Prediction method, prediction device and equipment for soil stratum distribution and medium

Abstract

The invention provides a prediction method, a prediction device, equipment and a medium for soil stratum distribution, and the method comprises the steps: for each target position point in a to-be-predicted soil region, determining a plurality of adjacent drilling points closest to the target position point from a drilling set; obtaining stratum category information corresponding to each adjacent drilling point, and generating a feature matrix corresponding to the target position point by using the multi-dimensional vector corresponding to each preset stratum category and the stratum category information corresponding to each adjacent drilling point; and inputting the feature matrix into a pre-trained stratigraphic distribution prediction model, determining a stratigraphic distribution prediction result corresponding to the target position point, and determining a target stratigraphic category to which the target position point belongs based on the stratigraphic distribution prediction result. According to the method and the device, the prediction accuracy, the boundary identification capability and the identification capability of minority stratums are improved, the manual intervention is reduced, and the exploration precision and efficiency are improved.

Description

Technical Field

[0001] This application relates to the technical field of formation distribution prediction, and in particular, to a method, device, equipment and medium for predicting soil formation distribution. Background Art

[0002] In the field of geotechnical engineering, accurately understanding the distribution of underground formations is crucial for engineering design and construction. Traditional geological modeling methods, such as spline interpolation, inverse distance weighting, and Kriging interpolation, usually perform modeling on the expected cross-section through two-dimensional formation interpolation. These methods are relatively effective in homogeneous regions, but under complex three-dimensional formation conditions, due to the high requirements for data volume and interpolation functions, their prediction accuracy and applicability are limited, and it is difficult to achieve automated processing. In recent years, with the rise of data-driven methods, the application of machine learning and deep learning in geological modeling has developed rapidly. In addition to traditional interpolation methods and probability models, methods such as random forest and deep neural network (DNN) show higher adaptability and prediction accuracy in dealing with complex formation characteristics. These methods do not rely on prior assumptions, but directly extract knowledge from existing data, and show high robustness in dealing with non-uniform borehole distributions or highly heterogeneous formations.

[0003] However, current formation modeling still faces many challenges. Many existing models rely too much on limited borehole data, resulting in insufficient generalization ability; the formation categories in some areas are scarce, making it difficult for the model to effectively identify a small number of formation categories. Therefore, it is particularly necessary to design a soil formation distribution prediction method with higher refinement and more applicable formation categories. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, device, equipment and medium for predicting soil formation distribution, which can quickly predict the formation distribution of any unknown point within the range only with multiple adjacent borehole points, improve the prediction accuracy, boundary recognition ability and the recognition ability of minority formation categories, improve the automation level of formation modeling, reduce manual intervention, and improve the exploration accuracy and efficiency.

[0005] In the first aspect, an embodiment of this application provides a method for predicting soil formation distribution, and the prediction method includes:

[0006] For each target position point in the soil area to be predicted, determine multiple adjacent borehole points closest to the target position point from the borehole set;

[0007] [[ID=2,5]]Obtain the formation category information corresponding to each adjacent borehole point, and generate a feature matrix corresponding to the target position point by using the multi-dimensional vector corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point;

[0008] Input the feature matrix into a pre-trained formation distribution prediction model to determine the formation distribution prediction result corresponding to the target location point, and determine the target formation category to which the target location point belongs based on the formation distribution prediction result; wherein, the formation prediction distribution result represents the probability that the target location point belongs to each preset formation category, and the training data used in the training process of the formation distribution prediction model is obtained by expanding through the SMOTE oversampling method.

[0009] Further, determining multiple neighboring borehole points closest to the target location point from the borehole set includes:

[0010] For each borehole point to be screened in the borehole set, determine the spatial position coordinates of the borehole point to be screened relative to the target location point, and calculate the Euclidean distance between the borehole point to be screened and the target location point using the spatial position coordinates;

[0011] Sort the multiple borehole points to be screened in ascending order of distance using the Euclidean distance corresponding to each borehole point to be screened, and select a preset number of borehole points to be screened from the sorting as the neighboring borehole points.

[0012] Further, generating the feature matrix corresponding to the target location point using the multi-dimensional vector corresponding to each preset formation category and the formation category information corresponding to each neighboring borehole point includes:

[0013] Perform grid processing on the formation distribution of the soil area to be predicted to form a high-resolution formation distribution matrix;

[0014] For each neighboring borehole point, determine the initial feature code corresponding to the neighboring borehole point from the multi-dimensional vectors corresponding to multiple preset formation categories based on the formation category information corresponding to the neighboring borehole point, and determine the depth information corresponding to the neighboring borehole point from the formation distribution matrix;

[0015] Concatenate the spatial position coordinates of the neighboring borehole point relative to the target location point, the Euclidean distance between the borehole point to be screened and the target location point, the depth information corresponding to the neighboring borehole point, and the initial feature code corresponding to the neighboring borehole point to generate the target feature code corresponding to the neighboring borehole point;

[0016] Generate the feature matrix corresponding to the target location point using the target feature codes corresponding to multiple neighboring borehole points.

[0017] Further, the formation distribution prediction model is trained through the following steps:

[0018] Determine a test set from the original drilling dataset, and use the remaining data in the original drilling dataset as the training set and the validation set;

[0019] Use the five-fold cross-validation method to divide the remaining data in the original drilling dataset to obtain a training set and a validation set;

[0020] Perform SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set;

[0021] Respectively use the target training set and the validation set to train and validate the original formation distribution prediction model until convergence;

[0022] Use the test set to test the converged original formation distribution prediction model, and use the converged original formation distribution prediction model that passes the test as the formation distribution prediction model.

