Petroleum drilling geomechanical property estimation method based on deep learning

Abstract

The invention provides a petroleum drilling geomechanical characteristic estimation method based on deep learning, and aims to solve the problems of high measurement cost and poor real-time performance of shear wave velocity Vs in traditional geomechanical analysis by fusing logging while drilling LWD data and real-time drilling engineering parameters. According to the method, a Transformer model is adopted to carry out depth correction on logging-while-drilling data, depth offset of a drill bit and a logging sensor is eliminated, and real-time drilling parameters such as torque T, bit pressure WOB and drilling speed ROP are combined and input into a multi-layer perceptron MLP network to predict the shear wave speed Vs. And further calculating key geomechanical parameters based on the predicted shear wave velocity Vs. In practical application, the method improves the shear wave velocity prediction accuracy to 97.2%, the mean absolute error MAE of the shear modulus is reduced from 0.186 to 0.059, and the bulk modulus is reduced from 0.189 to 0.040. The method can be used for outputting stratum elastic parameters, optimizing bit pressure, pre-warning well wall instability and adjusting a well track, well drilling safety and efficiency are improved, and meanwhile dependence on an expensive well logging technology is reduced.

Description

Technical Field

[0001] This invention belongs to the field of petroleum engineering technology, specifically relating to a method for estimating the geomechanical characteristics of petroleum drilling based on deep learning. Background Technology

[0002] In oil drilling engineering, obtaining accurate formation geomechanical parameters in real time (such as shear wave velocity Vs, shear modulus G, bulk modulus K, etc.) is crucial for ensuring wellbore stability, optimizing drilling pressure parameters, and preventing drilling accidents. Traditional methods mainly rely on logging-while-drilling or wireline logging to obtain longitudinal wave velocity (Vs). p The shear wave velocity (V) is calculated from the RHOB (Relative Tolerancing Surface) and density data, and then extrapolated using empirical formulas or laboratory core testing. s Other elastic parameters. However, measuring shear wave velocity requires specialized logging equipment (such as dipole acoustic logging tools), which is costly and cannot be implemented in all wells. Furthermore, logging-while-drilling data suffers from depth lag, making it difficult to provide timely support for real-time drilling decisions.

[0003] Rapid estimation of subsurface mechanical properties is the cornerstone of constructing geomechanical models of oil and gas reservoirs. These properties include pore pressure, vertical stress, and minimum stress magnitudes, as well as a set of rock elastic parameters such as shear wave velocity, p-wave velocity, and density (RHOB). These parameters collectively form the basis for assessing wellbore stability, fracture propagation, and other geomechanical behaviors crucial for safe and efficient drilling operations. Although V p Density is relatively more common and easier to measure, but V requires full-waveform sonic logging. s This method is particularly scarce. Due to its high cost and operational limitations, it is not frequently used in every well. Obtaining these elastic parameters, especially V... s The challenges highlighted by these methods underscore the limitations of traditional approaches such as logging while drilling, wireline logging, core sampling, and laboratory analysis. These methods not only require substantial financial investment but also introduce significant time delays into drilling operations. The use of machine learning (ML) algorithms has opened new opportunities for improving real-time data acquisition and prediction of rock properties, increasing speed and reducing costs. For example, Olkhovikov et al., in their paper "Geomechanical rock properties from surface drilling telemetry," demonstrated the use of surface drilling telemetry and logging data to feed ML algorithms, bridging the gap between real-time data and geomechanical modeling. This innovative approach emphasizes the importance of adapting machine learning models to the dynamic characteristics of the drilling environment, demonstrating the growing potential of combining spatial and temporal features to improve subsurface characteristics.

[0004] The main problems with existing technologies:

[0005] Current methods for predicting subsurface elastic parameters have significant limitations. They are costly, lack accuracy, and, in particular, cannot provide the real-time predictions necessary for drilling operations, wellbore stability analysis, and geological steering. Accurate estimation of V is challenging when relying solely on logging or well logging data. s Assessing key parameters for rock elasticity is challenging. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies in real-time prediction of geomechanical parameters, such as high measurement costs, data lag, and insufficient prediction accuracy, this invention aims to provide a real-time geomechanical property estimation method and system based on Transformer models and drilling parameters. This method can efficiently integrate multi-source drilling data to achieve accurate real-time prediction of formation mechanical parameters. It innovatively combines deep learning and knowledge from the petroleum engineering field, deeply exploring the dynamic correlation between drilling parameters and formation characteristics through a self-attention mechanism, thus solving the problem that traditional methods cannot respond to formation changes in real time. Specifically, this invention utilizes Transformer networks to process the depth correction problem of logging-while-drilling data, achieves high-precision prediction of shear wave velocity through a multilayer perceptron, and calculates key parameters such as shear modulus and bulk modulus in real time based on rock physics equations. The technical solution adopted by this invention is as follows:

