A method, device, medium and equipment for classifying and identifying surrounding rocks while drilling

Abstract

The present invention discloses a method, device, medium and equipment for classifying and identifying surrounding rocks while drilling, which relates to the technical field of geological exploration. The method includes: collecting surrounding rock data of different surrounding rock categories while drilling, and establishing an initial model introducing a dynamic preprocessor and a contrast learning engine; training the initial model by using the surrounding rock data; until the initial model meets the preset training end condition, determining the initial model that meets the training end condition as the surrounding rock classification and identification model while drilling; inputting the obtained surrounding rock data to be classified while drilling into the surrounding rock classification and identification model while drilling for processing, and obtaining the classification information corresponding to the surrounding rock data to be classified while drilling; realizing efficient and accurate analysis of the surrounding rock data, significantly improving the speed and accuracy of geological exploration; the method not only reduces the dependence on manual experience, reduces costs and workload, but also enhances the generalization ability and adaptability of the model, and can handle complex and diverse geological conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of geological exploration, and more specifically, it relates to a method, device, medium, and equipment for classifying and identifying surrounding rocks while drilling. Background Art

[0002] In the traditional process of classifying surrounding rocks while drilling, it usually relies on the experience of geologists and the intuitive analysis of core samples, logging data, etc.; however, this method is inefficient and subject to human subjective judgment. With the development of computer technology and machine learning, it has become possible to apply automated methods for classifying and identifying surrounding rocks while drilling, and it is expected to improve accuracy, consistency, and efficiency. Summary of the Invention

[0003] Aiming at the deficiencies in the prior art, the purpose of the present invention is to provide a method, device, medium, and equipment for classifying and identifying surrounding rocks while drilling, so as to achieve accurate classification of surrounding rocks while drilling.

[0004] The above technical objectives of the present invention are achieved through the following technical solutions:

[0005] In the first aspect, the present application provides a method for classifying and identifying surrounding rocks while drilling, including the following specific steps:

[0006] Collect surrounding rock data of different surrounding rock categories while drilling, and establish an initial model introducing a dynamic preprocessor and a contrast learning engine. The surrounding rock data includes logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data;

[0007] Use the surrounding rock data to train the initial model; among them, the dynamic preprocessor is used to preprocess the surrounding rock data with an optimized processing strategy, and the contrast learning engine is used to learn the feature representations for distinguishing different surrounding rock categories while drilling in the absence of class labels for the surrounding rock data;

[0008] Until the initial model meets the preset training end condition, determine the initial model that meets the training end condition as the surrounding rock classification and identification model while drilling;

[0009] Input the obtained surrounding rock data to be classified while drilling into the surrounding rock classification and identification model while drilling for processing, and obtain the classification information corresponding to the surrounding rock data to be classified while drilling.

[0010] Based on the above technical solutions, the present invention can also be improved as follows.

[0011] Further, the above initial model includes an input and preprocessing module, a self-supervised learning module, a meta-learning module, a feature fusion and classification module, and an output module. The preprocessing module includes a dynamic preprocessor, and the self-supervised learning module includes a contrast learning engine.

[0012] Furthermore, the above input and preprocessing module further includes a multi-modal sensor, where:

[0013] The multi-modal sensor is used to receive and integrate surrounding rock data or drill-bit surrounding rock data to be classified. The output of the multi-modal sensor is specifically:

[0014] ; , ;

[0015] In the formula, represents the output of the multi-modal sensor; represents the non-linear activation function; represents the weight matrix; is the bias vector; is the multi-modal feature representation of the data, representing the result of weighted summation of individual modal features ; represents the modal feature of importance weight, represents the total number of modalities; represents the th type of original input data of the modality; is the encoder function corresponding to the th type of modality;

[0016] The dynamic preprocessor is used to select the optimal processing strategy according to the data characteristics and preprocess the integrated data. The preprocessing includes feature extraction processing and noise reduction processing. The output of the dynamic preprocessor is specifically:

[0017] ;

[0018] In the formula, represents the output of the dynamic preprocessor; represents the original data matrix; represents the multi-modal sensor; is the adaptive enhancement parameter; represents the adaptive unit; represents the feature weight vector; represents feature selection and recombination; represents the adversarial deep autoencoder; represents the non-linear activation function in the autoencoder, represents the normalization process.

[0019] Furthermore, the above contrast learning engine learns the feature representation for distinguishing different drill-bit surrounding rock categories in the case where the surrounding rock data has no class labels, and is implemented through the following steps:

[0020] Randomly and reasonably transform the input data to form positive sample pairs and negative sample pairs. The positive sample pairs include a first enhanced sample and a second enhanced sample obtained through different transformation methods. The negative sample pairs include a third sample, and the first enhanced sample or the second enhanced sample. Moreover, the original data in the positive sample pairs and the original data in the negative sample pairs are surrounding rock data of different types of surrounding rocks while drilling.

[0021] Calculate the contrast loss between the positive sample pairs and the negative sample pairs using a preset contrast loss function to learn the feature representation for distinguishing different types of surrounding rocks while drilling in the case where the surrounding rock data has no class labels.

[0022] Furthermore, the above contrast loss function is specifically:

[0023] , where:

[0024] ;

[0025] , ;

[0026] In the formula, represents the contrast loss, represents a hyperparameter; represents the multi-view contrast loss, represents the local structure preservation term; represents the temperature parameter, represents the cosine similarity function, represents the number of data sources or views, represents a single view or data source, including logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, geochemical data; represents the positive sample pair, represents the negative sample pair, represents the batch size; represents, represents the degree matrix, represents the feature representation matrix of all samples after passing through the encoder, represents the matrix transpose.

[0027] Furthermore, the above feature fusion and classification module is used to effectively fuse the features of different data sources and compress them into a compact representation form, specifically:

[0028] , where:

[0029] ;

[0030] In the formula, represents the fused representation form, denotes the activation function, denotes the gating unit; denotes the gating parameter; denotes the weight coefficient of the denotes the feature transformation function of the denotes the bias term; denotes the parameter set of the feature transformation function denotes the scoring function; denotes the feature representation of the denotes the number of different modalities.

[0031] Furthermore, the above output module is used to output classification information, and the classification information is specifically:

[0032] , ; where:

[0033] , ;

[0034] In the formula, denotes the classification information, which is the class label output by the surrounding rock classification and recognition model while drilling; denotes the number of class labels, denotes the enhanced score of each class label obtained after the multi-class confidence weighting and context-dependent classification steps, denotes the classification confidence score, denotes the set of neighboring classes the sum of all class probabilities in, is the probability corresponding to a neighboring class , denotes the set of neighboring classes related to the denotes the adjustment parameter, which is used to balance the direct probability and the influence of related classes; denotes another adjustment parameter, which is used to balance the original score and the context information weight.

[0035] In a second aspect, the present application provides a surrounding rock classification and recognition device while drilling, which is applied to a surrounding rock classification and recognition method according to any one of the first aspect, and includes:

[0036] A model construction module, which is used to collect surrounding rock data of different types of surrounding rocks while drilling, and establish an initial model incorporating a dynamic preprocessor and a contrastive learning engine. The surrounding rock data includes logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data;

[0037] A model training module, which is used to train the initial model using the surrounding rock data. Among them, the dynamic preprocessor is used to preprocess the surrounding rock data with an optimized processing strategy, and the contrastive learning engine is used to learn the feature representations for distinguishing different types of surrounding rocks while drilling in the absence of class labels for the surrounding rock data;

[0038] A training end module, which is used to determine the initial model that meets the preset training end condition as the surrounding rock classification and recognition model while drilling until the initial model meets the preset training end condition;

[0039] A classification and recognition module, which is used to input the obtained surrounding rock data to be classified while drilling into the surrounding rock classification and recognition model for processing to obtain the classification information corresponding to the surrounding rock data to be classified while drilling.

