density of measuring points should be appropriately increased; measuring points should be laid out inside the roadbed or under the shoulder to ensure that the sensors can accurately reflect the roadbed settlement.

[0066] Collection station and sensor layout: Each collection station should include a data acquisition module that can collect sensor data in real time.

[0067] The collection station must have protective functions such as waterproof, moisture-proof, and shock-proof to adapt to the complex environment along the highway.

[0068] Control system layout: The central control console is arranged in the road command and dispatch control center or the ground survey department, and is manned by a dedicated person to issue monitoring commands.

[0069] Monitoring time: Determine the monitoring period based on project requirements and monitoring objectives; dynamically adjust the monitoring time based on real-time monitoring data and environmental changes; in areas with complex geological conditions or high subsidence risks, the monitoring frequency should be appropriately increased.

[0070] b. Micro-motion monitoring

[0071] Layout of measuring points and lines: Select a linear array, and the measuring lines should be laid out as close to the direction of the highway as possible to better monitor geological changes along the highway; increase the density of measuring lines in areas with complex geological conditions or high potential risks (such as landslides, collapses, and mined-out areas).

[0072] Collection station and sensor layout: Collection stations should be located in areas with flat and stable terrain; avoid being close to interference sources such as large machinery and high-voltage lines to ensure data stability and reliability; the equipment should have the function of on-site evaluation of the quality of collected data to ensure data quality.

[0073] Control system layout: The control system should be deployed in an area that is geologically stable and easy to maintain, and should be manually inspected regularly to ensure real-time continuous monitoring.

[0074] Monitoring time: Dynamically adjust monitoring time based on real-time monitoring data and environmental changes; when potential subsidence risks are discovered, increase monitoring frequency and duration to obtain more detailed data.

[0075] c. Natural potential monitoring

[0076] Electrode layout: The electrodes should be buried in moist soil as much as possible to avoid interference from debris such as gravel and grass. If the measuring point is located in a high-resistance medium such as rock or concrete ground, it needs to be grounded with moist soil. The electrode leads should be tightly connected to the measuring wires, and the connection points and exposed metal parts of the electrodes should be strictly prevented from contacting the ground or weeds to avoid electromagnetic interference.

[0077] Survey line layout: Survey lines are laid out along the direction of the highway to facilitate better real-time monitoring of ground subsidence.

[0078] Control system layout: When laying cables and installing equipment, effective anti-interference measures should be taken, such as using shielded cables and proper grounding.

[0079] Monitoring time: Adjust according to real-time data and external environment; during the monitoring time, analyze the collected data in real time, detect abnormal situations in time and adjust the monitoring strategy; reasonably allocate monitoring resources to avoid excessive monitoring during non-critical periods and improve monitoring efficiency.

[0080] S202 , performing feature extraction and multi-source data fusion on the Rayleigh wave time series signal, the time series signal of the electric field and magnetic field data, and the time series signal of the natural potential data through the edge computing node 102 to obtain a structured time series matrix.

[0081] Exemplarily, feature extraction is performed on the data acquired by the micro-motion sensor, including:

[0082] 1) Time domain characteristics:

[0083] Hilbert-Huang transform (HHT): Where x(t) is the original signal. IMF i (t) is the eigenmode function, r n (t) is the residual term;

[0084] Extract instantaneous energy For each IMF i (t) Perform Hilbert transform to obtain its instantaneous amplitude A i (t) and instantaneous phase φ i (t): IMF i (t) = A i (t)cos(φ i (t)), directly square the instantaneous amplitude:

[0085] Peak-to-Peak: Divide the signal into sliding windows of fixed length. The difference between the maximum and minimum values ​​of the entire signal is: Peak-to-Peak = max(x(t)) - min(x(t)).

[0086] RMS: Global RMS: The root mean square of the entire signal: Window RMS: Calculated within a sliding window: Window

[0087] Sliding window statistics:

[0088]

[0089] 2) Frequency domain characteristics:

[0090] Windowed FFT (Hanning window N=256, window length): Where X(f) is the spectrum amplitude at frequency f, t is the time index, N is the window length, and x win (t) is the value of the windowed signal at time t, e -j2πft / N is the complex exponential function (the basis function of Fourier transform);

[0091] Extract the main frequency f peak , Band Energy (0-10Hz, 10-20Hz): Within the full spectrum range (0≤f≤f s / 2), find the frequency point with the maximum amplitude: Focus on a specific frequency band (e.g. 0-10Hz, 10-20Hz) and search for the maximum value only within that range.