[0007] 1. A deep learning-based method for estimating the geomechanical properties of oil drilling wells, comprising the following steps:

[0008] 1) Collect real-time drilling parameters and logging-while-drilling data;

[0009] 2) Preprocess the data and divide the dataset;

[0010] 3) Based on the data processing in step 2), the Transformer model is used to perform depth correction on the logging-while-drilling data to solve the depth misalignment problem between the drill bit and the logging-while-drilling sensor.

[0011] 4) Based on the correction data obtained in step 3), the corrected logging-while-drilling data is fused with the real-time drilling parameters;

[0012] 5) Based on the data obtained in step 4), input the data into the multilayer perceptron (MLP) network to predict the shear wave velocity V at the current drill bit position. s ;

[0013] 6) Based on the V predicted in step 5) s Calculate the formation elastic parameters, including shear modulus G, bulk modulus K, Young's modulus E, and Poisson's ratio V.

[0014] Step 1) of collecting real-time drilling parameters and logging-while-drilling data includes:

[0015] Real-time drilling parameters include drilling pressure (WOB), drilling speed (ROP), rotational speed (RPM), hook load (HKL), and torque (T); the logging-while-drilling data includes gamma ray (GR), density (RHOB), longitudinal wave transit time (DTC), density difference (DRHO), photoelectric factor (PEF), and neutron porosity (NPHI).

[0016] Step 2) involves preprocessing the data collected in step 1) and dividing the dataset, specifically including the following:

[0017] 2.1) Outlier repair and missing value handling: Linear interpolation is used to repair data segments containing outliers or NaN values.

[0018] 2.2) A partitioning strategy is adopted, which designates an entire well for training or testing, selects one well for testing, and assigns the rest of the wells for training. This solves the problem of high accuracy of the test set caused by the traditional random partitioning due to differences in geology and operating conditions between wells, and the problem that the performance of the method may decrease when applied to new wells.

[0019] 2.3) Feature correlation screening: Feature correlation is analyzed using the Pearson correlation coefficient matrix to eliminate features correlated with shear wave velocity (V). s Features with an absolute value of linear correlation less than 0.2 (such as mud outlet density MDO) reduce the computational complexity of the model.

[0020] Step 3), based on the data processing in step 2), uses the Transformer model to perform depth correction on the logging-while-drilling data to solve the depth misalignment problem between the drill bit and the logging-while-drilling sensor. Specifically, it includes the following sub-steps:

[0021] 3.1) The Transformer network architecture consists of an input layer, an encoder layer, a decoder layer, and an output layer. It employs an encoder-decoder architecture for sequence feature extraction. The input sequence is passed through the encoder layer multiple times for feature extraction. The decoder layer uses the obtained encoded information to generate the output sequence. Positional encoding preserves the sequence's positional information. The specific positional encoding formula is as follows:

[0022]

[0023] Where pos is the position index, i is the dimension index, and d is the position index. model For the hidden layer dimension.

[0024] 3.2) Dynamic mapping of depth offset to establish drill bit depth (D) bit ) and logging-while-drilling sensor depth (D LWD Dynamic mapping function:

[0025] D bit =D LWD +ΔD

[0026] Where D bit D represents the drill bit depth. LWD The depth is represented by the logging-while-drilling sensor, and ΔD is the depth offset. It is dynamically adjusted by constructing a time series model using historical logging-while-drilling data and real-time drilling parameters.

[0027] 3.3) Perform time-depth conversion on the logging-while-drilling data to achieve real-time alignment with the drill bit position.

[0028] Step 4) involves fusing the corrected logging-while-drilling data with real-time drilling parameters based on the correction data obtained in step 3). This includes:

[0029] 4.1) The corrected logging-while-drilling data gamma, density, and P-wave transit time are combined with real-time drilling parameters torque, drilling pressure, drilling speed, and rotational speed by feature stitching.

[0030] 4.2) Standardize the spliced ​​features using the following formula:

[0031]

[0032] Where x is the original feature value, μ is the mean of the feature, and σ is the standard deviation.