[0040] In a third aspect, the present application provides a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the method of any one of the first aspects.

[0041] In a fourth aspect, the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method of any one of the first aspects is implemented.

[0042] Compared with the prior art, the present invention has at least the following beneficial effects:

[0043] In the present application, first, an initial model incorporating a dynamic preprocessor and a contrastive learning engine is constructed. Among them, the dynamic preprocessor is used to preprocess the surrounding rock data with an optimized processing strategy, and the contrastive learning engine is used to learn the feature representations for distinguishing different types of surrounding rocks while drilling in the absence of class labels for the surrounding rock data; second, the constructed initial model is trained using the surrounding rock data of different types of surrounding rocks while drilling, so that the model can identify the corresponding categories from the outputs such as logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data; finally, after the initial model is trained, the surrounding rock classification and recognition model while drilling can be obtained, and the obtained surrounding rock data to be classified while drilling is input into the surrounding rock classification and recognition model while drilling for processing, so as to obtain the classification information corresponding to the surrounding rock data to be classified while drilling.

[0044] In this application, by introducing the method for classifying and identifying surrounding rocks while drilling, efficient and accurate analysis of surrounding rock data is achieved, significantly improving the speed and accuracy of geological exploration. This method not only reduces the dependence on manual experience, lowers costs and workloads, but also enhances the generalization ability and adaptability of the model, enabling it to handle complex and diverse geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of this application, and do not limit the embodiments of the present invention. In the drawings:

[0046] Figure 1 is the flowchart of the classification and identification method in the embodiments of the present invention;

[0047] Figure 2 is a schematic diagram of the model for classifying and identifying surrounding rocks while drilling in the embodiments of the present invention;

[0048] Figure 3 is a schematic diagram of the input and preprocessing module in the embodiments of the present invention;

[0049] Figure 4 is a schematic diagram of the self-supervised learning module in the embodiments of the present invention;

[0050] Figure 5 is a schematic diagram of the meta-learning module in the embodiments of the present invention;

[0051] Figure 6 is a schematic diagram of the feature fusion and classification module in the embodiments of the present invention;

[0052] Figure 7 is a schematic diagram of the cross-domain attention network in the embodiments of the present invention;

[0053] Figure 8 is a schematic diagram of the output module in the embodiments of the present invention;

[0054] Figure 9 is a schematic diagram of the connection of the classification and identification device in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components of the embodiments of the present invention described and illustrated in the drawings here can be arranged and designed in various different configurations.

[0056] Accordingly, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0057] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0058] In the description of the embodiments of the present invention, "a plurality of" represents at least two.

[0059] Embodiment 1: In order to solve the problems in the traditional classification of surrounding rocks while drilling, which usually relies on the experience of geologists and the intuitive analysis of core samples, logging data, etc., resulting in low efficiency and being subject to subjective judgment of others. With the development of computer technology and machine learning, and in order to improve the accuracy, consistency and efficiency of the classification of surrounding rocks while drilling, this embodiment provides a method for classifying and identifying surrounding rocks while drilling, as Figure 1 shown, including the following specific steps:

[0060] S1. Collect surrounding rock data of different surrounding rock categories while drilling and establish an initial model introducing a dynamic preprocessor and a contrast learning engine. The surrounding rock data includes logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data.

[0061] Among them, the above-mentioned logging data includes resistivity logging data, density logging data, and natural gamma ray logging data; the core sample analysis data includes porosity and mineral composition; the seismic reflection data includes reflection amplitude and phase data; the surrounding rock thickness and depth data includes the buried depth data of the top and bottom plates of each layer and the variation data of the layered thickness; the geochemical data includes stable isotope ratios (such as isotope ratios of elements such as carbon, oxygen, sulfur, etc.) and trace element concentrations.

[0062] For the initial model introducing a dynamic preprocessor and a contrast learning engine, a multi-modal perceptron can be used to receive and integrate the surrounding rock data; the dynamic preprocessor can be used to automatically select the best preprocessing method according to the data characteristics and perform feature extraction and noise reduction on the integrated surrounding rock data using a deep autoencoder.

[0063] Optionally, the above-mentioned initial model is as Figure 2 shown, including an input and preprocessing module, a self-supervised learning module, a meta-learning module, a feature fusion and classification module, and an output module. The preprocessing module includes a dynamic preprocessor, and the self-supervised learning module includes a contrast learning engine.

[0064] Among them, the above-mentioned input and preprocessing module further includes a multi-modal sensor, such as Figure 3 shown in the schematic diagram of the input and preprocessing module. The input and preprocessing module will be further described below.

[0065] Among them, the above-mentioned multi-modal sensor is used to receive and integrate surrounding rock data or logging while drilling surrounding rock data to be classified. The output of the multi-modal sensor is specifically as follows:

[0066] ; , ;

[0067] In the formula, represents the output of the multi-modal sensor; represents a non-linear activation function, such as ReLU, Sigmoid; represents a weight matrix used to connect the multi-modal feature representation and the output of the multi-modal sensor y ; is a bias vector; is the multi-modal feature representation of the data, representing the result of the weighted sum of the single-modal features ; represents the modal feature The importance weight of represents the total number of modalities; represents the th type of original input data of the modality; is the encoder function corresponding to the th modality.

[0068] Specifically, the perception process of the multi-modal sensor for the surrounding rock data is described as follows:

[0069] For each type of surrounding rock data (such as logging data, core sample analysis data, seismic reflection data, etc.), first, through the corresponding encoder multi-modal sensor function Convert the original input data into a single-modal feature representation , this step ensures that each type of data is converted into a unified and processable form. Secondly, all single-modal feature representations are weighted and summed to obtain the multi-modal feature representation , where the importance of each modality is reflected by the weight . Finally, the multi-modal feature representation is transformed through the non-linear activation function , and combined with the weight matrix and the bias vector b to generate the final outputy By integrating different types of surrounding rock data using a multi-modal sensor, this application can improve the accuracy and robustness of the classification and recognition model for surrounding rock during drilling.

[0070] Furthermore, the above dynamic preprocessor is used to select an optimal processing strategy according to data characteristics and preprocess the integrated data. The preprocessing includes feature extraction and noise reduction. The output of the dynamic preprocessor is specifically:

[0071] ;

[0072] In the formula, represents the output of the dynamic preprocessor; represents the original data matrix, where each row represents a sample point and each column represents an observation feature (such as logging data, core images, seismic reflection data, etc.); represents the multi-modal sensor; is an adaptive enhancement parameter that intelligently selects a data enhancement strategy (such as rotation, scaling, brightness adjustment, etc.) suitable for the current sample set according to the statistical characteristics of the input data; represents the adaptive unit, which performs a series of random but reasonable transformations on the input data based on to increase data diversity and help the model generalize better; represents the feature weight vector, which is obtained by evaluating the importance of each feature through an interpretive analysis method (such as Shapley value or LIME); represents feature selection and recombination, which screens out the most valuable feature subset according to and recombines these features to discover potential feature interaction patterns; represents an adversarial deep autoencoder, which, through adversarial training (introducing the idea of the generative adversarial network GAN), enables the autoencoder to learn more robust feature representations in an adversarial environment to prevent overfitting; represents the non-linear activation function in the autoencoder, such as ReLU; represents normalization processing, which normalizes the features processed through the above steps to ensure that all features have the same scale range.

