Near-surface quality factor structure estimation method based on static sounding and micro-motion observation

Abstract

The invention discloses a near-surface quality factor structure estimation method based on static sounding and micro-motion observation. The near-surface quality factor structure estimation method is used for seismic excitation lithology optimization and near-surface attenuation compensation. Firstly, a near-surface field investigation scheme integrating three investigation means of micro-logging, surface wave and static sounding is provided, so that the arrangement density of micro-logging is reduced, and the investigation efficiency is improved. And then, providing a quality factor structure prediction method which takes a small number of micro-logging survey point location data as a training set, takes a long-short-term memory neural network as a learning model, restores the distribution continuity of rock-soil body parameters in a near-surface space through a time sequence learning mechanism of the long-short-term memory neural network, and extrapolates a non-linear relationship to a majority of micro-logging-free point locations. Finally, the method is applied to near-surface survey of a certain work area in northern Jiangsu, and experimental results prove that the method can accurately predict the underground quality factor structure of the single well and can be widely applied to regional quality factor structure survey.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of geophysical research, and relates to near-surface investigation data acquisition and preprocessing technology, artificial neural network training and generalization technology. The invention can serve as an important basis for distinguishing near-surface lithology, dividing surface low-velocity zone, optimizing excitation lithology and compensating attenuation, and support the improvement of seismic exploration data quality. BACKGROUND

[0002] Seismic exploration technology has important applications in near-surface investigation, which can effectively reveal underground structure information. However, the complexity of the near-surface layer and the influence on high-frequency signals make the seismic wave produce energy attenuation and scattering phenomena during the propagation in the near-surface, thereby affecting the resolution and signal-to-noise ratio of seismic data. Identifying the absorption attenuation characteristics of strata to waves is one of the basic tasks of near-surface investigation in seismic exploration. The strata quality factor (the quality factor Q is divided into longitudinal wave quality factor Q p and transverse wave quality factor Q s . In the field, the quality factor is written as Q p , and when describing physical quantities, it is simply referred to as Q value. Because the parameter that actually affects the processing is only Q p , we only calculate Q p as a key parameter to describe the energy absorption characteristics of rock media, and its numerical distribution provides an important basis for lithology optimization and seismic data attenuation compensation. In the near-surface investigation of exploration sites, it is difficult to directly measure the quality factor of the strata, so the micro-logging waveform analysis method (such as the spectral ratio method, the centroid frequency method and the waveform fitting method) is usually used for indirect inference. This kind of quality factor extraction method based on micro-logging data has high resolution, high reliability (without damaging the original state of the soil) and wide applicability (suitable for soft soil, hard rock and other strata). However, in actual field investigation, the implementation difficulty and extraction results of these methods are often limited by various site conditions, such as well distribution restrictions, lack of environmental friendliness, high construction cost, limited detection depth and noise interference. Therefore, without losing the accuracy of the investigation, reducing the number of micro-logging implementation points can help reduce engineering costs and protect the environment. SUMMARY

[0003] The absorption and attenuation of near-surface media to waves is one of the key factors affecting the quality of seismic data. The traditional micro-logging-based quality factor (referred to as Q value, symbol Q P ) estimation method often faces the application limitations of poor generalizability and high cost. In view of this, the present study proposes a near-surface rock-soil quality factor structure investigation method and process based on long short-term memory neural network to serve the seismic excitation lithology optimization and near-surface attenuation compensation.

[0004] TECHNICAL SCHEME

[0005] A method for estimating the structure of near-surface quality factors based on static cone penetration and micro-motion observations includes the following steps:

[0006] S1. Data acquisition: Obtain near-surface survey paired data, including: location and depth, stratigraphic lithology, cone tip resistance, side friction, shear wave velocity, and measured quality factor Q.

[0007] S2, Data Preprocessing;

[0008] S3. Construct a quality factor Q prediction model;

[0009] S4. Collect information on the location and depth of the near-surface to be predicted, lithology, cone tip resistance, side friction, and shear wave velocity, and estimate the near-surface quality factor Q using the quality factor Q prediction model.

