A method and system for logging a formation during a field survey

Abstract

This invention relates to the field of geological exploration technology, specifically to a method and system for field stratigraphic logging. The method includes the following steps: using a deep learning image segmentation model to identify stratigraphic boundaries in core images; extracting stratigraphic interfaces from well logging curves through wavelet transform and multi-scale analysis; analyzing the comprehensive characteristics of stratigraphic boundaries based on a multi-source feature fusion strategy using an attention mechanism and an autoencoder, combining the stratigraphic boundaries, stratigraphic interfaces, and seismic facies feature vectors; obtaining stratigraphic attribute map features by associating stratigraphic reflectance layers and lithological classification feature vectors from radar images according to a graph neural network stratigraphic attribute map construction strategy; and generating a stratigraphic logging report by constructing a comprehensive decision model combining a random forest classifier and a rule engine, combining geomorphic boundaries, the comprehensive characteristics of the stratigraphic boundaries, and the stratigraphic attribute map features. This invention improves the accuracy, objectivity, and efficiency of logging results through collaborative processing of multi-source data.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to a method and system for field exploration stratigraphy. Background Technology

[0002] Stratigraphic logging is a core component of geological exploration, and its results directly determine the accuracy of assessments of subsurface stratigraphic structure and lithological distribution. Current technologies suffer from the following key problems:

[0003] The traditional logging method is highly subjective: it relies heavily on technicians visually identifying core layers and judging lithology. This method is susceptible to differences in experience and visual fatigue. In practice, different technicians may produce significantly different results in identifying the same core sample, severely affecting the consistency of the logging results. Taking an oil and gas field exploration project as an example, manual logging of a single borehole can take 2-3 days. As the exploration workload increases, the long hours and high intensity of work further reduce the accuracy of judgments. More importantly, manual identification struggles to capture the microstructural features of core samples, making it prone to misjudgment in complex geological environments.

[0004] Limitations of a single data source: Existing methods often rely on single data sources for logging work. Core sampling has a point discontinuity, typically taking a sample every 10-20m. This discontinuous sampling method makes it difficult to reflect the true continuity of the strata. In complex structural areas, such as fault and fold zones, well logging curves are easily distorted and cannot accurately reflect the actual situation of the strata. When radar and seismic data are used alone, they lack effective correlation with lithological properties and can only provide general structural information of the strata, making it difficult to reveal in depth the lithological characteristics. This results in a "fragmented" state of logging results, making it impossible to form a complete and reliable stratigraphic model.

[0005] Poor technical adaptability: Current multi-source fusion technologies mostly remain at the level of simple data overlay, such as directly mapping well logging interfaces to seismic data. This approach does not fully consider the special properties of geological data. Core images suffer from blurred layer edges, making it difficult for traditional image processing algorithms to accurately identify boundaries. Radar waves experience significant attenuation in highly conductive strata, causing radar data to fail to accurately reflect stratigraphic information. Existing technologies lack customized processing strategies for these geological data characteristics, significantly reducing the reliability of the results after multi-source data fusion.

[0006] The decision-making mechanism is too simplistic: the current logging decision-making process relies heavily on a single model, such as using only clustering algorithms to classify stratigraphic data. This approach does not incorporate professional geological knowledge and rules, such as the important geological laws that terrace boundaries correspond to sedimentary discontinuities. In practical applications, the decision results of a single model often contradict common geological knowledge, leading to deviations in logging conclusions and failing to provide a scientific and accurate basis for subsequent resource exploration, engineering construction, and other work.

[0007] Therefore, there is an urgent need for a stratigraphic logging algorithm that can avoid the limitations of a single data source, reduce human intervention, and adapt to the characteristics of geological data, so as to improve the objectivity and efficiency of logging. Summary of the Invention

[0008] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides a field exploration stratigraphic logging method, comprising the following steps:

[0009] A deep learning image segmentation model is used to identify stratigraphic boundaries in core images; stratigraphic interfaces in well logging curves are extracted through wavelet transform and multi-scale analysis; based on an attention mechanism and an autoencoder multi-source feature fusion strategy, the comprehensive features of stratigraphic boundaries are analyzed by combining the stratigraphic boundaries, stratigraphic interfaces, and seismic facies feature vectors; according to the stratigraphic attribute map construction strategy of graph neural networks, stratigraphic reflection layers and lithological classification feature vectors in radar images are associated to obtain stratigraphic attribute map features; a comprehensive decision model combining a random forest classifier and a rule engine is constructed, and a stratigraphic logging report is generated by combining geomorphic boundaries, the comprehensive features of stratigraphic boundaries, and the features of the stratigraphic attribute map.