[0073] Specifically, the process of the dynamic preprocessor dynamically preprocessing the integrated surrounding rock data can be divided into the following steps:

[0074] First, the multi-modal sensor intelligently selects a data enhancement strategy suitable for the current sample set according to the statistical characteristics of the input data, such as including rotation, scaling, brightness adjustment, etc., aiming to improve the quality and diversity of the data. Next, the adaptive unit is based on the adaptive enhancement parameter Perform a series of random but reasonable transformations on the input data. Then, through the feature weight vector , use the interpretive analysis method to evaluate the importance of each feature. According to the evaluation results, screen out the most valuable feature subset, and reorganize these features to discover potential feature interaction patterns. This step aims to extract the feature combinations that contribute most to the model prediction. Subsequently, through adversarial training (introducing the idea of the generative adversarial network GAN), let the autoencoder learn more robust feature representations in an adversarial environment. This training method can prevent overfitting and ensure that the model can still maintain good performance when facing complex and changing data. Finally, perform standardization processing on the features processed through the above steps to ensure that all features have the same scale range, and avoid some features dominating the model learning process due to too large numerical ranges, thus affecting the fairness and accuracy of the model.

[0075] In the above, the dynamic preprocessor integrates multi-modal surrounding rock data, intelligently selects the optimal data augmentation strategy, dynamically adjusts feature selection and recombination, and uses an adversarial deep autoencoder for robust feature extraction and noise reduction, and finally outputs a standardized and high-quality feature matrix. This process can not only automatically optimize the preprocessing steps to significantly improve the quality of the surrounding rock data and the generalization ability of the model, but also enhance the understanding and recognition accuracy of complex geological features, laying a solid foundation for the subsequent classification and recognition tasks of the surrounding rock while drilling.

[0076] Optionally, the above contrast learning engine learns the feature representations for distinguishing different types of surrounding rock while drilling in the case where the surrounding rock data has no class labels, and is implemented through the following steps:

[0077] S11. Perform random and reasonable transformations on the input data to form positive sample pairs and negative sample pairs. The positive sample pairs include a first augmented sample and a second augmented sample obtained through different transformation methods. The negative sample pairs include a third sample, and the first augmented sample or the second augmented sample, and the original data in the positive sample pairs and the original data in the negative sample pairs are surrounding rock data of different types of surrounding rock while drilling.

[0078] S12. Use the preset contrast loss function to calculate the contrast loss between the positive sample pairs and the negative sample pairs, so as to learn the feature representations for distinguishing different types of surrounding rock while drilling in the case where the surrounding rock data has no class labels.

[0079] The above steps S11 - S12 are implemented through the contrast learning engine in the self-supervised learning module, such as Figure 4As shown, the self-supervised learning module may include a contrastive learning engine and a predictive modeling unit. Among them, the contrastive learning engine is used to construct positive sample pairs (similar samples of surrounding rock while drilling) and negative sample pairs (different samples of surrounding rock while drilling), so that the surrounding rock classification and recognition model can also learn useful representations from the surrounding rock data without labels. The predictive modeling unit is used to train the model to predict the missing data part or future time series values, so as to enhance the model's understanding of the internal structure of the surrounding rock data. For example, in seismic reflection data, the predictive modeling unit can be used to predict the change trend of the underground structure.

[0080] Furthermore, the contrastive learning engine includes an encoder, a projection head, a data augmentation unit, a positive and negative sample pair construction unit, and a contrastive loss calculation unit connected in sequence, as Figure 4 shown. Among them, the encoder uses a convolutional neural network (CNN) or a fully connected network (FCN) to convert the dynamically preprocessed surrounding rock data into a higher-level feature representation, aiming to capture the essential attributes of the input data, remove noise, and retain information crucial for distinguishing different samples. The projection head consists of 2 fully connected layers, and an activation function (such as ReLU) is set between the 2 fully connected layers to introduce non-linearity. The projection head is used to perform a non-linear transformation on the higher-level feature representation output by the encoder and map it to another feature space. This step helps to enhance the learning ability of the model, making similar data points closer in the new space and dissimilar data points farther away.

[0081] To construct positive sample pairs, the data augmentation unit needs to perform a series of random but reasonable transformations on the input data. For example, for core images, rotation, cropping, color jittering, etc. can be used to generate different views. Another example is that for logging curves, noise addition, smoothing processing, etc. can be used. By adopting a series of augmentation strategies, the data augmentation unit can ensure that even different observations of the same data point can generate rich learning signals, thus helping the model to learn more generalizable representations.

[0082] Specifically, the positive sample pair and negative sample pair construction unit is used to create positive sample pairs (i.e., different views from the same data point) for different views generated by the data augmentation unit. For example, given an original sample x, two augmented samples x1 and x2 are generated through two different augmentation methods (such as rotation and cropping), then (x1, x2) constitutes a positive sample pair; and samples are selected from different categories as negative sample pairs (i.e., views from different data points). For example, given an original sample x and its augmented sample x1, as well as an original sample y from a different category, then (x1, y) constitutes a negative sample pair. Positive sample pairs are used to guide the model to learn to make similar samples approach each other in the feature space, while negative sample pairs prompt the model to push dissimilar samples apart in the feature space.

[0083] Optionally, the above contrast loss function is specifically:

[0084] , where:

[0085] ;

[0086] , ;

[0087] In the formula, represents the contrast loss, represents a hyperparameter used to balance the importance of the multi-view contrast loss and the local structure preservation term; represents the multi-view contrast loss, represents the local structure preservation term; represents the temperature parameter, represents the cosine similarity function, represents the number of data sources or views (such as well logging data, core sample analysis data), represents a single view or data source, including well logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, geochemical data; represents the positive sample pair, represents the negative sample pair, represents the batch size; represents, represents the degree matrix, represents the feature representation matrix of all samples after passing through the encoder, represents the matrix transpose.

[0088] In summary, by using the above components to work together, the contrast learning engine can learn powerful feature representations from unlabeled surrounding rock data in a self-supervised manner; in this way, the contrast learning engine not only improves the quality of the feature representation but also enhances the performance of subsequent tasks (such as classification, regression, etc.).

[0089] As Figure 4 shown, the predictive modeling unit includes an input adaptation layer, a feature enhancement and transformation layer, a prediction core module, a shared layer, and an output layer connected in sequence. Among them, the input adaptation layer is responsible for receiving the feature representations generated by the contrast learning engine and can be customized according to specific geological data to adapt to different types of data formats. For example, well logging curves are usually one-dimensional time series or depth series data. For such data, one-dimensional convolution operations or recurrent neural networks (RNNs) can be used in the input adaptation layer to further extract features and adjust the feature dimensions to a size suitable for subsequent model use. Core images, as two-dimensional spatial data, need to be processed using two-dimensional convolution operations. The input adaptation layer contains appropriate pooling layers and smoothing layers to ensure that image features are correctly captured and converted into a format suitable for the prediction model. Seismic reflection data involves complex three-dimensional structural information, and 3D CNNs or other advanced techniques can be used to analyze these data and convert them into effective feature representations. Suppose a set of in-well surrounding rock datasets containing well logging curves and core images is being processed. For well logging curves, the input adaptation layer first extracts local patterns through 1D CNN, then uses a fully connected layer to compress these patterns into a fixed-length feature vector and standardizes it. For core images, 2D CNN is used to extract texture features, and after being processed by the pooling layer and smoothing layer, a fixed-length feature vector is also generated. Finally, these two feature vectors are combined and their dimensions are adjusted through appropriate linear transformations to match the requirements of the subsequent prediction model. This customized design not only improves the model's ability to process different types of geological data but also enhances its pertinence and accuracy for the in-well surrounding rock classification task.