[0092] 3) Spatial characteristics:

[0093] Kriging interpolation: in, The predicted value at the target point s0, λ i is the weight coefficient, z(s i ) is a known point s i The observation value at s0 is the target point, s i is a known point;

[0094] Combined with semivariogram modeling, the target point prediction value and spatial gradient are output. The specific steps include:

[0095] S31. Combined semivariogram modeling:

[0096] S311 calculates the experimental semivariogram:

[0097] Where h is the distance between the point pairs (hysteresis distance). N(h) is the number of point pairs with a distance of h, z(s i ) is the position s i The observed value at the location (such as the vibration intensity measured by the micro-motion sensor).

[0098] For a known data set, calculate the semivariogram γ(h) for different values ​​of h. Draw a plot of the experimental semivariogram (γ(h) vs. h).

[0099] S312 Selection Theory Semivariation Model:

[0100] Choose a spherical model for moderate spatial correlation: Where c0 is the nugget value (random noise), c is the structural variance (the spatially correlated part), and a is the range (the maximum distance of correlation).

[0101] S313. Fitting model parameters:

[0102] Fit the theoretical model to the experimental semivariogram using least squares or maximum likelihood methods.

[0103] Output: Determine parameters such as c0, c, a, and complete semivariogram modeling.

[0104] S32: Solving the Kriging Equations: Calculating the Weight λ i , get the predicted value of the target point s0

[0105] S321 constructs the Kriging equations:

[0106] The form of the system of equations is:

[0107] γ(s i ,s j ) is point s i and s j is the semivariogram between (calculated based on the fitted model), and μ is the Lagrange multiplier (used to constrain the sum of weights to 1).

[0108] S322 solves weight λ i : Write the system of equations in matrix form Aλ=b and solve it using linear algebra methods.

[0109] S323 calculation prediction value formula:

[0110] S33. Spatial gradient calculation: Estimate the spatial rate of change of the attribute value at the target point s0 (such as sedimentation gradient).

[0111] Based on Kriging weights: It is a linear combination of the known point values, and the gradient can be calculated directly from the weights and the known point coordinates:

[0112] Feature extraction of data acquired by the magnetotelluric field monitoring system, including:

[0113] 1) Time domain characteristics:

[0114] Abnormal pulse detection: Pulse strength = max(x(t)) - sliding mean(x(t)); where x(t) is the time domain data collected by the original signal sensor. max(x(t)) is the maximum instantaneous value in the positioning signal.

[0115] Trend fitting: Use third-order polynomial to fit signal trend;

[0116] Extract coefficients a, b, and c.

[0117] 2) Frequency domain characteristics:

[0118] Multi-band energy analysis: Divide the frequency bands (0-5Hz, 5-15Hz, 15-30Hz) and calculate the energy contribution of each band.

[0119] Determine the frequency index: according to the sampling rate f s And FFT length N, calculate the frequency index range of each frequency band:

[0120] 0-5Hz: k start =0,

[0121] 5-15Hz:

[0122] 15-30Hz: Energy calculation

[0123] Calculate the energy of each frequency band: sum the squares of the amplitudes within each frequency band:

[0124] 0-5Hz:

[0125] 5-15Hz:

[0126] 15-30Hz:

[0127] Calculate total energy

[0128] E total =E 0-5 +E 5-15 +E 15-30

[0129] Energy ratio calculation: calculate the energy ratio of each frequency band

[0130] 0-5Hz:

[0131] 5-15Hz:

[0132] 15-30Hz:

[0133] 3) Spatial characteristics:

[0134] Measurement point correlation matrix:

[0135] Output the eigenvalues ​​of the correlation matrix (such as the maximum eigenvalue, matrix rank). Among them, Cov(z(s i ),z(s j )): covariance of the observation values ​​of measurement point i and measurement point j, The standard deviation z(s i ):Observation value of measuring point i, si ,s j : The spatial position of measurement point i and measurement point j outputs the eigenvalues ​​of the correlation matrix (such as the maximum eigenvalue, matrix rank) for dimensionality reduction and feature extraction: simplifying data complexity and improving model efficiency;