[0010] Preferably, in S1, the soil is classified into five categories: silt, sand, clay, silty clay, and clayey silt by means of drilling core sampling and static cone penetration stratification.

[0011] Preferably, in S1, the shear wave velocity is obtained by inversion from passive source surface wave exploration data.

[0012] Preferably, the specific steps for obtaining the shear wave velocity include:

[0013] First, after preprocessing the surface wave record d(x,t), a Fourier transform is performed in the time domain:

[0014]

[0015] In the formula, f is the frequency. A Fourier transform of D(x,f) in the spatial domain is performed as follows:

[0016]

[0017] In the formula, k is the wave number, which satisfies the following relationship with the phase velocity v:

[0018] k = f / v

[0019] The above formula is used to interpolate the signal in the frequency-wavenumber domain to obtain the dispersion energy spectrum of the surface wave, where the frequency is along the horizontal axis and the wavenumber is along the vertical axis; the dispersion curve of the surface wave is extracted from the energy peak of the dispersion spectrum.

[0020] Subsequently, the shear wave velocity was inverted using the damped least squares method.

[0021] Preferably, in S2, the data preprocessing includes: normalization, interpolation and alignment, and data segmentation.

[0022] Preferably, in S3, a quality factor prediction model is established based on an LSTM neural network, with location and depth, formation lithology, cone tip drag, side friction drag, and shear wave velocity as inputs, and quality factor Q as output.

[0023] Preferably, the mean squared error (MSE) is used as the loss function of the model, and the loss function is defined as:

[0024]

[0025] In the formula, Q pred Q is the Q value predicted by the model. ture is the true Q value, and n is the number of samples.

[0026] Preferably, the LSTM parameters are updated using the backpropagation algorithm, and the update formula is:

[0027]

[0028] In the formula, θ t Let m be the value of the parameter at time t, η be the learning rate, and m be the value of the parameter at time t. t and v t These are the deviation correction estimates for the first and second momentum of the gradient, respectively. ε is a very small value used to prevent division by zero errors, and λ is the weight decay coefficient.

[0029] Preferably, the method further includes step S5: visualization of the results of quality factor structure inversion.

[0030] Preferably, the results are visualized through longitudinal profiles, two-dimensional plane contour maps, or three-dimensional quality factor distribution maps.

[0031] Beneficial effects of the present invention

[0032] This invention improves the traditional micrologging survey network into a multi-source data survey network, implementing micrologging, surface wave, and static cone penetration testing methods only at a few calibration points, thereby reducing survey costs and improving efficiency. Based on an LSTM model, a quality factor prediction model is trained and generalized to most common survey points, ultimately achieving visualized prediction of the quality factor structure, providing reference data for lithology optimization and attenuation compensation processing. Field measurements in the Subei exploration area demonstrate that the LSTM network model can utilize its own time-series learning mechanism to restore the continuity of the near-surface spatial distribution of soil and rock masses, and effectively construct the nonlinear mapping relationship between micrologging, static cone penetration testing, and passive source surface wave survey data. Therefore, the new invention achieves high single-well prediction accuracy for quality factors, and the multi-well extraction results can provide a time-varying quality factor model for attenuation compensation. Attached Figure Description

[0033] Figure 1This is the flowchart of the near-surface P-wave velocity conversion algorithm based on data fusion technology.

[0034] Figure 2 This is a schematic diagram of the layout of survey lines for near-surface investigation.

[0035] Figure 3 These are two near-surface survey lines laid out in the work area in the example.

[0036] Figure 4 This is a schematic diagram of the artificial neural network model structure used.

[0037] Figure 5 This is a schematic diagram of quality factor structure prediction based on training on the same test line.

[0038] Figure 6 This is a graph showing the convergence of the loss function.

[0039] Figure 7 This is a comparison chart of the predicted and measured values ​​of the quality factor distribution of well_1 and well_5 according to the present invention.