[0090] Specifically, the feature enhancement and transformation layer introduces a self-attention mechanism to help the model focus on the most predictive feature parts, which helps capture long-range dependencies. Especially when dealing with complex geological phenomena, this mechanism can help the model better understand the long-term trends of the data.

[0091] Furthermore, the prediction core module adopts a Transformer architecture that combines an encoder-decoder structure with a self-attention mechanism. This architecture can process long time series or multi-dimensional datasets more efficiently while maintaining high performance. Among them, the encoder is responsible for compressing the input feature representations, removing noise, and obtaining a more concentrated hidden representation; the decoder reconstructs the output based on the hidden representation generated by the encoder to predict future values or fill in missing data.

[0092] Furthermore, the common features among multiple prediction tasks can be learned through the shared layer to promote knowledge transfer and generalization ability. In in-well surrounding rock classification, this allows the model to use different but related geological information for comprehensive analysis, thereby improving classification accuracy.

[0093] Furthermore, the output layer can output corresponding prediction results according to different prediction tasks. For example, for numerical prediction, it outputs continuous values; for classification tasks, it outputs discrete class labels. In addition, it can also output metrics for measuring uncertainty, such as confidence intervals or probability distributions, to better understand the quality of the prediction.

[0094] In summary, the self-supervised learning module enables the model to learn the feature representations for distinguishing different types of formation surrounding the drill bit during drilling even without explicit labels by introducing a contrastive learning engine. This helps improve the model's understanding of complex and diverse formation surrounding the drill bit during drilling. When encountering new or rare types of formation surrounding the drill bit during drilling, the powerful feature representations learned by self-supervised learning can help the model adapt to the new situation faster and make effective adjustments using limited new samples to achieve accurate classification. In addition, due to the self-supervised approach, the model can rely to a large extent on large-scale unlabeled data for pre-training and then only requires a small amount of labeled data for fine-tuning to achieve good performance, thus greatly reducing the cost and time of data annotation.

[0095] Optionally, the schematic diagram of the meta-learning module in the above initial model is as Figure 5 shown. The meta-learning module includes a fast adaptation mechanism, a task generator, and a memory bank. Among them, the fast adaptation mechanism quickly adjusts the model parameters by using a small amount of sample data, enabling the model to accurately predict new types of formation surrounding the drill bit during drilling. Specifically, during the fast adaptation process, known geological principles are used as prior knowledge to guide the update of model parameters. For example, there are significant differences in physical properties (such as resistivity, density) among different types of surrounding rocks, and this information can be added as constraints to the model training. According to the characteristics of the formation surrounding the drill bit during drilling, the features that are most critical for distinguishing new types of surrounding rocks are selected and given higher weights. This can be achieved by analyzing the differential features between different types of surrounding rocks in historical data. A basic model is trained using a large amount of previously accumulated surrounding rock data, and when encountering a new type of surrounding rock, it is fine-tuned based on this basic model. This method can quickly improve the model performance with a small amount of samples. In addition, the learning rate can be dynamically adjusted according to the sample size. A larger initial learning rate is used for small sample datasets of new types of surrounding rocks to accelerate the convergence speed, and then the learning rate is gradually reduced to improve the accuracy. For example, when encountering a type of formation surrounding the drill bit during drilling that has never been seen before, the model can quickly adjust using a small amount of samples of this formation surrounding the drill bit during drilling, optimize its own parameters, and thus improve the classification accuracy of this type of formation surrounding the drill bit during drilling.

[0096] Among them, the task generator is used to randomly create a series of small-scale tasks, and each task corresponds to a classification problem of the formation surrounding the drill bit during drilling under different types or conditions, to help the model adapt to new types of formation surrounding the drill bit during drilling faster.

[0097] Specifically, the workflow of the task generator includes the following aspects:

[0098] Dataset partitioning: The original dataset is divided into multiple subsets, and each subset represents a possible task. For example, in the problem of classifying surrounding rocks while drilling, this may mean dividing the data according to different types of surrounding rocks while drilling, geological features, or other relevant attributes.

[0099] Sample selection: A certain number of samples are randomly selected from the above subsets as the support set and the query set. Among them, the support set is used to simulate the training data in the new task, and the query set is used to evaluate the performance of the model on this new task.

[0100] Task definition: Define a specific task for each selected support set - query set pair. For example, in a classification problem, the task may be to identify a specific type of surrounding rock while drilling; in a regression problem, the task may be to predict certain geochemical parameters.

[0101] Introduction of task diversity: To ensure that the model can handle various situations, the task generator deliberately introduces some variation factors, such as adding noise, changing the sample distribution, etc., to increase the challenge and diversity of the tasks.

[0102] Iterative update: As the model's learning progresses, the task generator can dynamically adjust the difficulty and complexity of the tasks according to the model's performance, such as adding more complex surrounding rock patterns while drilling or reducing the available amount of training samples.

[0103] In summary, the memory bank is used to store various surrounding rock patterns encountered in the past and their solutions, and relevant knowledge can be quickly retrieved and applied when facing similar new situations.

[0104] Optionally, the schematic diagram of the feature fusion and classification module of the above initial model is as Figure 6 shown. The feature fusion and classification module includes a cross - domain attention network and a hierarchical decision tree. Among them, the cross - domain attention network is used to enable the model to focus on the mutual relationships between different data sources, thereby constructing a more comprehensive representation of the surrounding rock features while drilling. The hierarchical decision tree is used to recursively divide the surrounding rock categories based on the importance of the features.

[0105] Specifically, the cross - domain attention network includes a multi - modal input integrator, a cross - modal interaction layer, a geological semantic enhancement module, an attention aggregator, and a classification head connected in sequence, as Figure 7As shown in the figure, the multimodal input integrator includes an encoder and a fusion layer. For different types of geological data, different encoders can be used to extract features. For example, for logging curves, 1D convolutional neural networks (1D CNNs) are used for feature extraction because logging data is essentially one-dimensional time series or depth series data; for core images, long short-term memory networks (LSTMs) or 2D convolutional neural networks (2D CNNs) are used for feature extraction because core images provide information on rock structure and texture, and different types of rock features such as fractures and grain size distributions can be identified through 2D convolutional neural networks. The fusion layer uses bilinear pooling, and the specific steps are as follows:

[0106] Suppose there are two modalities A and B, and after passing through their respective encoders, they obtain feature vectors and , where and are the dimensions of modalities A and B respectively.

[0107] Next, the feature vectors and are transformed through a bilinear mapping function, which specifically includes:

[0108] Calculate the outer product and of the feature vectors , and the calculation result is a matrix M of . Further, in order to reduce the number of parameters and computational complexity, the original feature vectors are first projected into a lower-dimensional space. Let the projection matrices be and , and the new feature vectors are and , where k represents the target dimension, and then calculate the outer product of and to obtain a smaller matrix .

[0109] Since the matrix generated by the outer product may be very large, it needs to be vectorized and normalized to maintain numerical stability: First, flatten the matrix G into a long vector v(G), and second, perform L2 norm normalization on the flattened vector v(G), that is .

[0110] Finally, perform a pooling operation on the flattened and normalized vector, that is, take the power root of each element in the normalized vector, such as the square root or a higher-order power root, and then perform L2 normalization again.