[0010] Optionally, the step of using a deep learning image segmentation model to identify the layer boundaries in the core image includes the following steps:

[0011] This invention utilizes a U-Net-based semantic segmentation network constructed using deep learning; combines Dice loss and cross-entropy loss to form a hybrid loss function; and applies morphological operations to the output pixel-level segmentation results to obtain pixel-level stratigraphic boundary coordinates. This invention achieves pixel-level identification of core stratigraphic layers through the U-Net semantic segmentation network, overcoming the limitations of manual visual identification; the hybrid loss function, combining Dice and cross-entropy, effectively solves the problem of unevenness in the stratigraphic regions of core samples, improving segmentation accuracy; morphological operations optimize the segmentation results, removing noise and connecting fracture boundaries, ensuring reliable boundary coordinates. Overall, this invention automates core stratigraphic identification, significantly reducing the error rate, improving processing efficiency, and minimizing the influence of human subjectivity and fatigue, providing precise point-scale basis for subsequent comprehensive stratigraphic boundary judgment and laying the foundation for multi-source data fusion.

[0012] Optionally, the step of extracting formation interfaces from well logging curves through wavelet transform and multi-scale analysis includes the following steps:

[0013] The original well logging curves are denoised; after multi-scale decomposition, high-frequency detail coefficients and low-frequency approximation coefficients at different scales are obtained; the modulus maxima of the high-frequency detail coefficients are used as the criterion for identifying potential formation interfaces to extract formation interfaces from the well logging curves. This invention first denoises the original well logging curves to remove instrument jitter and electromagnetic interference, ensuring data reliability; then, through multi-scale decomposition, it obtains high-frequency detail coefficients and low-frequency approximation coefficients at different depth resolutions, covering thin and thick formations, compensating for the limitations of discontinuities in core points, and achieving continuous interface extraction throughout the borehole; the high-frequency modulus maxima are used to identify potential interfaces, avoiding misidentification of thin interlayers. This not only provides a continuous linear-scale basis for subsequent multi-source feature fusion but also improves the accuracy of formation interface identification, reduces human subjectivity, and helps build a complete formation model, laying a key foundation for the objectivity and efficiency of formation logging.

[0014] Optionally, the multi-source feature fusion strategy based on attention mechanism and autoencoder, combining the stratigraphic boundary, the stratigraphic interface, and seismic facies feature vectors to analyze the comprehensive characteristics of the stratigraphic boundary, includes the following steps:

[0015] This invention constructs a feature correlation matrix to transform the dependencies between multi-source data into numerical weights; based on a multilayer perceptron, the weights are adaptively adjusted; and an encoder and decoder are designed to analyze the comprehensive characteristics of stratigraphic boundaries. By constructing a feature correlation matrix, the dependencies between multi-source data are transformed into numerical weights, avoiding subjective weight allocation; the multilayer perceptron adaptively adjusts the weights, dynamically adapting to differences in accuracy and changes in correlation among different data, overcoming the limitations of fixed weights; the encoder and decoder deepen feature extraction and redundancy removal, accurately analyzing the comprehensive characteristics of stratigraphic boundaries. Overall, this invention effectively solves the problem of multi-source feature conflicts, significantly improves the reliability and accuracy of comprehensive features, strengthens the synergistic effect of point-line-volume data, provides a more reliable core basis for subsequent stratigraphic logging and comprehensive decision-making, and further ensures the objectivity and efficiency of logging.