[0023] Further, the performing SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set includes:

[0024] Obtain the minority class samples in the training set, and determine a plurality of neighbor samples corresponding to each minority class sample;

[0025] For each minority class sample, determine a target neighbor sample from the plurality of neighbor samples corresponding to the minority class sample, and perform linear interpolation between the minority class sample and the target neighbor sample to generate augmented samples, so as to obtain the target training set.

[0026] In a second aspect, an embodiment of the present application further provides a prediction device for soil formation distribution, and the prediction device includes:

[0027] A drilling point determination module, configured to determine, for each target position point in the soil area to be predicted, a plurality of adjacent drilling points closest to the target position point from a set of drillings;

[0028] A feature matrix generation module, configured to obtain the formation category information corresponding to each adjacent drilling point, and generate a feature matrix corresponding to the target position point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent drilling point;

[0029] The formation type prediction module is used to input the feature matrix into a pre-trained formation distribution prediction model, determine the formation distribution prediction result corresponding to the target location point, and determine the target formation type to which the target location point belongs based on the formation distribution prediction result; wherein, the formation prediction distribution result represents the probability that the target location point belongs to each preset formation type, and the training data used in the training process of the formation distribution prediction model is obtained by expanding through the SMOTE oversampling method.

[0030] Further, when the borehole point determination module is used to determine multiple adjacent borehole points closest to the target location point from the borehole set, the borehole point determination module is further used for:

[0031] For each borehole point to be screened in the borehole set, determine the spatial position coordinates of the borehole point to be screened relative to the target location point, and calculate the Euclidean distance between the borehole point to be screened and the target location point using the spatial position coordinates;

[0032] Sort the multiple borehole points to be screened in ascending order of distance using the Euclidean distance corresponding to each borehole point to be screened, and select a preset number of borehole points to be screened from the sorting as the adjacent borehole points.

[0033] Further, when the feature matrix generation module is used to generate the feature matrix corresponding to the target location point using the multi-dimensional vector corresponding to each preset formation type and the formation type information corresponding to each adjacent borehole point, the feature matrix generation module is further used for:

[0034] Perform grid processing on the formation distribution of the soil area to be predicted to form a high-resolution formation distribution matrix;

[0035] For each adjacent borehole point, determine the initial feature code corresponding to the adjacent borehole point from the multi-dimensional vectors corresponding to multiple preset formation types based on the formation type information corresponding to the adjacent borehole point, and determine the depth information corresponding to the adjacent borehole point from the formation distribution matrix;

[0036] Concatenate the spatial position coordinates of the adjacent borehole point relative to the target location point, the Euclidean distance between the borehole point to be screened and the target location point, the depth information corresponding to the adjacent borehole point, and the initial feature code corresponding to the adjacent borehole point to generate the target feature code corresponding to the adjacent borehole point;

[0037] Generate the feature matrix corresponding to the target location point using the target feature codes corresponding to multiple adjacent borehole points.

[0038] In a third aspect, an embodiment of the present application further provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are executed by the processor, the steps of the prediction method for soil formation distribution as described above are executed.

[0039] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is run by a processor, the steps of the prediction method for soil formation distribution as described above are executed.

[0040] For a prediction method, a prediction device, a device, and a medium for soil formation distribution provided by an embodiment of the present application, first, for each target position point in a soil area to be predicted, a plurality of neighboring borehole points closest to the target position point are determined from a borehole set; then, formation category information corresponding to each neighboring borehole point is obtained, and a feature matrix corresponding to the target position point is generated by using a multi-dimensional vector corresponding to each preset formation category and the formation category information corresponding to each neighboring borehole point; finally, the feature matrix is input into a pre-trained formation distribution prediction model to determine a formation distribution prediction result corresponding to the target position point, and a target formation category to which the target position point belongs is determined based on the formation distribution prediction result; where the formation prediction distribution result represents the probability that the target position point belongs to each preset formation category, and the training data used in the training process of the formation distribution prediction model is obtained by expanding through the SMOTE oversampling method.

[0041] This application is applicable to sites containing multiple formation types through the feature extraction of adjacent borehole data. By combining multiple technologies in model design and data processing, it realizes the efficient prediction of shallow complex formations. By selecting the data of adjacent borehole points and combining with the spatial neighborhood strategy, multiple borehole points closest to the target point are selected as input features, enabling the model to effectively capture the local geological features of the target location and flexibly adapt to unevenly distributed borehole data. Encoding multiple formation types and mapping them into multi-dimensional vectors respectively can ensure that the model better identifies and differentiates complex multi-class formation types. For the rare formation types in the model training dataset, the Synthetic Minority Oversampling Technique (SMOTE) is used to augment the minority sample data. This application is applicable to complex site conditions containing multiple formation types. Only multiple adjacent borehole points are required to quickly predict the formation distribution of any unknown point within the range, providing technical support for refined geological modeling. It improves its prediction accuracy, boundary recognition ability, and the recognition ability of minority formation types, enhances the automation level of formation modeling, reduces manual intervention, and improves the exploration accuracy and efficiency.