[0111] After the above steps, a fixed-length vector can be finally output. This vector contains the information interaction from two different modalities and can be used as the input of the subsequent classifier or sent to a deeper network layer together with other features for further processing.

[0112] In the classification of surrounding rocks while drilling in this application, bilinear pooling can help the model better understand the relationships between different geological data, such as the connection between well logging curves and core images, and how these connections affect the types of surrounding rocks while drilling. In this way, the model can not only learn useful patterns from a single modality but also discover important cross-modal associations, thereby improving the classification performance. In addition, since bilinear pooling introduces additional non-linear transformations, it can increase the expressiveness of the model, making the model more robust and having better generalization ability.

[0113] Furthermore, as Figure 7 shown, the cross-modal interaction layer includes a multi-modal feature mapping layer, a feature interaction mechanism, and a dynamic context modeling component connected in sequence. Among them, the multi-modal feature mapping layer uses the adaptive linear transformation technology to map data from different modalities (such as well logging curves, core images, etc.) to a common feature space through adaptive linear transformation, and can also automatically adjust the mapping parameters according to the specific characteristics of the input data. For example, when processing well logging curves, considering their time series characteristics, a specific weight adjustment strategy will be applied; for core images, an adjustment method suitable for image feature extraction will be used (for example, using a model pre-trained on a large image dataset (such as ImageNet) and applying it to the feature extraction of core images through fine-tuning; or, considering that there may be important features of different scales in core images, a multi-scale feature extraction method can be designed, which means using receptive fields of different sizes to capture local and global information in the image, helping to understand the content of core images more comprehensively). The multi-modal feature mapping layer can ensure the effective interaction of different modal features in the same dimension by adopting different processing strategies for different modal features.

[0114] Specifically, in the feature interaction mechanism, a bilinear pooling method based on the attention mechanism is introduced. This method can not only capture the complex relationships between different data sources but also highlight the key features that are crucial for the analysis of geological phenomena. For example, when combining well logging data and core images, the model can automatically focus on the key areas that may indicate changes in mineral composition or geological structure mutations.

[0115] Among them, for the time evolution characteristics of surrounding rock data, the dynamic context modeling component can adopt, for example, GRU or LSTM to model long-term dependencies and local context information, enabling the model to more effectively learn and predict geological phenomena that change over time.

[0116] Among them, by setting up a cross-modal interaction layer, not only can the ability of the model to process different types of surrounding rock data be enhanced, but also the depth of understanding and prediction accuracy of the model for geological features can be improved, thereby achieving more accurate identification of surrounding rock data.

[0117] Furthermore, as Figure 7 shown, the geological semantic enhancement module includes a geological term embedding layer, a context awareness module, and a knowledge distillation component. Among them, the geological term embedding layer is used to utilize a pre-trained language model specifically for the geological field (such as GeoBERT, which has been trained on a large number of geological literatures and can capture geology-specific semantic information) to embed professional terms and concepts in the geological field into a low-dimensional vector space, enabling the model to use this prior knowledge for learning and prediction. For example, for specific markers in well logging curves or mineral compositions identified in core images, the model can understand their meanings based on existing geological terms and perform correlation analysis.

[0118] Among them, the context awareness module includes a geological background encoder, a conditional random field, a graph neural network (GNN), a self-attention mechanism, a geological rule constraint unit, and a dynamic context adapter. Among them, the geological background encoder is used to convert the geological background information (such as geological age, geographical location, etc.) of each sample into a vector representation that can be processed by the model. The conditional random field is used to model the sequential relationship between the surrounding rocks while drilling, especially in the case of an explicit hierarchical structure. For example, when the data presents a linear sequence (such as well logging curves), a chain CRF can be used to capture the dependencies between adjacent surrounding rocks while drilling. The graph neural network (GNN) is suitable for data that can be organized into a graph structure, such as the connections between geological bodies or the network of surrounding rocks while drilling. The self-attention mechanism enables the model to focus on the most critical parts among numerous information, thereby better understanding and processing geological data. The geological rule constraint unit is used to introduce rules and constraints in the geological field to ensure that the prediction results of the model follow known geological laws. A set of rules based on geological common sense can be predefined and integrated into the model, such as sedimentary facies change rules, mineral symbiotic combinations, etc. The dynamic context adapter enables the model to dynamically adjust its own parameters according to new or rare types of surrounding rocks while drilling to adapt to changing geological conditions. Specifically, a fast adaptation mechanism (such as the MAML algorithm) can be adopted, enabling the model to quickly update its weights or parameters on a small number of samples to cope with new types of surrounding rocks while drilling.

[0119] Furthermore, as Figure 7As shown, the knowledge distillation component includes an input layer, a domain knowledge encoding layer, and a knowledge fusion layer connected in sequence. Among them, the input layer is used to receive data from multiple sources (logging curves, core images, seismic reflection data, etc.) and domain knowledge (expert rules, knowledge graphs, literature mining results). The domain knowledge encoding layer includes an expert system module, a knowledge graph module, and a literature mining module. Among them, the expert system module is used to convert the professional knowledge of geologists into the form of rule sets or decision trees and embed them into the model as additional guiding information. The knowledge graph module uses GNNs to process the knowledge graph and extract the relationships between surrounding rock types. The literature mining module uses NLP technology to automatically extract important information about the characteristics and classification of surrounding rocks from geological literature and convert it into feature vectors or label distributions for the model to learn. The knowledge fusion layer includes a weighted fusion unit and a dynamic adjustment unit. Among them, the weighted fusion unit can assign different weights according to the reliability of the knowledge source and fuse this knowledge into the learning process of the main model. The dynamic adjustment unit is used to dynamically adjust the weights of each knowledge source according to the performance of the model to optimize the overall performance of the model.

[0120] Specifically, the processing process of the geological semantic enhancement module for surrounding rock data is as follows:

[0121] First, collect multi-source data including logging curves, core images, seismic reflection data, etc., and domain knowledge from expert systems, knowledge graphs, and literature mining through the input layer. Secondly, use the geological term embedding layer to convert geological professional terms into vector representations, and at the same time convert background information into a form that can be processed by the model through the geological background encoder. Then, in the context-aware module, apply CRF to capture the dependencies between linear sequence data, GNN is used for feature extraction of graph-structured data, the self-attention mechanism highlights key features, the geological rule constraint unit ensures that the prediction conforms to geological common sense, and the dynamic context adapter ensures that the model can quickly adapt to new surrounding rock types. Finally, in the knowledge distillation component, integrate multi-source knowledge through the weighted fusion unit, and continuously optimize the weights of the knowledge sources according to the model performance by the dynamic adjustment unit, ultimately improving the accuracy and reliability of surrounding rock data recognition.

[0122] Furthermore, as Figure 7As shown in the figure, the attention aggregator includes an input feature mapping layer, a multi-head attention mechanism, a context-aware attention mechanism, a cross-modal interaction attention mechanism, and an attention fusion and compression layer that are connected in sequence. Among them, the input feature mapping layer is used to receive the feature representations of the geological semantic enhancement module and perform L2 normalization on each feature representation to make the scales of different modal features consistent, so as to prevent certain modalities from dominating the attention mechanism due to a large numerical range. The multi-head self-attention mechanism obtains richer feature representations by creating multiple independent self-attention heads, each of which is used to capture different subspace information, and then combines the results of all self-attention heads through weighted summation to form the final self-attention output, which helps the model understand the complex relationships in the surrounding rock data from multiple perspectives. The context-aware attention mechanism uses a geological background encoder (such as a model based on GRU or LSTM) to capture the changes in geological conditions during the drilling process and provides this information as additional input to the attention mechanism, so that the attention mechanism can consider the specific geological environment of the sample and adjust the attention weights according to the context information provided by the geological background encoder. For example, higher attention is given to the data at a specific depth of the surrounding rock during drilling because this depth may correspond to a key geological interface or an important geological event. The cross-modal interaction attention mechanism promotes the effective exchange of information between different modalities through an interaction layer and captures the interdependent relationships between them. In addition, graph neural networks (GNNs) can be used to model the complex relationships between different modal data. For example, a graph structure containing nodes of different types of geological data is constructed, and the connection strengths between nodes are learned through GNNs. In addition, the cross-modal interaction attention mechanism can also adopt a feature selection algorithm to select the most representative features from the fused features, reduce redundant information, and improve the model efficiency.