[0040] Figure 8 This is a profile of the Q-value prediction results. Detailed Implementation

[0041] The present invention will be further described below with reference to embodiments, but the scope of protection of the present invention is not limited thereto:

[0042] This invention proposes a near-surface field survey scheme integrating three survey methods: micrologging, surface wave logging, and static cone penetration testing, to reduce the density of micrologging and improve survey efficiency. Subsequently, it proposes a method for predicting the quality factor structure by using data from a few micrologging points as a training set and a long short-term memory neural network as a learning model. This method utilizes the network's own temporal learning mechanism to restore the continuity of the spatial distribution of soil and rock parameters near the surface and extrapolates the nonlinear relationships to most locations without micrologging points. Finally, this method was applied to a near-surface survey in a work area in northern Jiangsu Province. Experimental results demonstrate that the proposed method can accurately predict the subsurface quality factor structure of single wells and can be widely applied to regional quality factor structure surveys. Figure 1 It includes the following implementation steps:

[0043] (1) Field collection strategy for multi-source near-surface survey data:

[0044] When establishing a near-surface survey network, survey lines should be divided into ordinary survey lines and calibration survey lines. It is recommended that the ratio of these two types be set between 3:1 and 10:1 to ensure prediction accuracy while controlling costs. For example... Figure 2As shown, only two survey methods, static cone penetration and passive surface wave probing, are used at survey points along ordinary survey lines. However, three survey methods—static cone penetration, passive surface wave probing, and micrologging—are used at survey points along calibration survey lines. The main data types and acquisition methods are as follows:

[0045] Location and depth: This includes the relative position (Offset, m) of the acquisition point within the survey line, and the depth (Depth, m) above the ground. Given continuous sedimentary deposition, geophysical parameter observations from observation points at similar distances and depths should exhibit similarity, thereby constraining the trend and range of the Q-value fitting results.

[0046] Stratigraphy: The soil was classified into five categories, namely silt, sand, clay, silty clay and clayey silt, by means of drilling core and static cone penetration testing.

[0047] Static cone penetration test: A dual-bridge probe is used to record the cone tip resistance (qc) and side friction resistance (fs). Records are taken every 0.1m from the wellhead to the bottom of the well, up to the specified depth, with an estimated 300-400 records.

[0048] Shear wave velocity: The shear wave velocity, representing the propagation velocity of shear waves in the formation, is obtained by inversion from passive source surface wave exploration data. First, the surface wave record d(x,t) is preprocessed, and then a Fourier transform is performed in the time domain.

[0049]

[0050] In the formula, f is the frequency. Perform a Fourier transform on D(x,f) in the spatial domain:

[0051]

[0052] In the formula, k is the wave number, which satisfies the following relationship with the phase velocity v:

[0053] k = f / v (3)

[0054] The dispersion energy spectrum of the surface wave can be obtained by interpolating the signal in the frequency-wavenumber domain using the above formula, where frequency is along the horizontal axis and wavenumber is along the vertical axis. The dispersion curve of the surface wave can be extracted from the energy peak value of the dispersion spectrum. Subsequently, the shear wave velocity is inverted using the damped least squares method. Assuming that the subsurface medium satisfies an n-layered model, the phase velocity c Rj Satisfies the implicit equation:

[0055] F(f j ,c Rj ,v s ,v p ,ρ,h)=0 (4)

[0056] In the formula, f is the frequency, and v s vp Let ρ be the shear wave velocity and h be the longitudinal wave velocity, respectively, ρ be the density, and h be the thickness. Equation (1) can be solved linearly using the following formula:

[0057] JΔx=Δb (5)

[0058] In the formula, vector x represents the shear wave velocity of n strata x=[v s1 ,v s2 ,v s3 ,v s4 ,...,v sn ] T Vector b represents the phase velocity at the m-th frequency point, b = [b s1 ,b s2 ,b s3 ,b s4 ,...,b sn ] T Δb = b - c R (x0) represents the difference between the observed data and the model-predicted data, c R (x0) represents the model response to the initial shear wave velocity, Δx represents the model modification, and J is an m x n Jacobian matrix whose elements are the phase velocities c. R The first-order partial derivative with respect to the transverse wave velocity. Equation (5) can be solved using an optimization algorithm, with the objective function defined as:

[0059]

[0060] In the formula, ||JΔx-Δb||2 is the 2-norm of the vector, and W is the weighting matrix, which can be represented by a diagonal matrix L as W=L T L and α are damping factors that control the convergence direction and inversion speed of Δx. In actual inversion, the solution to equation (5) can be obtained by minimizing the objective function through singular value decomposition of the data.