[0042] To make the above objects, features, and advantages of this application more obvious and understandable, the following specific preferred embodiments are given, in conjunction with the accompanying drawings, and are described in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0044] Figure 1 It is a flowchart of a prediction method for soil formation distribution provided by an embodiment of this application;

[0045] Figure 2 It is a schematic diagram of the spatial distribution of borehole points and the neighborhood selection strategy provided by an embodiment of this application;

[0046] Figure 3 It is a schematic diagram of the conversion of a 9-dimensional vector of a formation based on One-hot encoding provided by an embodiment of this application;

[0047] Figure 4 It is a schematic diagram of the data format of a feature matrix and the architecture of a formation distribution prediction model provided by an embodiment of this application;

[0048] Figure 5Schematic diagram of the loss curve and evaluation index curve of a CNN model during the cross-validation process provided by the embodiments of the present application;

[0049] Figure 6 Schematic diagram of the structure of a prediction device for soil stratum distribution provided by the embodiments of the present application;

[0050] Figure 7 Schematic diagram of the structure of an electronic device provided by the embodiments of the present application. Detailed implementation manners

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Generally, the components of the embodiments of the present application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, every other embodiment obtained by those skilled in the art without creative efforts belongs to the scope of protection of the present application.

[0052] First, the applicable application scenarios of the present application are introduced. The present application can be applied to the technical field of stratum distribution prediction.

[0053] In the field of geotechnical engineering, accurately understanding the distribution of underground strata is crucial for engineering design and construction. Traditional geological modeling methods, such as spline interpolation method, inverse distance weighting method, and Kriging interpolation method, usually perform modeling on the expected cross-section through two-dimensional stratum interpolation. These methods are relatively effective in homogeneous regions, but under complex three-dimensional stratum conditions, due to the high requirements for the amount of data and interpolation functions, their prediction accuracy and applicability are limited, and it is difficult to achieve automated processing. In recent years, with the rise of data-driven methods, the application of machine learning and deep learning in geological modeling has developed rapidly. In addition to traditional interpolation methods and probability models, methods such as random forest and deep neural network (DNN) show higher adaptability and prediction accuracy in dealing with complex stratum features. These methods do not rely on prior assumptions, but directly extract knowledge from existing data, and show high robustness in dealing with non-uniform borehole distributions or highly heterogeneous strata.

[0054] Nevertheless, current formation modeling still faces many challenges. Many existing models rely too much on limited borehole data, resulting in insufficient generalization ability; the formation categories in some areas are scarce, making it difficult for the model to effectively identify a small number of formation categories. Therefore, it is particularly necessary to design a soil formation distribution prediction method with higher refinement and applicable to more formation categories.

[0055] Based on this, the embodiments of the present application provide a prediction method for soil formation distribution, which can quickly predict the formation distribution of any unknown point within the range only with multiple adjacent borehole points, improving the prediction accuracy, boundary recognition ability, and the recognition ability for minority formation categories, enhancing the automation level of formation modeling, reducing manual intervention, and improving the exploration accuracy and efficiency.

[0056] Please refer to Figure 1 , Figure 1 which is a flowchart of a prediction method for soil formation distribution provided by the embodiments of the present application. As shown in Figure 1 , the prediction method provided by the embodiments of the present application includes:

[0057] S101, for each target position point in the soil area to be predicted, determine multiple adjacent borehole points closest to the target position point from the borehole set.

[0058] For the above step S101, in specific implementation, for each target position point in the soil area to be predicted, during the feature extraction process of the target position point, a spatial neighborhood strategy is adopted, and the data of multiple adjacent borehole points closest to the target position point in the borehole set are selected as feature inputs. In this way, the present application selects the borehole points closest to the target position point, improving the subsequent model's ability to capture local geological features and adapting to unevenly distributed borehole data, enhancing the adaptability and generalization ability of the model in complex geological environments.

[0059] As an optional embodiment, for the above step S101, the determining multiple adjacent borehole points closest to the target position point from the borehole set includes:

[0060] Step 1011, for each borehole point to be screened in the borehole set, determine the spatial position coordinates of the borehole point to be screened relative to the target position point, and calculate the Euclidean distance between the borehole point to be screened and the target position point using the spatial position coordinates.

[0061] For the above step 1011, in specific implementation, for each borehole point to be screened in the borehole set, determine the spatial position coordinates of the borehole point to be screened relative to the target position point, and calculate the Euclidean distance between the borehole point to be screened and the target position point using the spatial position coordinates. Specifically, set a certain target position point P, and the borehole set It is represented by the following formula:

[0062] ={B1(x1,y1),B2(x2,t2),…,B n (x n ,y n )}

[0063] where (x i , y i ) represents the spatial position coordinates of the l-th drill hole to be screened in the drill hole set relative to the target position point P.

[0064] The Euclidean distance between the drill hole to be screened B i and the target position point P is calculated by the following formula:

[0065]

[0066] where d(P,B i ) represents the Euclidean distance between the i-th drill hole to be screened in the drill hole set and the target position point P, and (x0, y0) represents the position coordinates of the target position point P.

[0067] Step 1012: Sort the multiple drill holes to be screened in ascending order of distance according to the Euclidean distance corresponding to each drill hole to be screened, and select a preset number of drill holes to be screened from the sorting as the adjacent drill holes.

[0068] For the above step 1012, in specific implementation, the multiple drill holes to be screened are sorted in ascending order of distance according to the Euclidean distance corresponding to each drill hole to be screened, and then a preset number of drill holes to be screened are selected from the sorting as adjacent drill holes. Here, according to the example provided in this application, the preset number is 5, that is, the 5 drill holes to be screened closest to the target position point are selected as adjacent drill holes. Specifically, the 5 adjacent drill holes sorted by distance are represented by the following formula:

[0069]

[0070] where k is the numbering item, is the five adjacent drill holes sorted by distance.