[0136] Feature extraction is performed on the data collected by the natural potential detection system, including:

[0137] 1) Time domain characteristics:

[0138] o Trend and Seasonality Decomposition: Use STL decomposition (Seasonal-Trend Decomposition): z(t) = Trend(t) + Seasonal(t) + Residual(t), where z(t) represents the raw data reflecting actual observations. Trend(t) represents the long-term trend quantifying slow geological evolution, Seasonal(t) represents cyclical fluctuations associated with external environmental factors (such as rainfall and temperature), and Residual(t) represents random noise or anomalies, supporting real-time monitoring and early warning.

[0139] Extract the trend slope:

[0140] 2) Spatial characteristics:

[0141] Fluid diffusion model interpolation: Construct groundwater velocity field based on Darcy's law and predict the spatial distribution of potential: in, is the gradient sign, K is the permeability coefficient, h is the head height, and S is the water storage coefficient.

[0142] In some implementations, multi-source data fusion is achieved using feature-level fusion and spatiotemporal alignment mechanisms, and feature-level fusion is performed, specifically including:

[0143] S41. Use feature-level fusion and spatiotemporal alignment mechanisms to achieve multi-source data fusion.

[0144] 1) Space-time alignment

[0145] (a) Time synchronization:

[0146] The timestamps are unified through the GPS / NTP protocol, high-frequency data (micro-motion) are downsampled to 10 Hz, and low-frequency data (spontaneous potential) are interpolated to 10 Hz.

[0147] Interpolation method: x interp (t) = CubicSpline(t raw ,x raw )(t), where t raw : original time point series, x raw: original observation sequence, CubicSpline: cubic spline interpolation function, t target interpolation time point, x interp (t): estimated value after interpolation.

[0148] (b) Spatial alignment:

[0149] The sensor coordinates are mapped to the UTM grid coordinate system to construct a local spatial grid (resolution 1m×1m).

[0150] Coordinate transformation formula: (x utm ,y utm )=UTM_from_WGS84(lon,lat); where lon,lat: geographic coordinates → used for global positioning. utm ,y utm : Plane coordinates → used for local spatial analysis and engineering applications. UTM_from_WGS84: Coordinate conversion bridge → ensures spatial alignment of multi-source data, supporting feature fusion and visualization.

[0151] S42. Feature-level fusion

[0152] 1) Feature stitching:

[0153] Align the time domain, frequency domain, and spatial features of each sensor according to the time window to form a high-dimensional time series matrix:

[0154] X=[F 微动 , F 电磁 , F 电位 ]∈R T×D ; Where T is the number of time steps, D is the total feature dimension), and R is a set of real numbers, emphasizing the numerical characteristics of the input data (continuous and computable);

[0155] 2) Feature Normalization:

[0156] Perform Z-Score normalization independently for each feature dimension:

[0157]

[0158] Where x: original data value μ: mean data set center position, σ: standard deviation, x norm : The normalized value.

[0159] S203 , dividing the structured time series matrix into a training set, a validation set, and a test set in a ratio of 8:1:1 through the training platform 103 .

[0160] It should be noted that the output content corresponding to the structured timing matrix is ​​the same as the output content of the output layer described below.

[0161] In some implementations, stratified sampling is required according to proportion when executing step S203 to ensure that the distribution of samples of different categories or different time periods in each subset (training, validation, test) is consistent with the overall data.

[0162] In some implementations, time travel is strictly prohibited when dividing the training set, validation set, and test set to ensure that the time range of the test set is completely later than the training set and validation set, preventing the model from being exposed to data from future time points during the training phase.

[0163] S204: Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the physical parameters of the road surface.

[0164] In some implementations, the loss function for N rounds of pre-training is: Among them, except for the loss mixing coefficient of the physical parameter mapping layer which is non-zero, S phys is the collected physical value; y true is the pre-trained prediction value, and MSE is the mean squared error. By optimizing only the mapping layer of physical parameters, we can ensure that the prediction results are physically reasonable and avoid the outputs that violate physical common sense that may be produced by purely data-driven models.

[0165] S205: Perform joint training on the pre-trained prediction model based on all parameters.