[0061] Δx=V(Λ 2 +αΙ) -1 ΛU T d, (7)

[0062] In the formula, matrix A = UΛV T d = Lb, where I is a unit diagonal matrix. The initial velocity model can be defined by selecting the phase velocity at a specific frequency.

[0063]

[0064] In the formula, i = 2, 3, ..., n-1, β is a constant between 0.874 and 0.955, and c R (high) and c R(low) represents the phase velocities corresponding to the highest and lowest frequencies, respectively. Global random search, by randomly sampling throughout the solution space, can effectively avoid getting trapped in local optima, thus making it more likely to find a solution close to the global optimum of formula (7). Specifically: First, several initial models are randomly generated in the model space, each model representing the distribution of shear wave velocity with depth in the underground medium; then, for each randomly generated model, its theoretical dispersion curve is calculated and compared with the observed actual dispersion curve, and the error between the two is calculated (usually using an error function, such as mean square error); finally, based on the magnitude of the error function, the random search continues to explore new models. Each search generates a new model in the solution space through some random mechanism (such as the Monte Carlo method) to explore new parameter combinations. When the algorithm gradually converges to a model with a small error, it is considered to be the best inversion result of the shear wave velocity.

[0065] Formation quality factor (Q value): The Q value is estimated by comparing the ratio of the spectral frequencies of seismic waves at different propagation paths or times in micrologging records to calculate the reduction in high-frequency components caused by attenuation. The basic steps are as follows: First, select reference and target bands. Typically, waveforms of the same seismic event at different receivers or waveforms at the same receiver within different time windows are selected for comparison. Assume the reference band is S0(f) and the target band is S1(f), where f is the frequency. Then, calculate the spectral ratio by performing Fourier transforms on the seismic records of the reference and target bands respectively, obtaining their amplitude spectra A0(f) and A1(f) in the frequency domain. The ratio of the amplitude spectra of the two bands can be expressed as:

[0066]

[0067] This ratio reflects the frequency-dependent attenuation characteristics of seismic waves during propagation. Finally, the theoretical model is fitted and solved using linear regression. According to the theoretical model of seismic wave attenuation, the relationship between the spectral ratio R(f) and the frequency f can be expressed as:

[0068]

[0069] Where Δt is the propagation time difference between the reference band and the target band, and Q is the quality factor to be determined.

[0070] (2) Preprocessing of multi-source near-surface survey data:

[0071] Normalization: Due to the large differences in the numerical ranges of different physical quantities, we need to normalize the input data. For example, depth, horizontal distance, static cone tip drag and side friction drag, shear wave velocity and quality factor can be normalized by min-max normalization or Z-score normalization so that they fall within the same numerical range (0 to 1), which is convenient for model processing.

[0072] Interpolation and Alignment: Data sources differ for different survey lines. Borehole velocity data from micro-logging, penetration resistance data from static cone penetration tests, and shear wave velocity from passive source inversion all require formatting. Data density may vary across survey lines, necessitating standardization to the same depth range (0–30 m) and resolution (0.1 m). For missing data points, interpolation methods (such as spline interpolation or linear interpolation) can be used to complete the data. Measurements with the same depth and offset are paired and combined into a single data point, ensuring that the input data is aligned time-series data.

[0073] D i = [lith i , d i , x i , qc i , fs i , vs i Qp i ], (11)

[0074] Where i represents the i-th observation data, lith represents lithology, d represents depth distance, x represents horizontal distance, qc represents cone tip resistance, fs represents side friction resistance, vs represents shear wave velocity, and Qp represents quality factor.

[0075] Data Splitting: During model training, we split the data into two parts: one part for training and the other for validating the model's generalization ability. First, we used most of the data from the same survey line (including shear wave velocity data from micrologging, static cone penetration testing, and passive source method inversion) for training, and the remainder for testing. Subsequently, we used all the data from one survey line for training, and the entire other survey line as test data.