[0123] Specifically, the attention fusion and compression layer is used to effectively fuse the features from different data sources and compress them into a compact representation form. Specifically, methods such as max pooling and average pooling can be used to retain the most important feature information while reducing the feature dimension. For example, for the attention results of each modality, the maximum value or the average value can be selected as the final feature representation. This application introduces a functional expression of the attention fusion and compression layer, which is specifically expressed as follows:

[0124] , where:

[0125] ;

[0126] In the formula, represents the fused representation form, represents the activation function, such as ReLU or sigmoid; Denotes a gating unit, which is used to evaluate the importance of each modality and perform weighting according to its importance; Denotes gating parameters that control the importance scores of each modality; Denotes the weight coefficient of the Denotes the feature transformation function of the Denotes a bias term; Denotes the parameter set of the feature transformation function Denotes a scoring function, which is used to evaluate the importance of each modality; Denotes the feature representation of the Denotes the number of different modalities.

[0127] Among them, the attention fusion and compression layer can obtain more comprehensive information by effectively fusing the features of different data sources, enabling the model to comprehensively utilize this information to make more accurate predictions for the surrounding rock classification and recognition tasks; in addition, the attention fusion and compression layer also allows the model to dynamically adjust the weights of each modality according to the characteristics of the input data, thereby improving the adaptability and robustness of the model; furthermore, the compressed feature representation can also reduce the computational burden of the model to enhance the speed and stability of model training.

[0128] Specifically, the processing process of the attention aggregator for surrounding rock data is described as follows:

[0129] First, the input feature mapping layer performs L2 normalization on the received surrounding rock data features of different modalities to ensure scale consistency and avoid a certain modality dominating subsequent analysis. Secondly, multi-perspective feature extraction uses the multi-head self-attention mechanism to capture features in the surrounding rock data from multiple angles to form a more comprehensive and rich feature representation. Thirdly, the context-aware attention mechanism uses a geological background encoder (such as a model based on GRU / LSTM) to capture geological changes during the drilling process and dynamically adjust the attention weights according to the geological background to ensure attention to key geological interfaces. Then, the cross-modal interaction attention mechanism models the complex relationships between different types of geological data through GNNs to achieve effective information exchange, and at the same time applies a feature selection algorithm to remove redundant information. Finally, the attention fusion and compression layer fuses and compresses the attention weights from multiple sources into a compact representation form to provide support for the accurate classification of surrounding rock data.

[0130] In summary, based on the above components, the attention aggregator can not only enhance the model's ability to process different types of surrounding rock data, but also closely combine geological background information to improve the accuracy and reliability of the model's recognition of surrounding rock data.

[0131] Furthermore, as Figure 7 shown, the classification head includes an input adaptation layer, a fully connected layer, a feature selection and compression layer, and an output layer connected in sequence. Among them, the input adaptation layer is used to receive the compact feature representation from the attention aggregator and perform a linear transformation on the input features to adjust their dimensions to match the requirements of subsequent layers. The fully connected layer is used to further compress the feature space to capture higher-level abstract features. The feature selection and compression layer is used to select the most representative features from the output of the fully connected layer to reduce redundant information. The output layer is used to output the corresponding classification results according to the task requirements and pass them to the hierarchical decision tree for more detailed classification or regression analysis. Specifically, for multi-class classification tasks, the classification head can use the softmax activation function to output the probability distribution of each class; for binary classification tasks, the classification head can use the sigmoid activation function to output a value between 0 and 1, representing the probability of belonging to a certain class.

[0132] The hierarchical decision tree is represented as follows:

[0133] ;

[0134] In the formula, represents the hierarchical decision tree model, represents the input feature matrix, represents the target label vector, represents the feature importance evaluation function, represents the process of selecting the optimal splitting feature and threshold, represents the growth process of the tree, including recursive splitting and pruning, represents the ensemble learning method, such as random forest or gradient boosting decision tree, represents the features obtained from the cross-domain attention network weights of, represents the reduction in Gini impurity after splitting using this feature, which is used to measure the contribution of the feature to the classification task.

[0135] Among them, the above-mentioned hierarchical decision tree combines the weights output by the cross-domain attention network (CDAN) and the reduction in Gini impurity To calculate the importance score of features, ensuring that the model can identify the most discriminative features. This adaptive method enables the model to dynamically adjust the weights of each feature during the training process, thus better reflecting their actual impact on the classification task. In addition, by selecting the feature and its threshold that maximize the feature importance score at each node as the splitting criterion, this hierarchical decision tree can ensure that each split can maximize the separation of samples of different classes, thereby improving the classification accuracy of the surrounding rock while drilling. Further, the hierarchical decision tree uses a cross-domain attention network to capture the interrelationships between different data sources, enabling the construction of a more comprehensive representation of the surrounding rock features while drilling. This not only enhances the expressiveness of a single decision tree but also provides an additional information source for the entire model, enabling the model to handle more complex data patterns.

[0136] Furthermore, the output module of the above initial model, such as Figure 8 shown, the output module includes a fully connected layer, an activation function (Softmax), and a classification decision layer. Among them, the fully connected layer includes 5 neurons, each neuron representing a type of surrounding rock while drilling. The activation function is used to output the probability distribution of the surrounding rock types while drilling, and the classification decision layer is used to select the class with the highest probability as the classification result according to the probability vector output by the activation function. Among them, the classification information output by the output module is specifically:

[0137] , ; where:

[0138] , ;

[0139] In the formula, represents the classification information, which is the class label output by the surrounding rock classification and recognition model while drilling; represents the number of class labels, represents the enhanced score of each class label obtained after the multi-class confidence weighting and context-dependent classification steps of represents the classification confidence score, indicating the confidence level of the model for , and taking the maximum value in as the classification confidence score; represents the sum of the probabilities of all classes in the adjacent class set , is the probability corresponding to an adjacent class , represents the th adjacent class set related to the represents the adjustment parameter, which is used to balance the direct probability and the influence of related classes; represents another adjustment parameter for balancing the original score and context information weights.

[0140] It should be noted that in the classification decision layer shown above, by considering the probabilities of neighboring classes , additional context information can be introduced, enabling the classification decision to not only rely on the probability of a single class but also comprehensively consider the influences of multiple related classes. This helps the model capture more subtle pattern changes, thereby improving the accuracy and reliability of model classification. In addition C provides a quantitative metric to represent the confidence level of the model in the classification result and helps decision-makers quickly judge the quality of model predictions, thus enabling better assessment of potential risks.

[0141] S2. Use the surrounding rock data to train the initial model; among them, the dynamic preprocessor is used to preprocess the surrounding rock data with an optimized processing strategy, and the contrast learning engine is used to learn the feature representations for distinguishing different types of surrounding rock during drilling in the case where the surrounding rock data has no class labels.