[0166] In some implementations, after multiple rounds of joint training, all parameters are released, and the loss function of the joint training based on all parameters is: in, is the predicted settlement sequence, Y true is the true settlement sequence, R is the residual vector, ‖‖2 is the L2 norm, which is used to quantify the amplitude of the residual or vector. is the first-order difference of the predicted value, is the first-order difference of historical data.

[0167] S206. After the joint training is completed, fine-tuning the model is triggered based on the newly added Rayleigh wave time series signal, the time series signal of the electric and magnetic field data, and the time series signal of the spontaneous potential data to obtain a trained prediction model.

[0168] In some implementations, in the process of triggering model fine-tuning based on the newly added Rayleigh wave time series signals, electric field and magnetic field data time series signals, and natural potential data time series signals, elastic weight integration (EWC) is used to prevent catastrophic forgetting; wherein, in, is the total loss function, F iis the diagonal element of the ith parameter in the Fisher information matrix, θ i is the current value of the i-th parameter of the model, θ i,old is the historical value of the i-th parameter of the model.

[0169] In some implementations, the training platform determines whether the loss function of the joint training meets a pre-set training completion condition. If so, the training platform stops iterative training and uses the model obtained from the current training as the final prediction model.

[0170] If not, the training platform adjusts the model parameters of the prediction model obtained by the current training, and uses the prediction model after adjusting the parameters as the prediction model of the current training to continue iterating the above steps 205-206 until the total loss function meets the training completion conditions of the pre-equipment, and uses the prediction model obtained by the current training as the final prediction model completed by the training.

[0171] In some implementations, the architecture of the prediction model includes an input layer, a backbone network TCN, a physical constraint branch, a residual fusion mechanism, a loss function, and an output layer.

[0172] In some implementations, the input layer is a high-dimensional time series matrix obtained by aligning and fusion of structured time series matrices according to time windows;

[0173] The shape of the high-dimensional time series matrix: X∈R T×D , where T is the time step, such as 300 steps, corresponding to a 5-minute window, and D is the total feature dimension.

[0174] For example: micro motion 10 dimensions + electromagnetic 8 dimensions + natural potential 5 dimensions → D = 23), and Z-Score normalization needs to be performed independently for each feature dimension: where μ train , σ train Calculated based on the data of the corresponding group in the training set.

[0175] The output layer includes predicted values, real-time risk scores, and early warning signals;

[0176] It should be noted that the predicted value is a real-time prediction of road subsidence;

[0177] The real-time risk score is obtained by calculation. The specific calculation method of the real-time risk score is: in, Where p t For real-time risk scoring, is the predicted settlement at the current time step, is the predicted settlement at the i-th time step in the past 30 days;

[0178] The above-mentioned warning signal is a warning level divided by the real-time score and dynamic threshold according to the warning rules. Specifically, the dynamic threshold and warning rules are as follows:

[0179] a. Dynamic threshold: p threshold =μ+2σ, where μ is the mean value of the corresponding group data and σ is the variance of the corresponding group data.

[0180] b. Environmental correction factor:

[0181] When the rainfall is >50mm for 3 consecutive days: p threshold ←p threshold ×0.7.

[0182] When the daily rise of groundwater level is greater than 10cm: threshold ←p threshold ×0.8.

[0183] c. Multi-level warning rules:

[0184] Level 1 (monitoring level): p t ≥0.4 or autoencoder abnormality persists for >2 hours.

[0185] Level 2 (warning level): p t ≥0.7 or Level 1 lasting for >12 hours.

[0186] Level 3 (Emergency): p t ≥0.9 or the micro-signal RMS week-on-week growth rate is >50%.

[0187] In some implementations, the backbone network TCN consists of 4 layers of dilated causal convolutions, residual connections, and skip connections;

[0188] The calculation formula of dilated causal convolution: H l =ReLU(W l * d H l-1 +b l ); where H l is the output of the lth layer, W l is the convolution kernel weight matrix of the i-th layer, * d represents the convolution operation with a dilation factor of d, b l is the bias term of the lth layer;

[0189] The calculation formula of residual connection is: in, is the final output of the lth layer;

[0190] The calculation formula of skip connection is: Among them, H skip The final output of the backbone network.