[0033] Step 5) involves inputting the data obtained in step 4) into a multilayer perceptron (MLP) network to predict the shear wave velocity V at the current drill bit position. s The specific content includes:

[0034] 5.1) MLP Network Construction:

[0035] 5.1.1) Input layer configuration: The standardized and combined depth-corrected time characteristics and drilling operation data are used as inputs;

[0036] 5.1.2) Hidden layer configuration: Hidden layer 1: 1024 neurons, with a Dropout rate of 20%; Hidden layer 2: 512 neurons, with L2 regularization λ = 0.001; Hidden layer 3: 256 neurons, followed by a batch normalization layer; Hidden layer 4: 128 neurons; Hidden layer 5: 32 neurons. All hidden layers use the ReLU activation function.

[0037] 5.1.3) Output layer configuration: Output layer: 1 neuron, using a linear activation function;

[0038] 5.1.4) Loss function configuration: Use Mean Absolute Error (MAE).

[0039] 5.2) Predicting Vs The processed data is input into the trained MLP network, and then passed through each hidden layer for nonlinear transformation to obtain the predicted V at the output layer. s .

[0040] Step 6) is based on the V predicted in step 5). s Calculate formation elastic parameters, including shear modulus G, bulk modulus K, Young's modulus E, and Poisson's ratio V, based on the predicted shear wave velocity (V). s ), longitudinal wave velocity (V) p Given the rock density (ρ), the elastic parameters of the formation are calculated using the following formula:

[0041] Shear modulus (G):

[0042] G=ρ×Vs 2

[0043] The unit GPa reflects the rock's ability to resist shear deformation.

[0044] Poisson's ratio (V):

[0045]

[0046] Dimensionless, representing the ratio of transverse strain to longitudinal strain in rocks.

[0047] Bulk modulus (K):

[0048]

[0049] The unit GPa measures the ability of a rock to resist volumetric compression.

[0050] Young's modulus (E):

[0051]

[0052] The unit is GPa, which describes the elastic properties of rocks under tension or compression.

[0053] The advantages of this invention compared to the prior art are mainly reflected in:

[0054] (1) Realize the real-time prediction and feedback of formation elastic parameters.

[0055] Traditional methods typically rely on the time difference between P-waves and shear waves measured by logging-while-drilling (LOD) instruments to infer formation parameters through static models or experiments. This approach suffers from significant delays and physical location biases, failing to meet real-time decision-making requirements. This invention utilizes a deep learning model to fuse drilling parameters and LOD data, achieving the inverse relationship between shear wave velocity (Vt) and shear wave velocity (Vt). s The real-time, high-frequency prediction of elastic parameters significantly improves response speed.

[0056] (2) Improve the accuracy and universality of shear wave velocity Vs prediction.

[0057] This invention introduces a Transformer-based depth correction mechanism to dynamically align logging-while-drilling data, resolving the hysteresis error between the sensor and the drill bit. Simultaneously, a shear wave velocity prediction model is constructed using a multilayer perceptron (MLP), trained and validated on multi-well datasets, demonstrating good generalization ability and accuracy.

[0058] (3) Reduce reliance on high-cost logging methods.

[0059] Traditional shear wave velocity V s The acquisition of data typically relies on expensive multipole acoustic logging or wireline logging methods. However, this invention achieves high-quality prediction results by integrating existing drilling and logging data for intelligent reasoning, without the need for additional hardware or high-cost instruments. This has good economic efficiency and promotional value.

[0060] In summary, this invention, by combining deep learning algorithms with multi-source real-time data, achieves rapid, high-precision, low-latency, and low-cost prediction of formation elastic parameters during drilling, effectively overcoming the lag, dependency, and inefficiency problems of traditional methods. It possesses good engineering practicality, model stability, and field adaptability, making it suitable for oil and gas drilling projects under various well types and formation conditions, and has broad industrial application prospects. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below.