[0076] (3) Establish a quality factor prediction model based on LSTM neural network:

[0077] like Figure 1 Assuming the number of paired data is n, the lithology is lith, the combined depth is d, the offset is x, and the cone tip resistance is q. c Side friction resistance f s Constructing a six-dimensional characteristic space with transverse wave velocity vs. quality factor Q p As the target output, it is denoted as:

[0078]

[0079] LSTM models are defined using deep learning frameworks (such as TensorFlow, Keras, PyTorch, etc.). An LSTM unit typically consists of three parts: an input gate, a forget gate, and an output gate. Each gate has a weight matrix, and an optimization algorithm learns the relationship between the input and output. The architecture of an LSTM network typically includes: an input layer that receives time-series data; hidden layers containing LSTM units, which can be single or multi-layered; and an output layer that predicts the output, which can be the next value of the time series or a classification label. During training, input data is fed into the LSTM network. At each time step, the LSTM determines how much information to retain through its gating mechanism (forget gate, input gate, output gate). Forward propagation step: through the forget gate f... t Determine how much information from the previous time step to retain; this is done by inputting gate i. t and candidate memory unit C t Determines the effect of the current input; through the output gate o t Calculate the current hidden state h t The output is then sent to the next layer.

[0080] LSTM updates memory cell states and output values ​​using the following four basic formulas. Forgetting old memories: via the forget gate f... t The decision is made on how much of the data from the previous moment C is discarded. t-1 Information. Calculate new memory: via input gate i t and candidate memory units Introduce the input information from the current moment. Update the memory state: Based on the results of forgetting and input, update the current memory unit state C. t Calculate the output: through the output gate o t and the current memory state C t The output h at the current moment is determined. t .

[0081] The forget gate determines the state C of the memory cell before the current time t. t-1 How much information needs to be forgotten?

[0082] f t =σ(W f ·[h t-1 ,x t ]+b f (13)

[0083] Among them, f t This is the output of the forget gate, with values ​​ranging from [0,1], where 1 indicates complete retention and 0 indicates complete forgetting. W f and b f These are the weight matrix and bias terms of the forget gate. h t-1It is the hidden state from the previous moment, x t This is the input at the current time step. σ is the sigmoid activation function, used to compress values ​​to the range [0,1].

[0084] The input gate determines the new information at the current moment. How much is added to memory units?

[0085]

[0086] Among them, i t It is the output of the input gate, which controls the amount of current input information, and its value range is [0,1]. By inputting x t And the previous hidden state h t -1 is used to calculate the candidate memory cell states, which range from [-1, 1]. W i and W C These are the weight matrices for the input gate and the candidate memory state, respectively. i and b C These are their bias terms.

[0087] Update the memory cell state; LSTM memory cell C t-1 It updates by forgetting old information and adding new information. The formula is as follows:

[0088]

[0089] Among them, C t It represents the current state of the memory unit. t *C t-1 This indicates that the previous memory unit C is retained. t-1 Partial information. This indicates the new candidate memory. Add to the memory cell.

[0090] The output gate controls the output h of the LSTM at the current time. t This part of the output depends not only on the memory unit C. t Also by output gate o t To adjust:

[0091]

[0092] Among them, o t It is the output of the output gate. h t It represents the hidden state or output of the LSTM at the current time step. W o and b o These are the weight matrix and bias terms of the output gate.

[0093] Mean squared error (MSE) is used as the loss function of the model to measure the difference between the predicted quality factor and the actual quality factor. The loss function is defined as follows:

[0094]

[0095] Among them, Q pred Q is the Q value predicted by the model. ture Let be the true Q-value and n be the number of samples. The parameters of the LSTM (such as weights and biases) are updated using the backpropagation algorithm. Specifically, this is done by calculating the gradient of the loss at each time step using backpropagation time-shaping (BPTT). The AdamW optimizer is chosen as the optimization algorithm, which accelerates convergence and avoids overfitting while maintaining weight decay. The gradient update formula for AdamW is:

[0096]

[0097] Where, θ t Let m be the value of the parameter at time t, η be the learning rate, and m be the value of the parameter at time t. t and v t These are the bias correction estimates for the first and second momentum of the gradient, respectively. ε is a very small value used to prevent division by zero errors, and λ is the weight decay coefficient.