[0142] S3. Until the initial model meets the preset training end condition, determine the initial model that meets the training end condition as the classification and recognition model for surrounding rock during drilling.

[0143] Among them, the above classification and recognition model for surrounding rock during drilling can be trained through the following steps

[0144] S21: Collect surrounding rock data of different types of surrounding rock during drilling and divide it into a training set and a validation set according to a ratio. For example, the division ratio can be 7:3. Among them, the types of surrounding rock during drilling include, for example, sandstone, shale, and limestone

[0145] S22: Set training parameters. For example, the training batch size is 32, the number of training epochs is set to 100, and the learning rate is set to 0.001. Use the training set to train the classification and recognition model for surrounding rock during drilling until the maximum number of training times is reached

[0146] S23: Use the validation set to validate the classification and recognition model for surrounding rock during drilling, and use accuracy, recall rate, and F1 score as evaluation metrics to evaluate the performance of the model. When each metric reaches 0.9 or above, the model validation passes; otherwise, adjust the training parameters (for example, adjust the number of training epochs to 200 and adjust the learning rate to 0.05) or expand the training samples and retrain the model until the model validation passes

[0147] S4. Input the obtained surrounding rock data during drilling to be classified into the classification and recognition model for surrounding rock during drilling for processing to obtain the classification information corresponding to the surrounding rock data during drilling to be classified

[0148] The following uses examples to elaborate in detail on the solutions in this application. Different sample data was collected on-site, specifically including:

[0149] Sample 1:

[0150] Well logging data: Resistivity logging: 2.7 ohm·m, Density logging: 2.66 g / cm 3 ,

[0151] Natural gamma ray logging: 32 API;

[0152] Core sample analysis data: Porosity: 18%, Mineral composition: Quartz 82%, Feldspar 13%, Mica 5%;

[0153] Seismic reflection data: Reflection amplitude: 0.72, Phase: 115°;

[0154] Surrounding rock thickness and depth data: Roof depth: 1250 m, Layer thickness: 32 m;

[0155] Geochemical data: Carbon isotope ratio: δ 13 C = -26‰, Oxygen isotope ratio: δ 18 O = -9‰, Trace element concentration: Sr = 42 ppm, Ba = 155 ppm.

[0156] Sample 2:

[0157] Well logging data: Resistivity logging: 0.9 ohm·m, Density logging: 2.72 g / cm 3 ,

[0158] Natural gamma ray logging: 88 API;

[0159] Core sample analysis data: Porosity: 6%, Mineral composition: Clay minerals 62%, Quartz 28%, Calcite 10%;

[0160] Seismic reflection data: Reflection amplitude: 0.31, Phase: 265°;

[0161] Surrounding rock thickness and depth data: Roof depth: 1520 m, Layer thickness: 16 m;

[0162] Geochemical data: Carbon isotope ratio: δ 13 C = -27‰, Oxygen isotope ratio: δ 18 O = -11‰, Trace element concentration: Sr = 118 ppm, Ba = 295 ppm.

[0163] Sample 3:

[0164] Well logging data: Resistivity logging: 4.1 ohm·m, Density logging: 2.72 g / cm 3 ,

[0165] Natural gamma ray logging: 52 API;

[0166] Core sample analysis data: Porosity: 11%, Mineral composition: Calcite 94%, Dolomite 6%;

[0167] Seismic reflection data: Reflection amplitude: 0.88, Phase: 50°;

[0168] Surrounding rock thickness and depth data: Roof depth: 1820 m, Layer thickness: 48 m;

[0169] Geochemical data: Carbon isotope ratio: δ 13 C = -4‰, Oxygen isotope ratio: δ 18 O = +3‰, Trace element concentration: Sr = 22 ppm, Ba = 85 ppm.

[0170] Based on the trained while-drilling surrounding rock classification and recognition model of this application, the prediction results shown in Table 1 can be obtained:

[0171] Table 1

[0172] Sample Number Sandstone Probability Shale Probability Limestone Probability Granite Probability Basalt Probability Final Classification Sample 1 85% 10% 3% 1% 1% Sandstone Sample 2 5% 85% 7% 1% 2% Shale Sample 3 2% 5% 88% 3% 2% Limestone

[0173] As can be seen from Table 1, the while-drilling surrounding rock classification and recognition model provided by this application calculates the probability of each sample belonging to different while-drilling surrounding rock types by analyzing the input while-drilling surrounding rock characteristic data, and makes a final classification decision based on the principle of maximum probability. After laboratory tests, the experimental test results of Sample 1, Sample 2 and Sample 3 are consistent with the final prediction results given by this application, indicating that the technical solution provided by this application has significant advantages in improving the accuracy, consistency and efficiency of while-drilling surrounding rock classification.

[0174] It should be noted that by accurately identifying and classifying the surrounding rock while drilling, this application can achieve multiple functions and significant benefits, including: First, it can greatly shorten the time required for traditional manual analysis, making the exploration process more efficient. By quickly and accurately classifying the surrounding rock while drilling, it can immediately provide support for drilling decisions, reduce unnecessary drilling work, and improve the speed of resource development. Second, by using machine learning algorithms to process complex geological data, including multi-source information such as well logging, core sample analysis, seismic reflection, thickness and depth, and geochemistry, patterns and relationships that are difficult to detect by the human eye can be discovered, thereby achieving more accurate classification of the surrounding rock and improving the accuracy of understanding the underground structure. Third, more efficient exploration means fewer ineffective boreholes, which helps to reduce interference with the natural environment. In addition, accurate classification results can help better plan mining activities, optimize resource allocation, and protect the ecological environment.

[0175] In summary, the method for classifying and identifying the surrounding rock while drilling provided by this application not only revolutionizes the traditional geological exploration method, but also lays a solid foundation for the digital transformation of the resource exploration industry, providing strong technical support for realizing intelligent, green, and efficient modern resource development.

[0176] Embodiment 2: This application provides a device for classifying and identifying the surrounding rock while drilling, which is applied to a method for classifying and identifying the surrounding rock while drilling in Embodiment 1, as Figure 9 shown, including:

[0177] A model construction module, which is used to collect surrounding rock data of different surrounding rock categories while drilling and establish an initial model introducing a dynamic preprocessor and a contrast learning engine. The surrounding rock data includes well logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemistry data;

[0178] A model training module, which is used to train the initial model using the surrounding rock data; among them, the dynamic preprocessor is used to preprocess the surrounding rock data with an optimized processing strategy, and the contrast learning engine is used to learn the feature representations for distinguishing different surrounding rock categories while drilling in the case where the surrounding rock data has no category labels;

[0179] A training end module, which is used to determine the initial model that meets the preset training end condition as the surrounding rock classification and identification model while drilling until the initial model meets the preset training end condition;

[0180] A classification and identification module, which is used to input the obtained surrounding rock data to be classified while drilling into the surrounding rock classification and identification model for processing to obtain the classification information corresponding to the surrounding rock data to be classified.

[0181] Embodiment 3: An embodiment of the present application provides a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the method in Embodiment 1.

[0182] Embodiment 4: An embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method in Embodiment 1 is implemented.