[0062] Appendix Figure 1 This is an overall schematic diagram of the present invention;

[0063] Appendix Figure 2 This is a diagram of the Transformer model framework used in this invention;

[0064] Appendix Figure 3 This is the MLP model framework used in this invention;

[0065] Appendix Figure 4 This is a curve comparing the model prediction results; Detailed Implementation

[0066] Example 1

[0067] This embodiment demonstrates a deep learning-based method for estimating the geomechanical properties of oil wells. It has been validated at multiple well locations in the Volve oilfield and can achieve the estimation of shear wave velocity (V0) from raw data processing. s From prediction to the calculation and visualization of elasticity parameters. See also Figure 1 This is an overall schematic diagram of the present invention, which illustrates the following steps:

[0068] 1) Collect real-time drilling parameters and logging-while-drilling data;

[0069] 2) Preprocess the data from step 1) and divide the dataset;

[0070] 3) Based on step 2), use the Transformer model to perform depth correction on the logging-while-drilling data to solve the depth misalignment problem between the drill bit and the logging-while-drilling sensor;

[0071] 4) Based on step 3), fuse the corrected logging-while-drilling data with real-time drilling parameters;

[0072] 5) Based on the input from step 4), the multilayer perceptron (MLP) network predicts the shear wave velocity (V) at the current drill bit position. s );

[0073] 6) Based on step 5), calculate the formation elastic parameters based on the predicted Vs, including shear modulus (G), bulk modulus (K), Young's modulus (E), and Poisson's ratio (V).

[0074] In this embodiment, step 1) involves acquiring real-time drilling parameters and logging-while-drilling data. First, raw drilling parameters and logging-while-drilling data are obtained from wells F1-A, F1-B, F11-A, and F11-T2 in the Volve oilfield. Table 1 shows the information for each well in the dataset. This dataset includes data from four wells: F1-A, F1-B, F11-T2, and F11-A. The data depth range provided by well F1-A is 2620 to 3641 meters, F1B is 2999 to 3425 meters, F11-T2 is 2622 to 4513 meters, and F11-A is 2586 to 3623 meters. The drilling parameters in the dataset, as shown in Table 2, include hook load (HKL), mud inflow density (MDI), mud outflow density (MDO), mud inflow rate (MFI), mud outflow rate (MFO), drilling speed (ROP), rotational speed (RPM), and bit weight (WOB). Table 3 shows the logging-while-drilling (WSD) data, which includes parameters such as gamma ray velocity (GR), porosity (NPHI), density (RHOB), shear wave velocity (DTS), and compression wave velocity (DTC). These parameters are combined with lithology at different depths. The original formats of these data are LAS, LIS, and DLIS, which need to be read and converted into structured data formats using tools such as lasio or PetroPy for standardized processing.

[0075] Table 1 shows the well information in the dataset.

[0076]

[0077] Table 2. Detailed information on drilling parameter data.

[0078]

[0079]

[0080] Table 3. Detailed information on logging-while-drilling parameters.

[0081]

[0082] In this embodiment, step 2) involves preprocessing the well data collected in step 1) to divide the dataset, specifically including the following:

[0083] 2.1) To ensure the quality of model training and prediction accuracy, the raw data is cleaned and normalized. First, drilling parameters and logging parameters are aligned and merged according to sampling points based on the depth field (DEPT).

[0084] 2.2) Next, delete data segments with more than 10% missing values, and fill in the few missing points using linear interpolation. Use the Z-score method to remove outliers and replace the outlier data with the median of each variable.

[0085] 2.3) Calculate the Pearson correlation coefficient and remove variables that correlate with the target variable shear wave velocity (V). s The characteristic variable with weak correlation was excluded because MDO (mud outflow density) lacked linear correlation with other variables and had significant lag, as it reflects the mud density after circulation rather than the real-time conditions of the drill bit.

[0086] 2.4) To unify the scale and improve model stability, min-max normalization is applied to all input variables to compress them to the range [0, 1]. The formula for min-max normalization is:

[0087]

[0088] Among them, X normalized The dataset consists of normalized data; X represents the original data; min(X) is the minimum value of the original data; and max(X) is the maximum value of the original data. The final dataset employs a partitioning strategy, assigning entire wells for either training or testing. Wells F1-B, F11-A, and F11-T2 are allocated for training, while well F1-A is used for testing. This approach ensures the model is tested on completely unseen data from different wells, resulting in a more robust evaluation of its generalization ability under real-world conditions. Using entire wells more accurately simulates field conditions than using random partitions of the entire well data.

[0089] In this embodiment, step 3) uses the Transformer model to perform depth correction on the logging-while-drilling data based on step 2), thus resolving the depth misalignment issue between the drill bit and the logging-while-drilling sensor. (See [link to previous section]). Figure 2 The Transformer model framework used in this invention specifically includes the following:

[0090] 3.1) Data from the 32 depth points preceding each prediction point are used as sequential input. Position encoding is introduced to preserve sequential features, employing a two-layer Transformer encoder. Each layer contains an 8-head multi-attention structure with a hidden dimension of 256. The model output is the predicted logging values ​​for GR, RHOB, and DTC at the current drill bit location. During training, mean squared error (MSE) is used as the loss function, and the Adam optimizer is employed with a learning rate of 0.0001. The corrected data more closely reflects the actual geological response at the drill bit location, significantly improving the accuracy of subsequent predictions.