[0071] Please refer to Figure 2 , Figure 2 which is a schematic diagram of the spatial distribution of drill holes and the neighborhood selection strategy provided by the embodiment of the present application. As shown in Figure 2As shown, it shows the spatial distribution of boreholes in the soil area to be predicted and the distance distribution characteristics of adjacent borehole points. Among them, the black points represent all target location points, the hollow circles represent the borehole points in the borehole set, and the points connected by the gray lines represent the five nearest neighbor borehole points of the target location point. The inset shows the distance distribution histogram of adjacent borehole points. The distance distribution is mainly concentrated in the range of 20 - 60m, and the farthest distance reaches 89m.

[0072] S102. Obtain the formation category information corresponding to each adjacent borehole point, and generate a feature matrix corresponding to the target location point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point.

[0073] Here, according to the embodiments provided in the present application, there are a total of nine preset formation categories, namely plain fill, silt, silty fine sand, silty soil, silty clay, medium sand, residual soil silty clay, strongly weathered argillaceous siltstone, and moderately weathered argillaceous siltstone. Each preset formation category is respectively mapped to a 9-dimensional vector. Please refer to Figure 3 , Figure 3 which is a schematic diagram of the conversion of the 9-dimensional vector of the formation based on One-hot coding provided by the embodiments of the present application. As Figure 3 shown, each formation corresponds to a set of vector data. One-hot coding is performed according to the formation category, and One-hot coding is used to convert the category information of different formations into a standardized matrix form. The nine formations are respectively mapped to 9-dimensional vectors. Optionally, the nine-layer formation category can be defined differently according to the actual scenario, as long as different formations are defined as unique 9-dimensional vectors. One-hot coding is performed on multiple formation categories, and the category information is mapped to multi-dimensional vectors to ensure the model's ability to distinguish different formation types.

[0074] Regarding the above step S102, in specific implementation, after determining multiple adjacent borehole points in step S101, the formation category information corresponding to each adjacent borehole point, and then generate a feature matrix corresponding to the target location point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point.

[0075] As an optional embodiment, regarding the above step S102, the generating a feature matrix corresponding to the target location point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point includes:

[0076] Step 1021. Perform grid processing on the formation distribution of the soil area to be predicted to form a high-resolution formation distribution matrix.

[0077] For the above-mentioned step 1021, in specific implementation, the formation distribution of the soil area to be predicted is gridded with high-precision resolution to form a high-resolution formation distribution matrix. Here, according to the example provided in this application, a depth interval of 0.1 m is used as the unit to grid the formation distribution of each borehole. The high resolution means that the grid interval of depth is 0.1 m. The specific method is to resample the soil layer information from the surface layer to the maximum drilling depth of the soil area to be predicted into equally spaced depth data points (divided by 0.1 m) to form a high-resolution formation distribution matrix. In this way, the depth distribution of the formation is refined using high-resolution depth data. Such high-resolution data enables the model to more accurately capture the variation characteristics of the formation in the depth direction. High resolution not only improves the model's ability to capture subtle formation changes, enhances the model's recognition ability for thin layer structures and boundary regions, but also makes its prediction in the boundary region more accurate, reducing the misclassification phenomenon of adjacent formations. Especially in the transition region, the model can more finely depict the continuity and gradual change characteristics of the formation, reducing the misclassification phenomenon of adjacent formations. The results show that the developed model has significant advantages in the extraction of most class formation characteristics and spatial distribution learning, and can provide reliable support for the automated prediction of complex formations.

[0078] Step 1022: For each adjacent borehole point, based on the formation category information corresponding to this adjacent borehole point, determine the initial feature code corresponding to this adjacent borehole point from the multi-dimensional vectors corresponding to multiple preset formation categories, and determine the depth information corresponding to this adjacent borehole point from the formation distribution matrix.

[0079] For the above-mentioned step 1022, in specific implementation, for each adjacent borehole point, based on the formation category information corresponding to this adjacent borehole point, determine the initial feature code corresponding to this adjacent borehole point from the multi-dimensional vectors corresponding to multiple preset formation categories. Here, as an example, when the formation category information corresponding to this adjacent borehole point is plain fill, the multi-dimensional vector corresponding to plain fill is the initial feature code corresponding to this adjacent borehole point. Then determine the depth information z corresponding to this adjacent borehole point from the formation distribution matrix i 。

[0080] Step 1023: Concatenate the spatial position coordinates of this adjacent borehole point relative to the target position point, the Euclidean distance between this borehole point to be screened and the target position point, the depth information corresponding to this adjacent borehole point, and the initial feature code corresponding to this adjacent borehole point to generate the target feature code corresponding to this adjacent borehole point.

[0081] For the above-mentioned step 1023, in specific implementation, the spatial position coordinates (x i , y i) The Euclidean distance d(P, B) between the to-be-screened drilling point and the target position point i ) The depth information z corresponding to the adjacent drilling point i and the initial feature encoding corresponding to the adjacent drilling point are concatenated to generate a target feature encoding that concatenates the adjacent drilling points row by row.

[0082] Step 1024: Generate a feature matrix corresponding to the target position point by using the target feature encodings corresponding to multiple adjacent drilling points.

[0083] For the above step 1024, in specific implementation, after the target feature encoding corresponding to each adjacent drilling point is constructed, a feature matrix corresponding to the target position point is generated by using the target feature encodings corresponding to multiple adjacent drilling points. Please refer to Figure 4 , Figure 4 which is a schematic diagram of a feature matrix data format and a formation distribution prediction model architecture provided by an embodiment of the present application. As Figure 4 shown, continuing the above embodiment, when the multi-dimensional vector corresponding to the preset formation category is a 9-dimensional vector, a 5×13 feature matrix can be constructed by selecting data of five adjacent drilling points as the feature input of the formation distribution prediction model. In this way, the present application constructs a multi-dimensional and multi-level formation feature matrix by combining the spatial position of adjacent drill holes, the formation depth, and the one-hot encoding features of the formation category. This feature matrix can capture the spatial correlation of the geological sequence and improve the classification accuracy of the subsequent model.