[0016] Optionally, analyzing the seismic facies feature vector includes the following steps:

[0017] Data augmentation processing is performed on the seismic reflection profile data; a deep neural network model for seismic data feature extraction is constructed, and the seismic phase feature vector is analyzed by combining the deep neural network model and the augmented data.

[0018] Optionally, the step of constructing a stratigraphic attribute map based on a graph neural network, associating stratigraphic reflectance layers and lithological classification feature vectors in radar images to obtain stratigraphic attribute map features, includes the following steps:

[0019] This invention structures radar reflectance layers and lithological classification features into a graph model. Based on a graph neural network (GNN), an information transmission mechanism is constructed. Through multi-layer message aggregation and updating, the spatial distribution patterns and attribute evolution trends of complex geological bodies are captured, resulting in stratigraphic attribute map features. This invention solves the problem of significant differences in the forms of the two types of data and the difficulty in directly associating them, by structuring radar reflectance layers (spatial structure data) and lithological classification features (attribute data) into a graph model, laying the foundation for the fusion of spatial and attribute information. By using a GNN to construct an information transmission mechanism, and through multi-layer message aggregation and updating, the spatial distribution patterns (such as lateral extension range) and attribute evolution trends (such as lithological gradation characteristics) of geological bodies (such as sandstone bodies and shale layers) can be accurately captured, balancing spatial correlation and attribute continuity. The resulting stratigraphic attribute map features provide complete stratigraphic information at both the surface scale and attribute levels for subsequent comprehensive decision-making, significantly improving the spatial accuracy and attribute completeness of stratigraphic logging, and ensuring that the logging results conform to geological reality.

[0020] Optionally, identifying the formation reflective layer includes the following steps:

[0021] A subsurface medium velocity model is constructed; based on the subsurface medium velocity model, the received wave field is propagated in reverse using the acoustic wave equation to generate an offset profile; the Canny edge detection algorithm is used to extract the contour of the reflection layer of the profile.

[0022] Extracting the lithological classification feature vector includes the following steps:

[0023] Principal component analysis is used to reduce the dimensionality of lithological features; cluster analysis is performed based on the dimensionality-reduced feature vectors; and the center coordinates of each cluster are combined with the principal component loading matrix to generate lithological classification feature vectors.

[0024] Optionally, the step of constructing a comprehensive decision model combining a random forest classifier and a rule engine, and generating a stratigraphic logging report by combining the comprehensive features of the geomorphic boundary, the stratigraphic boundary, and the stratigraphic attribute map features, includes the following steps:

[0025] This invention integrates the comprehensive features of geomorphic boundaries, stratigraphic limits, and stratigraphic attribute maps; constructs a rule engine; and builds an uncertainty reasoning framework based on the classification probability matrix and the rule engine, using DS evidence theory. Monte Carlo simulation is then introduced to assess decision stability and generate a stratigraphic logging report. This invention integrates spatial positioning and attribute information by fusing geomorphic boundaries, comprehensive features of stratigraphic limits, and stratigraphic attribute maps, avoiding the fragmented limitations of single-feature decisions. The rule engine, combined with DS evidence theory, embeds geological prior rules to resolve conflicts between multiple features and reduce decision uncertainty. Monte Carlo simulation assesses decision stability and identifies potential risks. Ultimately, this improves the credibility and scientific rigor of stratigraphic logging reports, providing reliable evidence for mineral exploration and engineering surveys, strengthening the advantages of algorithmic data-driven and rule-constrained approaches, and compensating for the shortcomings of existing single-decision models.

[0026] Optionally, obtaining the terrain boundary includes the following steps:

[0027] Calculate curvature and slope; set dual threshold conditions for curvature and slope to initially identify the boundary; optimize the initial identification results through morphology to obtain the landform boundary.