[0191] In some implementations, the physical constraint branch is embedded in a geomechanical equation;

[0192] The embedded geomechanical equation is: Where τ is the one-dimensional consolidation equation fitted by the historical settlement curve: E is the elastic modulus, E = ρ·v 2 , ρ is soil density, v is electromagnetic wave velocity; ΔP is pore pressure change, g is the acceleration due to gravity, is the natural potential gradient, k is the permeability coefficient, σ is the electrical conductivity of the soil, h is the thickness of the soil layer, α is the regional experience coefficient, S max is the maximum allowable compression of the soil or the design settlement limit, and t is the duration of the consolidation process.

[0193] The above mainly introduces the scheme of the embodiment of the present application from the perspective of device implementation. It is understandable that each device, for example, an electronic device, in order to implement the above functions, includes at least one of the hardware structure and software modules corresponding to the execution of each function. Those skilled in the art should easily appreciate that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software driven hardware manner depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0194] The embodiment of the present application can divide the functional units of the electronic device according to the above method example. For example, each functional unit can be divided according to each function, or two or more functions can be integrated into one processing unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional units. It should be noted that the division of units in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0195] In the case of an integrated unit, Figure 3 A possible structural diagram of the electronic device involved in the above embodiment is shown. The electronic device 30 includes a processing unit 301 and a communication unit 302 , and may further include a storage unit 303 . Figure 3 The structural schematic diagram shown can be used to illustrate the structure of the electronic device involved in the above embodiments.

[0196] when Figure 3The structural diagram shown is used to illustrate the structure of the electronic device involved in the above embodiment. The processing unit 301 is used to control and manage the actions of the electronic device, the communication unit 502 is used for an electronic device to communicate with other devices, and the storage unit 503 is used to store program codes and data of the electronic device.

[0197] For example, the communication unit 502 is used to obtain the Rayleigh wave time series signals, the electric field and magnetic field data time series signals, and the self-potential data time series signals collected by the micro-motion, magnetotelluric field and self-potential detection system;

[0198] The processing unit 502 is used to extract features and fuse multi-source data on the Rayleigh wave time series signal, the electric field and magnetic field data time series signal, and the spontaneous potential data time series signal to obtain a structured time series matrix;

[0199] The structured time series matrix and the road settlement values ​​are divided into a training set, a validation set, and a test set according to a preset ratio;

[0200] Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters;

[0201] Perform joint training based on all parameters of the pre-trained prediction model;

[0202] After the joint training is completed, the model is fine-tuned based on the newly added Rayleigh wave time series signals, the time series signals of electric and magnetic field data, and the time series signals of natural potential data to obtain the trained prediction model.

[0203] Among them, the processing unit 501 can be a processor or a controller, and the communication unit 502 can be a communication interface, a transceiver, a transceiver, a transceiver circuit, a transceiver device, etc. Among them, the communication interface is a general term and can include one or more interfaces. The storage unit 503 can be a memory. When the electronic device 50 is a chip, the processing unit 501 can be a processor or a controller, and the communication unit 502 can be an input interface and / or output interface, a pin or a circuit, etc. The storage unit 503 can be a storage unit within the chip (for example, a register, a cache, etc.), or it can be a storage unit located outside the chip (for example, a read-only memory (ROM), a random access memory (RAM), etc.).

[0204] Among them, the communication unit can also be called a transceiver unit. The antenna and control circuit with transceiver functions in the electronic device 50 can be regarded as the communication unit 502 of the electronic device 50, and the processor with processing function can be regarded as the processing unit 501 of the electronic device 50. Optionally, the device used to implement the receiving function in the communication unit 502 can be regarded as the communication unit, and the communication unit is used to perform the receiving steps in the embodiment of the present application. The communication unit can be a receiver, a receiver, a receiving circuit, etc. The device used to implement the sending function in the communication unit 502 can be regarded as a sending unit, and the sending unit is used to perform the sending steps in the embodiment of the present application. The sending unit can be a transmitter, a transmitter, a sending circuit, etc.

[0205] Figure 3 If the integrated units are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The storage medium for storing computer software products includes various media that can store program codes, such as USB flash drives, mobile hard drives, read-only memories, random access memories, magnetic disks or optical disks.

[0206] Figure 3 A unit in a can also be called a module, for example, a processing unit can be called a processing module.