[0084] S103: Input the feature matrix into a pre-trained formation distribution prediction model, determine the formation distribution prediction result corresponding to the target position point, and determine the target formation category to which the target position point belongs based on the formation distribution prediction result.

[0085] For the above step S103, in specific implementation, the feature matrix is input into a pre-trained formation distribution prediction model to determine the formation distribution prediction result corresponding to the target position point. The formation prediction distribution result represents the probability that the target position point belongs to each preset formation category, and the target formation category to which the target position point belongs is determined based on the formation distribution prediction result, that is, the preset formation category with the highest probability in the formation prediction distribution result is used as the target formation category to which the target position point belongs.

[0086] According to the embodiment provided by the present application, the formation distribution prediction model in the present application is a CNN model structure for predicting shallow geological formations. As Figure 4 shown in the formation distribution prediction model architecture, the basic structure of this CNN model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and the specific design is as follows:

[0087] The input layer receives the preprocessed 5×13 feature matrix as input; the form of the input feature matrix is as follows:

[0088]

[0089] Among them, X is the input feature matrix, and C, H, and W are the sizes of the matrix.

[0090] The formation distribution prediction model provided by this application is designed with a total of 5 convolutional layers. Each layer extracts features through a 3×3 convolutional kernel and uses the Leaky ReLU (ReLU with a negative slope) activation function, which can retain small-range negative value information and help avoid the problem of gradient disappearance. The output of each convolutional layer can be expressed by the following formula:

[0091]

[0092] Among them, Y represents the output features of each convolutional layer. It is the result after convolution operation, bias addition, and activation function. X represents the input feature map of each layer, which is the output of the previous layer or the initial input data. c represents the index of the input channel. k represents the index of the output channel. For example, if the convolutional layer has 16 convolutional kernels, then 16 output channels are correspondingly generated. b represents the bias value, which is used to adjust the base value of the convolution result and enhance the expression ability of the model. f is the activation function (Leaky ReLU activation function), which is used to introduce non-linear mapping so that the model can learn complex features.

[0093] The Leaky ReLU activation function is shown as follows:

[0094]

[0095] The high-dimensional features output by the convolutional layer are further reduced in dimension through the global average pooling (GAP) layer, and the feature values of each channel are compressed into a scalar. Global pooling is used to reduce the number of parameters and retain the most significant global features. The output of the pooling layer can be expressed by the following formula:

[0096]

[0097] Finally, the fully connected layer is designed as two fully connected layers, which map the extracted features to the target output. The first layer compresses the features to 32 dimensions, and the second layer maps to the final 9 formation categories. The category with the largest prediction factor in the final output is taken as the formation prediction result of the model.

[0098] Further, according to the embodiments provided by the present application, the training data used in the training process of the formation distribution prediction model is obtained by augmenting through the SMOTE oversampling method. In the formation prediction task, the unbalanced distribution of formation categories is one of the main challenges in model training. The data samples of certain formations (such as silty fine sand, medium sand, etc.) are relatively scarce, while the proportion of data samples of other formations (such as silt) is too large. This may cause the model to perform well on the majority-class formations during training, but have poor prediction ability for the minority-class formations, thus affecting the overall performance of the model. To solve this problem, the present application introduces the SMOTE technology to balance the data distribution of each formation category by generating adjacent synthetic samples of real formation data in the feature space.

[0099] According to the embodiments provided by the present application, the formation distribution prediction model is trained through the following steps:

[0100] A: Determine a test set from the original borehole dataset, and use the remaining data in the original borehole dataset as the training set and the validation set.

[0101] For the above step A, in specific implementation, the samples in the original borehole dataset are divided to determine the test set, and the remaining data in the original borehole dataset is used as the training set and the validation set. Here, during the data division process, to ensure the generalization ability of the model and the fairness of training, for the borehole data within the total data range, 20% of it is selected as the independent test set. This test set is used to evaluate the generalization ability of the final model and does not participate in the model training and validation of the following steps. This test set is used to evaluate the generalization ability of the final model.

[0102] B: Use the five-fold cross-validation method to divide the remaining data in the original borehole dataset to obtain the training set and the validation set.

[0103] Here, the five-fold cross-validation method (5-Fold Cross-Validation) is a commonly used model evaluation method, and its core idea is to more reliably evaluate the generalization performance of the model through multiple data divisions and model validations.

[0104] For the above-mentioned step B, in specific implementation, the five-fold cross-validation method is used to divide the remaining data. The remaining data is evenly divided into 5 subsets for cross-validation. During the cross-validation process, one of the subsets is selected as the validation set each time, and the remaining 4 subsets are used as the training set. This method can not only maximize the utilization rate of data but also effectively detect the stability and effectiveness of the model on different data partitions. During the data partitioning process, special attention should be paid to ensuring the uniformity of the spatial distribution of each data set. Avoid that a certain data set only contains data in a certain concentrated area, so as to ensure that the formation data in training and testing can fully represent the geological distribution of the entire research area.

[0105] C: Perform SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set.

[0106] Here, the basic principle of SMOTE oversampling is to generate new samples by interpolating the features of minority class samples, so as to increase the number of minority class samples and achieve the purpose of balancing the class distribution.