[0098] After defining the model and optimization method, data is fed into the LSTM network in mini-batches for training. As training progresses, the model gradually learns how to extract useful features from the input sequences and optimizes the model parameters through backpropagation. First, the data is randomly split into training, validation, and test sets. Then, the training set is used to train the model, while the validation set is used to adjust hyperparameters (such as learning rate and number of LSTM layers).

[0099] (4) Quality factor prediction and result visualization:

[0100] After training is completed, this invention will use the unknown observation points ( Figure 3 ) or another nearby survey line ( Figure 2 The data from these observation points were used as a test set for generalization prediction. These observation points were not tested with micrologging, therefore there was no direct observational data on quality factors; only location, depth, cone tip resistance, side friction, and shear wave data were available. By inputting this data, the model predicted the quality factor distribution for the survey line. The purpose of this process was to validate the model's predictive ability under unseen geological conditions.

[0101] Visualizing the results of quality factor structure inversion is a crucial step in geophysics and engineering, used to demonstrate the absorption characteristics and wave propagation attenuation of subsurface media. Common visualization methods include depth profiles and 2D or 3D contour maps. These include: Longitudinal profiles: These create a profile using the inverted quality factor results, displaying the quality factor structure at different depths by showing the relationship between depth (vertical axis) and quality factor (horizontal axis); Two-dimensional contour maps: These generate contour maps of quality factors based on their lateral distribution. This method can be used to display the distribution of quality factors across different survey lines or locations; Three-dimensional quality factor distribution maps: These display the distribution of quality factors with depth and horizontal position in three-dimensional space, commonly used for multi-survey line data analysis or the study of three-dimensional stratigraphic structures.

[0102] The visualization results can be analyzed from the following perspectives: areas with higher quality factors usually indicate areas with lower wave propagation loss (low absorption), which may be related to harder strata; low quality factors usually represent high absorption, and these areas may be related to geological layers such as mud and clay; layered structure: the visualization map can clearly show the changes in quality factors at different depths, and combined with other geological information, further explain the underground structure.

[0103] Example:

[0104] The work area is located at the border of Jiangsu and Anhui provinces, with Yangzhou City to the southeast and Tianchang City, Chuzhou City, Anhui Province to the northwest. It is 17.2 km from downtown Yangzhou and 14.1 km from Tianchang City. The terrain is typical hilly, with gentle slopes and a surface layer entirely covered by Quaternary silty clay, approximately 300 m thick, with some exposed rocks. The terrain slopes from southwest to northeast, with elevations ranging from 10 to 50 m. The water table is shallow, generally 2-5 meters.

[0105] To ensure the accuracy of near-surface data and provide reliable low-velocity zone data for field excitation parameters, such as... Figure 3 In this study, two survey lines were laid out along the west and east sides of the work area (survey lines 5199 and 5999), each consisting of 8 survey points spaced 1 km apart. The dual-well micrologging was conducted using an electric spark well-initiated method with in-well and surface-received data. The western survey line (wells 1-8) corresponds to three methods: dual-well micrologging, static cone penetration testing, and surface wave testing. The eastern survey line (wells 9-16) only corresponds to static cone penetration testing and surface wave testing data. Figure 4 The input layer receives six input features: lithology (lith), combined depth (d), offset (x), and cone tip resistance (q). c Side friction resistance f sThe transverse wave velocity vs. is compared with a 6-layer LSTM model, with 256 neurons per layer and a learning rate of 0.0001. The model is preset to 500 iterations (if the loss function remains unchanged for 5 consecutive iterations, an early stopping mechanism is triggered to stop the iteration). 10% of neurons are randomly dropped in each iteration to prevent overfitting. The mean squared error loss function and AdamW optimizer are used, and the output layer is set to a quality factor.