[0091] 3.2) To address the asynchronous formation response caused by the physical lag of the logging-while-drilling sensor relative to the drill bit, a time-depth mapping correction method using a Transformer neural network is employed. Leveraging the superior time-series modeling capabilities of the Transformer model, the temporal dependencies between drilling parameters and logging-while-drilling data are analyzed in depth, thereby enabling dynamic prediction of formation parameters at the current drill bit position and relocation of the logging response.

[0092] Specifically, by training a time-series prediction model, using historical logging-while-drilling curves (such as RHOB, GR, DTC, etc.) along the depth direction as the input sequence, the model outputs the logging response value corresponding to the current drill bit position, achieving predictive "pre-interpolation". Based on this, a mapping relationship is established between the depth measured by the logging-while-drilling sensor and the actual depth of the drill bit, where D... bit D represents the drill bit depth. LWD For the logging-while-drilling sensor depth, the following offset mapping function applies between the two:

[0093] D bit =D LWD +ΔD

[0094] Wherein, ΔD is a depth offset constant determined based on the drill string structure and sensor position. By constructing the above mapping relationship and combining it with the Transformer network output, time-depth synchronous correction of logging-while-drilling data is achieved, enabling the formation physical parameters measured by logging-while-drilling to accurately reflect the geological environment at the current drill bit location.

[0095] In this embodiment, step 4) involves fusing the corrected logging-while-drilling data with real-time drilling parameters based on step 3), specifically including the following:

[0096] 4.1) Perform feature stitching on the corrected logging-while-drilling data GR, RHOB, DTC and real-time drilling parameters T, WOB, ROP, RPM;

[0097] 4.2) Standardize the spliced ​​features using the following formula:

[0098]

[0099] Where x represents the original feature value, μ is the mean of the feature in the training set, and σ is the standard deviation. Standardization can effectively improve the synergy of features with different dimensions during model training, avoid gradient skew, and enhance the stability and prediction accuracy of model training.

[0100] In this embodiment, step 5) involves inputting the data obtained in step 4) into a multilayer perceptron (MLP) network based on step 4) to predict the shear wave velocity (V) at the current drill bit position. s See also Figure 3 The MLP model framework used in this invention specifically includes the following:

[0101] 5.1) The depth-corrected logging data and drilling parameters are fused to form the model input, with a total of 12 variables in the input dimensions. A multilayer perceptron (MLP) model is constructed for V... s The predicted model structure includes: Hidden layer 1: 1024 neurons, with a 20% Dropout rate; Hidden layer 2: 512 neurons, with L2 regularization λ = 0.001; Hidden layer 3: 256 neurons followed by a batch normalization layer; Hidden layer 4: 128 neurons; Hidden layer 5: 32 neurons; Output layer: 1 neuron using a linear activation function. All hidden layers use the ReLU activation function, calculated as follows:

[0102] ReLU(x) = max(0, x)

[0103] The model was trained using the MSE loss function, with the Adam optimizer, a learning rate of 0.00001, a batch size of 32, and a maximum iteration count of 100 epochs. A validation set was used for early stopping monitoring to prevent overfitting. After training, the model was able to predict the shear wave velocity V at various well depths with high accuracy. s See Figure 4 The model prediction results are compared using curves. To illustrate the difference in prediction results, the Transformer model without logging-while-drilling data and the LSTM model with logging-while-drilling data are selected as a comparison. It can be seen that the Transformer model using logging-while-drilling data performs the best. Finally, the prediction results will be used in subsequent steps to calculate formation elastic parameters.

[0104] In this embodiment, step 6) calculates the formation elastic parameters based on the Vs predicted in step 6), including shear modulus (G), bulk modulus (K), Young's modulus (E), and Poisson's ratio (V), specifically including the following:

[0105] 6.1) Using the V predicted in step 5 s Combined with the V after Transformer correction p (Calculated from DTC) and density ρ (obtained from RHOB), a series of formation elastic parameters are calculated. These parameters include:

[0106] shear modulus

[0107] G=ρ×Vs 2

[0108] Poisson's ratio

[0109]

[0110] bulk modulus

[0111]

[0112] and Young's modulus

[0113]

[0114] These calculations are based on classical rock physics models and can quantitatively assess formation mechanical properties, which is of great significance for wellbore stability analysis and fracturing scheme design.