[0107] For the above-mentioned step C, in specific implementation, in each cross-validation, perform SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set. In this way, by introducing the SMOTE technology, this application can generate synthetic minority class samples, effectively balance the sample distribution of various formations, thus avoiding the training bias caused by data imbalance and improving the prediction accuracy of the model for the minority class.

[0108] Specifically, for the above-mentioned step C, the performing SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set includes:

[0109] a: Obtain the minority class samples in the training set and determine multiple neighbor samples corresponding to each minority class sample.

[0110] b: For each minority class sample, determine a target neighbor sample from the multiple neighbor samples corresponding to the minority class sample, and perform linear interpolation between the minority class sample and the target neighbor sample to generate an augmented sample, so as to obtain the target training set.

[0111] For the above-mentioned steps a - b, in specific implementation, first identify the minority class samples, determine the minority class samples in the training set, and determine the k neighbor samples of each minority class sample in the training set. Then, for each minority class sample, randomly select one from the k neighbor samples corresponding to the minority class sample as the target neighbor sample. Generate an augmented sample through linear interpolation between the minority class sample and the target neighbor sample, as shown in the following formula:

[0112] p new =pi +λ·(p neighbor -p i )

[0113] where p i is a minority-class sample; p neighbor is its target neighbor sample; λ is a random number in [0, 1].

[0114] In this way, the minority-class samples in the training set are augmented to obtain the augmented target training set. The Synthetic Minority Over-sampling Technique (SMOTE) is used to generate synthetic samples of scarce classes (such as silt, strongly weathered argillaceous siltstone, etc.) to balance the formation data set. The problem of class imbalance is solved, the prediction ability of the model for minority-class formations is improved, and the training bias is reduced.

[0115] D: The original formation distribution prediction model is trained and verified using the target training set and the validation set respectively until convergence.

[0116] E: The converged original formation distribution prediction model is tested using the test set, and the converged original formation distribution prediction model that passes the test is used as the formation distribution prediction model.

[0117] Here, as an example, during the model optimization process, the influence of various parameters on the model performance was observed through multiple rounds of pre-training experiments, and the following parameter configurations were finally determined. The learning rate is set to 0.0001; the batch size is set to 4096; the number of training epochs is set to 200; the convolutional kernel size of each neural network layer is set to 3×3, and 16 to 256 convolutional kernels are used in each layer.

[0118] For the above steps D - E, in specific implementation, by dividing the training set into five equal parts, each time one part is selected as the validation set, and the remaining four parts are used as the training set, the training and validation process is repeated, and a total of five rounds of training are performed, and finally the performance metrics are calculated. By comparing the results of cross-validation, the finally converged original formation distribution prediction model is determined. The converged original formation distribution prediction model is tested using the test set, and the converged original formation distribution prediction model that passes the test is used as the formation distribution prediction model. Here, for the classification problem of multi-class uneven samples, this application uses three metrics, namely Precision, Recall, and F1-score, to comprehensively evaluate the performance of the model. The formulas for the evaluation metrics are as follows:

[0119]

[0120] Please refer to Figure 5 , Figure 5Schematic diagram of the loss curve and evaluation index curve of a CNN model during the cross-validation process provided by an embodiment of the present application. As Figure 5 shown, the change of the loss function curve and each evaluation index with the number of training iterations. It can be seen from Figure 5 that the difference between the training set loss and the validation set loss always remains within a certain range, indicating that the model has not experienced overfitting or underfitting during the training process. In addition, the steady decline of the validation set loss further proves that the model structure and training strategy can effectively adapt to the data characteristics and have high stability under different data segmentation conditions. In addition, the evaluation index also quickly tends to be stable. In the five-fold cross-validation, the values of the three evaluation indexes are all stable at about 0.7-0.75, and the precision and recall rate are almost the same. This shows that the developed model has achieved a good balance in the correct identification of the formation and the ability to reduce false alarms, and has good prediction ability for the multi-classification complex sample problem in the present application.

[0121] The prediction method for soil formation distribution provided by the embodiment of the present application. First, for each target position point in the soil area to be predicted, multiple adjacent borehole points closest to the target position point are determined from the borehole set; then, the formation category information corresponding to each adjacent borehole point is obtained, and a feature matrix corresponding to the target position point is generated by using the multi-dimensional vector corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point; finally, the feature matrix is input into a pre-trained formation distribution prediction model to determine the formation distribution prediction result corresponding to the target position point, and the target formation category to which the target position point belongs is determined based on the formation distribution prediction result; wherein, the formation prediction distribution result represents the probability that the target position point belongs to each preset formation category, and the training data used in the training process of the formation distribution prediction model is obtained by expanding through the SMOTE oversampling method.

[0122] This application realizes the efficient prediction of shallow complex strata by introducing feature extraction of adjacent borehole data, which is applicable to sites containing multiple strata categories. Through the combination of multiple technologies in model design and data processing, it can effectively capture the local geological features of the target location, enabling the model to flexibly adapt to unevenly distributed borehole data. By selecting the data of adjacent borehole points and combining with the spatial neighborhood strategy, multiple borehole points closest to the target point are selected as input features. Encoding multiple strata categories into multi-dimensional vectors respectively can ensure that the model can better identify and distinguish complex multi-class strata types. For the rare strata categories in the model training dataset, the Synthetic Minority Oversampling Technique (SMOTE) is used to augment the minority sample data. This application is applicable to complex site conditions containing multiple strata categories. Only multiple adjacent borehole points are required to quickly predict the strata distribution of any unknown point within the range, providing technical support for refined geological modeling. It improves its prediction accuracy, boundary recognition ability, and the recognition ability of minority strata types, enhances the automation level of strata modeling, reduces manual intervention, and improves the exploration accuracy and efficiency.