[0106] In Experiment 1, such as Figure 5 As shown, observation points well_1 and well_5 are used as the test set, and the remaining 6 observation points are used as the training set. Figure 6 As shown, the early stopping mechanism was triggered after 135 iterations, and the model achieved an accuracy of approximately 92.0% on the validation set. Figure 7 The image shows the predicted results (blue curve) and measured quality factors (red curve) for well_1 and well_5. It can be seen that the LSTM model achieves good predictions of quality factors at different depths and accurately judges the changing trends of formation quality factors. Leveraging the advantage of LSTM networks in capturing time-series features, the model effectively captures the time dependence in the data. Subsurface parameters (such as soil properties and geological layer properties) often have time or depth correlations, making LSTM particularly suitable for this task.

[0107] In Experiment 2, such as Figure 1 As shown, this is the Q-value prediction module for inputting data at 9 measuring points along the eastern survey line. Figure 8 As shown, a two-dimensional profile of the quality factor values ​​was obtained and plotted. The quality factor structure in this area exhibits a three-layer distribution pattern with high values ​​in the middle and low values ​​at the top and bottom. The dominant activating lithology is located approximately 10 to 25 meters underground, roughly corresponding to the clay layer distribution.

[0108] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for estimating the structure of near-surface quality factors based on static cone penetration and micro-motion observation, characterized in that... It includes the following steps: S1. Data acquisition: Obtain near-surface survey paired data, including: location and depth, stratigraphic lithology, cone tip resistance, side friction, shear wave velocity, and measured quality factor Q. S2, Data Preprocessing; S3. Construct a quality factor Q prediction model; S4. Collect information on the location and depth of the near-surface to be predicted, lithology, cone tip resistance, side friction, and shear wave velocity, and estimate the near-surface quality factor Q using the quality factor Q prediction model.

2. The method according to claim 1, characterized in that... In S1, the soil was classified into five categories: silt, sand, clay, silty clay, and clayey silt by means of drilling core sampling and static cone penetration stratification.

3. The method according to claim 1, characterized in that... In S1, the shear wave velocity is obtained by inversion from passive source surface wave exploration data.

4. The method according to claim 3, characterized in that... The specific steps for obtaining shear wave velocity include: First, after preprocessing the surface wave record d(x,t), a Fourier transform is performed in the time domain: In the formula, f is the frequency. A Fourier transform of D(x,f) in the spatial domain is performed as follows: In the formula, k is the wave number, which satisfies the following relationship with the phase velocity v: k = f / v The above formula is used to interpolate the signal in the frequency-wavenumber domain to obtain the dispersion energy spectrum of the surface wave, where the frequency is along the horizontal axis and the wavenumber is along the vertical axis; the dispersion curve of the surface wave is extracted from the energy peak of the dispersion spectrum. Subsequently, the shear wave velocity was inverted using the damped least squares method.

5. The method according to claim 1, characterized in that... In S2, the data preprocessing includes: normalization, interpolation and alignment, and data segmentation.

6. The method according to claim 1, characterized in that... In S3, a quality factor prediction model is established based on an LSTM neural network, with location and depth, formation lithology, cone tip drag, side friction drag, and shear wave velocity as inputs and quality factor Q as output.

7. The method according to claim 6, characterized in that... The mean squared error (MSE) is used as the loss function for the model, and the loss function is defined as follows: In the formula, Q pred Q is the Q value predicted by the model. ture is the true Q value, and n is the number of samples.

8. The method according to claim 6, characterized in that... The parameters of the LSTM are updated using the backpropagation algorithm, and the update formula is: In the formula, θ t Let m be the value of the parameter at time t, η be the learning rate, and m be the value of the parameter at time t. t and v t These are the deviation correction estimates for the first and second momentum of the gradient, respectively. ε is a very small value used to prevent division by zero errors, and λ is the weight decay coefficient.

9. The method according to claim 1, characterized in that... The method also includes step S5: visualization of the results of quality factor structure inversion.

10. The method according to claim 9, characterized in that... Results are visualized through longitudinal profiles, two-dimensional contour maps, or three-dimensional quality factor distribution maps.

Images