Claims

1. A deep learning-based method for estimating the geomechanical properties of oil drilling wells, comprising the following steps: 1) Collect real-time drilling parameters and logging-while-drilling data; 2) Preprocess the data and divide the dataset; 3) Use the Transformer model to perform depth correction on the logging-while-drilling data to correct the depth misalignment between the drill bit and the logging-while-drilling sensor; 4) Integrate the corrected logging-while-drilling data with real-time drilling parameters; 5) Construct a multilayer perceptron (MLP) network to input fused data of corrected logging-while-drilling data and real-time drilling parameters to predict the shear wave velocity V at the current drill bit position. s ; 6) Based on the predicted shear wave velocity V s Calculate geomechanical properties.

2. The geomechanical property estimation method according to claim 1, characterized in that, The real-time drilling parameters in step 1) include torque T, drilling pressure WOB, drilling speed ROP, rotational speed RPM, and hook load HKL; the logging-while-drilling data include gamma ray GR, density RHOB, longitudinal wave transit time DTC, density difference DRHO, photoelectric factor PEF, and neutron porosity NPHI.

3. The geomechanical property estimation method according to claim 1, characterized in that, Step 2) further includes: 2.1) Linear interpolation is used to process data segments with outliers and NaN values; 2.2) The dataset is divided by well, with one well selected for testing and the remaining wells used for training; 2.3) Perform correlation analysis on the features of the dataset and remove features that lack linear correlation.

4. The geomechanical property estimation method according to claim 1, characterized in that, Step 3) further includes: 3.1) Constructing the prediction model structure: The Transformer network consists of four main components: model input, encoder, decoder, and output. The Transformer network model adopts an encoder-decoder architecture. The input sequence is passed through the encoder layer multiple times for feature extraction. The decoder layer uses the obtained encoding information to generate the output sequence. Position encoding is integrated to preserve the position information of the elements. 3.2) Analyze the time-series relationship between drilling parameters and logging-while-drilling data using the Transformer model, and establish a mapping function between the drill bit position and the logging-while-drilling sensor position. The mapping function is as follows: D bit =D LWD +ΔD Among them, D bit D represents the drill bit depth. LWD The depth is the logging-while-drilling sensor depth, and ΔD is the depth offset. 3.3) Perform time-depth conversion on the logging-while-drilling data to achieve real-time alignment with the drill bit position.

5. The geomechanical property estimation method according to claim 1, characterized in that, The data fusion in step 4) specifically includes: 4.1) The corrected logging-while-drilling data gamma, density, and P-wave transit time are combined with real-time drilling parameters torque, drilling pressure, drilling speed, and rotational speed by feature stitching. 4.2) Standardize the spliced ​​features using the following formula: Where x is the original feature value, μ is the mean of the feature, and σ is the standard deviation.

6. The geomechanical property estimation method according to claim 1, characterized in that, Step 5) further includes the following steps: 5.1) MLP Network Construction: 5.1.1) Input layer configuration: The standardized and combined depth-corrected time characteristics and drilling operation data are used as inputs; 5.1.2) Hidden layer configuration: Hidden layer 1: 1024 neurons, with a Dropout rate of 20%; Hidden layer 2: 512 neurons, with L2 regularization λ = 0.001; Hidden layer 3: 256 neurons, followed by a batch normalization layer; Hidden layer 4: 128 neurons; Hidden layer 5: 32 neurons. All hidden layers use the ReLU activation function. 5.1.3) Output layer configuration: Output layer: 1 neuron, using a linear activation function; 5.1.4) Loss function configuration: Use Mean Absolute Error (MAE); 5.2) Predicting Vs: The processed data is input into the trained MLP network, and nonlinear transformations are performed sequentially through each hidden layer to obtain the predicted Vs at the output layer. s .

7. The geomechanical property estimation method according to claim 1, characterized in that, The formula for calculating the geomechanical characteristics in step 6) is as follows: Shear modulus: Poisson's ratio: Bulk modulus: And Young's modulus: Where ρ is the rock density, V p V is the longitudinal wave velocity. s Let G be the transverse wave velocity, G be the shear modulus, K be the bulk modulus, E be the Young's modulus, and V be the Poisson's ratio.

Images