[0123] Please refer to Figure 6 , Figure 6 which is a schematic structural diagram of a prediction device for soil strata distribution provided by an embodiment of this application. As Figure 6 shown in

[0124] The borehole point determination module 601 is configured to determine, for each target position point in the soil area to be predicted, multiple adjacent borehole points closest to the target position point from the borehole set;

[0125] The feature matrix generation module 602 is configured to obtain the strata category information corresponding to each adjacent borehole point, and generate a feature matrix corresponding to the target position point by using the multi-dimensional vector corresponding to each preset strata category and the strata category information corresponding to each adjacent borehole point;

[0126] The strata category prediction module 603 is configured to input the feature matrix into a pre-trained strata distribution prediction model, determine the strata distribution prediction result corresponding to the target position point, and determine the target strata category to which the target position point belongs based on the strata distribution prediction result; wherein, the strata prediction distribution result represents the probability that the target position point belongs to each preset strata category, and the training data used in the training process of the strata distribution prediction model is obtained by augmenting through the SMOTE oversampling method.

[0127] Furthermore, when the drilling point determination module 601 is used to determine multiple neighboring drilling points closest to the target position point from a set of drilling points, the drilling point determination module 601 is further used to:

[0128] For each drilling point to be screened in the set of drilling points, determine the spatial position coordinates of the drilling point to be screened relative to the target position point, and calculate the Euclidean distance between the drilling point to be screened and the target position point using the spatial position coordinates;

[0129] Sort the multiple drilling points to be screened in ascending order of distance using the Euclidean distance corresponding to each drilling point to be screened, and select a preset number of drilling points to be screened from the sorting as the neighboring drilling points.

[0130] Furthermore, when the feature matrix generation module 602 is used to generate a feature matrix corresponding to the target position point using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each neighboring drilling point, the feature matrix generation module 602 is further used to:

[0131] Perform grid processing on the formation distribution of the soil area to be predicted to form a high-resolution formation distribution matrix;

[0132] For each neighboring drilling point, determine the initial feature code corresponding to the neighboring drilling point from the multi-dimensional vectors corresponding to multiple preset formation categories based on the formation category information corresponding to the neighboring drilling point, and determine the depth information corresponding to the neighboring drilling point from the formation distribution matrix;

[0133] Concatenate the spatial position coordinates of the neighboring drilling point relative to the target position point, the Euclidean distance between the drilling point to be screened and the target position point, the depth information corresponding to the neighboring drilling point, and the initial feature code corresponding to the neighboring drilling point to generate the target feature code corresponding to the neighboring drilling point;

[0134] Generate a feature matrix corresponding to the target position point using the target feature codes corresponding to multiple neighboring drilling points.

[0135] Furthermore, the prediction device 600 further includes a model training module, and the model training module is used to train the formation distribution prediction model through the following steps:

[0136] Determine a test set from the original drilling dataset, and use the remaining data in the original drilling dataset as the training set and the validation set;

[0137] Use five-fold cross-validation to divide the remaining data in the original drilling dataset to obtain a training set and a validation set;

[0138] Perform SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set;

[0139] Train and validate the original formation distribution prediction model using the target training set and the validation set respectively until convergence;

[0140] Test the converged original formation distribution prediction model using the test set, and use the converged original formation distribution prediction model that passes the test as the formation distribution prediction model.

[0141] Further, when the model training module is used to perform SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set, the model training module is further used to:

[0142] Obtain the minority class samples in the training set and determine multiple neighbor samples corresponding to each minority class sample;

[0143] For each minority class sample, determine a target neighbor sample from the multiple neighbor samples corresponding to the minority class sample, and perform linear interpolation between the minority class sample and the target neighbor sample to generate an augmented sample, so as to obtain the target training set.

[0144] Please refer to Figure 7 , Figure 7 which is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As Figure 7 shown in, the electronic device 700 includes a processor 710, a memory 720, and a bus 730.

[0145] The memory 720 stores machine-readable instructions executable by the processor 710. When the electronic device 700 runs, the processor 710 communicates with the memory 720 through the bus 730. When the machine-readable instructions are executed by the processor 710, the steps of the prediction method for soil formation distribution in the method embodiment as shown above can be executed. The specific implementation manner can refer to the method embodiment and will not be elaborated here. Figure 1 shown.

[0146] An embodiment of the present application further provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is run by a processor, the steps of the prediction method for soil formation distribution in the method embodiment as shown above can be executed. The specific implementation manner can refer to the method embodiment and will not be elaborated here. Figure 1 shown.

[0147] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0148] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some communication interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0149] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0150] In addition, in each embodiment of the present application, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0151] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0152] Finally, it should be noted that the above-described embodiments are only specific implementation manners of the present application, used to illustrate the technical solutions of the present application, rather than limiting it. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the technical field can still modify the technical solutions described in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for predicting soil stratum distribution, characterized in that, The prediction method includes: For each target location point in the soil area to be predicted, determine multiple adjacent borehole points closest to the target location point from the borehole set; Obtain the formation category information corresponding to each adjacent borehole point, and generate a feature matrix corresponding to the target location point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point; Input the feature matrix into a pre-trained formation distribution prediction model, determine the formation distribution prediction result corresponding to the target location point, and determine the target formation category to which the target location point belongs based on the formation distribution prediction result; wherein, the formation prediction distribution result represents the probability that the target location point belongs to each preset formation category, and the training data used in the training process of the formation distribution prediction model is obtained by expanding through the SMOTE oversampling method.