[0183] The above specific implementation manners further elaborate on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are only specific implementation manners of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for classification and identification of surrounding rocks while drilling, characterized in that: The specific steps include: Collect surrounding rock data of different types of surrounding rock while drilling, and establish an initial model that introduces a dynamic preprocessor and a comparative learning engine. The surrounding rock data includes well logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data; The initial model is trained using the surrounding rock data; wherein the dynamic preprocessor is used to preprocess the surrounding rock data with a preferred processing strategy, and the comparative learning engine is used to learn the feature representations that distinguish different types of surrounding rock while drilling when the surrounding rock data has no category labels; the dynamic preprocessor is used to select the preferred processing strategy according to the data characteristics and preprocess the integrated data, and the preprocessing includes feature extraction processing and noise reduction processing, and the output of the dynamic preprocessor is specifically: Where Y represents the output of the dynamic preprocessor; X represents the original data matrix; MME(X) represents the multimodal perceptron, which intelligently selects the data enhancement strategy suitable for the current sample set according to the statistical characteristics of the input data. The data enhancement strategy includes rotation adjustment, scaling adjustment, and brightness adjustment; θ is the adaptive enhancement parameter; AAU(·,θ) represents the adaptive unit; w represents the feature weight vector; FSR(·,w) represents feature selection and recombination; represents the adversarial deep autoencoder; φ represents the nonlinear activation function in the autoencoder, and Standardize(·) represents the standardization process; Until the initial model meets a preset training end condition, the initial model that meets the training end condition is determined as a surrounding rock classification and identification model while drilling; The acquired drilling-while-rock data to be classified is input into the drilling-while-rock classification and identification model for processing to obtain classification information corresponding to the drilling-while-rock data to be classified.

2. The method for classification and identification of surrounding rocks while drilling according to claim 1, characterized in that: The initial model includes an input and preprocessing module, a self-supervised learning module, a meta-learning module, a feature fusion and classification module and an output module. The preprocessing module includes a dynamic preprocessor, and the self-supervised learning module includes a contrastive learning engine.

3. A method for classification and identification of surrounding rocks while drilling according to claim 2, characterized in that: The input and preprocessing module also includes a multimodal sensor, wherein: The multimodal sensor is used to receive and integrate surrounding rock data or surrounding rock data to be classified while drilling. The output of the multimodal sensor is specifically: Where y represents the output of the multimodal perceptron; g(·) represents the nonlinear activation function; W represents the weight matrix; b is the bias vector; z multi is the multimodal feature representation of the data, representing a single modal feature z i The result of weighted summation; w i Represents the modal characteristic z i The importance weight of M represents the total number of modes; x i represents the original input data of the i-th mode; E i (·) is the encoder function corresponding to the i-th mode.

4. The method for classification and identification of surrounding rocks while drilling according to claim 2, characterized in that: The contrastive learning engine learns the feature representations that distinguish different types of surrounding rock while drilling without class labels in the surrounding rock data. This is achieved through the following steps: The input data is randomly and reasonably transformed to form positive sample pairs and negative sample pairs. The random and reasonable transformation is used to generate rich learning signals from different observations of the same data point through random and reasonable transformation, so as to help the model learn a representation with better generalization ability. The positive sample pair includes a first enhanced sample and a second enhanced sample obtained by different transformation methods, and the negative sample pair includes a third sample and the first enhanced sample or the second enhanced sample. The original data in the positive sample pair and the original data in the negative sample pair are surrounding rock data of different types of surrounding rock while drilling. The preset contrast loss function is used to calculate the contrast loss between positive sample pairs and negative sample pairs, so as to learn the feature representations that distinguish different types of surrounding rock while drilling when the surrounding rock data has no category labels.

5. A method for classification and identification of surrounding rocks while drilling according to claim 4, characterized in that: The contrast loss function is specifically: L final =αL multi +(1-α)L local , among which: L local =tr(Z T LZ),L=D-A; Where, L final represents contrast loss, α represents hyperparameter; L multi represents the multi-view contrast loss, L local represents the local structure preservation term; τ represents the temperature parameter, sim(·) represents the cosine similarity function, V represents the number of data sources or perspectives, and v represents a single perspective or data source, including well logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data; represents a positive sample pair, Represents negative sample pairs, N represents the batch size; A represents the adjacency matrix, D represents the degree matrix, Z represents the feature representation matrix of all samples after passing through the encoder, and T represents the matrix transpose.

6. The method for classification and identification of surrounding rocks while drilling according to claim 2, characterized in that: The feature fusion and classification module is used to effectively fuse the features of different data sources and compress them into a compact representation, specifically: in: In the formula, Z represents the fused representation, σ represents the activation function, G(A i ; γ i ) represents a gate control unit; γ i represents the gating parameter; w i represents the weight coefficient of the i-th mode; f(A i ; φ i ) represents the characteristic transfer function of the i-th mode; b represents the bias term; φ i Denotes the feature conversion function f(A i ; φ i ) parameter set; g(A j ) represents the scoring function; A j represents the feature representation of the i-th mode, and n represents the number of different modes.

7. The method for classification and identification of surrounding rocks while drilling according to claim 2, characterized in that: The output module is used to output the classification information, and the classification information is specifically: in: s' i s i +β·c i ,s i (p i +α∑ i∈N(i) p j 100. In the formula, represents classification information, which is the category label output by the surrounding rock classification and identification model while drilling; K represents the number of category labels, s' i represents the enhanced score of each category label i after the multi-category confidence weighting and context-dependent classification steps, C represents the classification confidence score, α∑ i∈N(i) p j represents the sum of all category probabilities in the adjacent category set N(i), p j is the probability corresponding to an adjacent category j, N(i) represents the set of adjacent categories related to the i-th category, and α represents the adjustment parameter used to balance the direct probability p i and the influence of related categories; β represents another adjustment parameter used to balance the original score s i and context information c i The weight of .

8. A device for classifying and identifying surrounding rocks while drilling, applied to a method for classifying and identifying surrounding rocks while drilling according to any one of claims 1 to 7, characterized in that: include: A model building module is used to collect surrounding rock data of different types of surrounding rock while drilling, and to establish an initial model that introduces a dynamic preprocessor and a comparative learning engine. The surrounding rock data includes well logging data, core sample analysis data, seismic reflection data, surrounding rock thickness, depth data, and geochemical data; The model training module is used to train the initial model using the surrounding rock data; wherein the dynamic preprocessor is used to preprocess the surrounding rock data with a preferred processing strategy, and the comparative learning engine is used to learn the feature representations that distinguish different types of surrounding rock while drilling when the surrounding rock data has no category labels; the dynamic preprocessor is used to select the preferred processing strategy according to the data characteristics and preprocess the integrated data, and the preprocessing includes feature extraction processing and noise reduction processing. The output of the dynamic preprocessor is specifically: Where Y represents the output of the dynamic preprocessor; X represents the original data matrix; MME(X) represents the multimodal perceptron, which intelligently selects the data enhancement strategy suitable for the current sample set according to the statistical characteristics of the input data. The data enhancement strategy includes rotation adjustment, scaling adjustment, and brightness adjustment; θ is the adaptive enhancement parameter; AAU(·,θ) represents the adaptive unit; w represents the feature weight vector; FSR(·,w) represents feature selection and recombination; represents the adversarial deep autoencoder; φ represents the nonlinear activation function in the autoencoder, and Standardize(·) represents the standardization process; A training end module, used for determining the initial model satisfying a preset training end condition as a surrounding rock classification and identification model while drilling until the initial model satisfies the preset training end condition; The classification and identification module is used to input the acquired drilling surrounding rock data to be classified into the drilling surrounding rock classification and identification model for processing, so as to obtain classification information corresponding to the drilling surrounding rock data to be classified.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which enable a computer to execute the method of any one of claims 1-7.

10. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the method according to any one of claims 1 to 7 is implemented when the processor executes the computer program.

Images