[0207] The present application also provides a hardware structure diagram of an electronic device, see Figure 4 The electronic device 60 includes a processor 601 and, optionally, a memory 602 connected to the processor 601 .

[0208] In the first possible implementation, see Figure 4 The electronic device 60 further includes a transceiver 603. The processor 601, the memory 602, and the transceiver 603 are connected via a bus. The transceiver 603 is used to communicate with other devices or a communication network. Optionally, the transceiver 603 may include a transmitter and a receiver. The device used to implement the receiving function in the transceiver 603 can be considered a receiver, and the receiver is used to perform the receiving steps in the embodiments of the present application. The device used to implement the transmitting function in the transceiver 603 can be considered a transmitter, and the transmitter is used to perform the transmitting steps in the embodiments of the present application.

[0209] Based on the first possible implementation, Figure 4 The structural schematic diagram shown can be used to illustrate the structure of the electronic device involved in the above embodiments.

[0210] in, Figure 4 It can also represent a system chip in an electronic device. In this case, the actions performed by the electronic device can be implemented by the system chip. The specific actions performed can be found above and will not be repeated here.

[0211] During implementation, each step of the method provided in this embodiment can be completed by hardware integrated logic circuits in a processor or by software instructions. The steps of the method disclosed in the embodiments of this application can be directly implemented as execution by a hardware processor, or as a combination of hardware and software modules in a processor.

[0212] The processor in this application may include but is not limited to at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a microcontroller unit (MCU), or an artificial intelligence processor, etc., various types of computing devices that run software, each of which may include one or more cores for executing software instructions to perform operations or processing. The processor may be a separate semiconductor chip, or it may be integrated into a semiconductor chip together with other circuits. For example, it may form an SoC (system on chip) with other circuits (such as a codec circuit, a hardware acceleration circuit, or various bus and interface circuits), or it may be integrated into the ASIC as a built-in processor of an ASIC. The ASIC with the integrated processor may be packaged separately or with other circuits. In addition to including a core for executing software instructions to perform operations or processing, the processor may further include necessary hardware accelerators, such as a field programmable gate array (FPGA), a PLD (programmable logic device), or a logic circuit that implements dedicated logic operations.

[0213] The memory in the embodiments of the present application may include at least one of the following types: read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or electrically erasable programmable read-only memory (EEPROM). In some scenarios, the memory may also be a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to this.

[0214] An embodiment of the present application also provides a computer-readable storage medium, comprising instructions, which, when executed on a computer, enables the computer to execute any of the above methods.

[0215] An embodiment of the present application also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute any of the above methods.

[0216] An embodiment of the present application also provides a chip, which includes a processor and an interface circuit, the interface circuit is coupled to the processor, the processor is used to run a computer program or instruction to implement the above method, and the interface circuit is used to communicate with other modules outside the chip.

[0217] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using a software program, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more media that can be integrated. The available media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVDs), or semiconductor media (eg, solid state disks (SSDs)).

[0218] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0219] Although the present application has been described with reference to specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely illustrative of the present application as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art may make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, the present application is intended to include such modifications and variations as fall within the scope of the claims of the present application and their equivalents.

Claims

1. A road settlement real-time monitoring and prediction model training method based on multi-physics fields, characterized in that: include: Obtain Rayleigh wave time series signals, electric and magnetic field data time series signals, and natural potential data time series signals; Perform feature extraction and multi-source data fusion on the time series signals of Rayleigh waves, electric and magnetic field data, and natural potential data, construct weighted feature vectors, and divide them into time windows to form a structured time series matrix. The structured time series matrix and the road settlement values ​​are divided into a training set, a validation set, and a test set according to a preset ratio; Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters; Perform joint training based on all parameters of the pre-trained prediction model; After the joint training is completed, the model fine-tuning is triggered based on the newly added time series signal to obtain the trained prediction model.

2. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to claim 1, characterized in that: The loss function of the N-round pre-training process is: Among them, except for the loss mixing coefficient of the physical parameter mapping layer which is non-zero, S phys is the collected physical value; y true is the pre-trained prediction value, and MSE is the mean square error.

3. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to claim 1, characterized in that: The loss function of the joint training based on all parameters is: in, is the predicted settlement sequence, Y true is the true settlement sequence, R is the residual vector, ‖‖2 is the L2 norm, which is used to quantify the amplitude of the residual or vector. is the first-order difference of the predicted value, is the first-order difference of historical data.

4. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to claim 1, characterized in that: In the process of fine-tuning the model triggered by the newly added Rayleigh wave time series signal, the time series signal of electric and magnetic field data, and the time series signal of natural potential data, elastic weight integration (EWC) is used to prevent catastrophic forgetting; wherein, in, is the total loss function, F i is the diagonal element of the ith parameter in the Fisher information matrix, θ i is the current value of the i-th parameter of the model, θ i,old is the historical value of the i-th parameter of the model.

5. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to claim 1, characterized in that: After fine-tuning the triggering model based on the newly added Rayleigh wave time series signal, the time series signal of the electric and magnetic field data, and the time series signal of the natural potential data, the method further includes: lightweight design of the prediction model; The lightweight design method is to add a channel compression module to the last layer of the backbone network: H compressed =ConvlD 1×1 (H TCN , channels = 16); the backbone network uses the time convolution network TCN; Conv1D 1×1 To apply 1×1 convolution to the output of layer 1, H TCN is the output of the temporal convolutional network, channels=16 means the number of output channels is set to 16.

6. A road subsidence real-time monitoring and prediction model training method based on multi-physics fields according to any one of claims 1 to 5, characterized in that: The architecture of the prediction model includes an input layer, a backbone network TCN, a physical constraint branch, a residual fusion mechanism, a loss function and an output layer; The output content of the output layer includes predicted values, real-time risk scores and early warning signals; Method for obtaining the real-time risk score: in, Where p t For real-time risk scoring, is the predicted settlement at the current time step, is the predicted settlement amount at the i-th time step in the past 30 days. The warning signal is a warning level divided according to the real-time score.

7. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to any one of claim 6, characterized in that: The input layer is a high-dimensional time series matrix after the structured time series matrix is ​​aligned and fused according to the time window.

8. The method for training a road subsidence real-time monitoring and prediction model based on multi-physics fields according to any one of claim 6, characterized in that: The backbone network TCN consists of 4 layers of dilated causal convolution, residual connection and skip connection; The calculation formula of the dilated causal convolution is: H l =ReLU(W l * d H l-1 +b l ); where H l is the output of the lth layer, W l is the convolution kernel weight matrix of the i-th layer, * d represents the convolution operation with a dilation factor of d, b l is the bias term of the lth layer; The calculation formula of the residual connection is: in, is the final output of the lth layer; The calculation formula of the skip connection is: Among them, H skip The final output of the backbone network.

9. The method for training a road subsidence real-time monitoring prediction model based on multi-physics fields according to any one of claim 6, characterized in that: The physical constraint branch is embedded in the geomechanical equation; The embedded geomechanical equation is: Where τ is the one-dimensional consolidation equation fitted by the historical settlement curve: E is the elastic modulus, E = ρ·v 2 , ρ is soil density, v is electromagnetic wave velocity; ΔP is pore pressure change, g is the acceleration due to gravity, is the natural potential gradient, k is the permeability coefficient, σ is the electrical conductivity of the soil, h is the thickness of the soil layer, α is the regional experience coefficient, S max is the maximum allowable compression of the soil or the design settlement limit, and t is the duration of the consolidation process.

10. An electronic device, characterized in that: include: a communication unit and a processing unit; The communication unit is used to obtain the Rayleigh wave time series signals, the electric field and magnetic field data time series signals, and the self-potential data time series signals collected by the micro-motion, magnetotelluric field and self-potential detection system; The processing unit is used to extract features and fuse multi-source data on the Rayleigh wave time series signal, the time series signal of the electric field and magnetic field data, and the time series signal of the spontaneous potential data to obtain a structured time series matrix; The structured time series matrix and the road settlement values ​​are divided into a training set, a validation set, and a test set according to a preset ratio; Build an initial prediction model and perform N rounds of pre-training based on the physical branch of the road surface physical parameters; Perform joint training based on all parameters of the pre-trained prediction model; After the joint training is completed, the model is fine-tuned based on the newly added Rayleigh wave time series signals, the time series signals of electric and magnetic field data, and the time series signals of natural potential data to obtain the trained prediction model.

Images