2. The prediction method according to claim 1, wherein The determining multiple adjacent borehole points closest to the target location point from the borehole set includes: For each borehole point to be screened in the borehole set, determine the spatial position coordinates of the borehole point to be screened relative to the target location point, and calculate the Euclidean distance between the borehole point to be screened and the target location point by using the spatial position coordinates; Sort the multiple borehole points to be screened in ascending order of distance by using the Euclidean distance corresponding to each borehole point to be screened, and select a preset number of borehole points to be screened from the sorting as the adjacent borehole points.

3. The prediction method according to claim 2, wherein The generating a feature matrix corresponding to the target location point by using the multi-dimensional vectors corresponding to each preset formation category and the formation category information corresponding to each adjacent borehole point includes: Perform grid processing on the formation distribution of the soil area to be predicted to form a high-resolution formation distribution matrix; For each adjacent borehole point, determine the initial feature code corresponding to the adjacent borehole point from the multi-dimensional vectors corresponding to multiple preset formation categories based on the formation category information corresponding to the adjacent borehole point, and determine the depth information corresponding to the adjacent borehole point from the formation distribution matrix; Concatenate the spatial position coordinates of the adjacent borehole point relative to the target location point, the Euclidean distance between the borehole point to be screened and the target location point, the depth information corresponding to the adjacent borehole point, and the initial feature code corresponding to the adjacent borehole point to generate the target feature code corresponding to the adjacent borehole point; Generate a feature matrix corresponding to the target location point by using the target feature codes corresponding to multiple adjacent borehole points.

4. The prediction method according to claim 1, wherein The formation distribution prediction model is trained through the following steps: Determine a test set from the original borehole dataset, and use the remaining data in the original borehole dataset as the training set and the validation set; ​ ​ ​ The original prediction model of the stratum distribution after convergence is tested using the test set, and the original prediction model of the stratum distribution after convergence that passes the test is used as the stratum distribution prediction model.

5. The prediction method according to claim 4, wherein Performing SMOTE oversampling on the minority class samples in the training set to obtain an augmented target training set, including: Obtaining the minority class samples in the training set and determining a plurality of neighbor samples corresponding to each minority class sample; For each minority class sample, determining a target neighbor sample from the plurality of neighbor samples corresponding to the minority class sample, and performing linear interpolation between the minority class sample and the target neighbor sample to generate an augmented sample, so as to obtain the target training set.

6. A prediction device for soil stratum distribution, characterized in that, The prediction device includes: A borehole point determination module, configured to determine, for each target position point in the soil area to be predicted, a plurality of adjacent borehole points closest to the target position point from a borehole set; A feature matrix generation module, configured to obtain the stratum category information corresponding to each adjacent borehole point, and generate a feature matrix corresponding to the target position point by using the multi-dimensional vectors corresponding to each preset stratum category and the stratum category information corresponding to each adjacent borehole point; A stratum category prediction module, configured to input the feature matrix into a pre-trained stratum distribution prediction model, determine the stratum distribution prediction result corresponding to the target position point, and determine the target stratum category to which the target position point belongs based on the stratum distribution prediction result; wherein, the stratum prediction distribution result represents the probability that the target position point belongs to each preset stratum category, and the training data used in the training process of the stratum distribution prediction model is obtained by augmentation using the SMOTE oversampling method.

7. The prediction device according to claim 6, characterized in that, When the borehole point determination module is used to determine a plurality of adjacent borehole points closest to the target position point from the borehole set, the borehole point determination module is further configured to: For each borehole point to be screened in the borehole set, determining the spatial position coordinates of the borehole point to be screened relative to the target position point, and calculating the Euclidean distance between the borehole point to be screened and the target position point by using the spatial position coordinates; Sorting the plurality of borehole points to be screened in ascending order of distance by using the Euclidean distance corresponding to each borehole point to be screened, and selecting a preset number of borehole points to be screened from the sorting as the adjacent borehole points.

8. The prediction device according to claim 7, characterized in that, When the feature matrix generation module is used to generate a feature matrix corresponding to the target position point by using the multi-dimensional vectors corresponding to each preset stratum category and the stratum category information corresponding to each adjacent borehole point, the feature matrix generation module is further configured to: Performing grid processing on the stratum distribution of the soil area to be predicted to form a high-resolution stratum distribution matrix; For each adjacent borehole point, determining the initial feature code corresponding to the adjacent borehole point from the multi-dimensional vectors corresponding to multiple preset stratum categories based on the stratum category information corresponding to the adjacent borehole point, and determining the depth information corresponding to the adjacent borehole point from the stratum distribution matrix; Concatenate the spatial position coordinates of the adjacent borehole point relative to the target position point, the Euclidean distance between the borehole point to be screened and the target position point, the depth information corresponding to the adjacent borehole point, and the initial feature encoding corresponding to the adjacent borehole point to generate the target feature encoding corresponding to the adjacent borehole point; Generate a feature matrix corresponding to the target position point by using the target feature encodings corresponding to multiple adjacent borehole points.

9. An electronic device, characterized in that, Including: A processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are run by the processor, the steps of the prediction method for soil stratum distribution according to any one of claims 1 to 5 are executed.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium. When the computer program is run by the processor, the steps of the prediction method for soil stratum distribution according to any one of claims 1 to 5 are executed.

Images