[0028] Beneficial Effects: This invention replaces manual identification of core layers with deep learning, avoiding subjective errors caused by differences in human experience and visual fatigue, thus improving the accuracy and efficiency of layering. Wavelet transform is used to process logging curves, compensating for discontinuities in core points and achieving continuous extraction of stratigraphic interfaces throughout the borehole, improving data continuity. A multi-source feature fusion strategy integrates layer boundaries, stratigraphic interfaces, and seismic facies features, overcoming the limitations of single data sources, highlighting high-reliability features, and improving the accuracy of comprehensive stratigraphic boundary judgment. Based on graph neural networks linking radar reflectance layers and lithological features, a stratigraphic attribute map is constructed, clearly presenting the spatial distribution patterns of strata. The comprehensive decision model combines multi-dimensional features such as geomorphic boundaries, balancing data-driven approaches with geological rules, ensuring that the logging results conform to geological understanding. Overall, this invention overcomes the problems of strong subjectivity and data fragmentation in traditional logging, significantly improving the objectivity, efficiency, and reliability of logging, providing scientific and accurate stratigraphic evidence for mineral exploration and engineering surveys.

[0029] Secondly, to efficiently execute the field exploration stratigraphic logging method provided by this invention, this invention also provides a field exploration stratigraphic logging system, including a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory stores a computer program containing program instructions. The processor is configured to call the program instructions to execute the field exploration stratigraphic logging method as described in the first aspect of this invention. The field exploration stratigraphic logging system of this invention has a compact structure and stable performance, and can stably execute the field exploration stratigraphic logging method provided by this invention, further enhancing the overall applicability and practical application capability of this invention. Attached Figure Description

[0030] Figure 1 A flowchart of a field exploration stratigraphic logging method provided in an embodiment of the present invention;

[0031] Figure 2 This is a framework diagram of a field exploration stratigraphic logging system provided in an embodiment of the present invention. Detailed Implementation

[0032] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0033] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0034] Please see Figure 1 To address the above problems, this invention provides a method for in-situ exploration stratigraphic logging, such as... Figure 1 As shown, in one embodiment, the method includes the following steps:

[0035] S1. Using a deep learning image segmentation model, identify the layer boundaries in core images.

[0036] In this embodiment, a semantic segmentation network based on U-Net is constructed using deep learning. During the training phase, a large dataset of labeled core images is used as a foundation. By optimizing the cross-entropy loss function, the model can effectively capture subtle differences in texture, color, and structure among different lithologies. In practical applications, this network segments core images with pixel-level precision, assigning specific labels to typical lithologies such as granite, sandstone, and shale. Edge detection algorithms are used to enhance boundary features at lithological transitions, and morphological operations are combined to remove noise interference, accurately locating stratigraphic boundaries and providing a reliable basis for subsequent geological analysis.

[0037] Specifically, the method of using a deep learning image segmentation model to identify the layer boundaries in a core image includes the following steps:

[0038] S11. Construct a semantic segmentation network based on U-Net using deep learning.

[0039] An attention-gated module (AGM) is introduced at the encoding end. By calculating the attention weights of different regions, it focuses on key features such as layer edges, improving the ability to capture subtle stratigraphic boundaries. Residual connections are added at the decoding end to directly transfer features from shallow networks to deeper layers. This effectively alleviates the gradient vanishing and feature degradation problems during the training of deep networks, enabling the model to learn richer stratigraphic semantic information and enhancing the network's expressive power and robustness.

[0040] S12. Combining Dice loss and cross-entropy loss, a hybrid loss function is formed.

[0041] High-precision, manually annotated hierarchical boundaries are used as training labels. To address the imbalance in the number of samples of different categories in stratigraphic images, a hybrid loss function combining Dice loss and cross-entropy loss is employed. Dice loss focuses on the overlap of segmented regions, significantly improving the segmentation effect for small target regions; cross-entropy loss optimizes the classification results from a probabilistic perspective. The combination of the two makes model training more stable. The Adam optimizer is used, with a dynamic learning rate strategy. The initial learning rate is set to 0.001, and the learning rate decays by 10% every 50 training epochs. This accelerates convergence in the early stages of training and prevents the model from getting trapped in local optima in the later stages, improving training efficiency and model generalization ability.

[0042] In this embodiment, the original image is susceptible to noise interference from the camera sensor during acquisition. Gaussian filtering is used for noise reduction. By constructing a two-dimensional Gaussian kernel, the image is smoothed based on the distance weights between pixels and their neighbors, effectively suppressing random noise. To highlight lithological differences, adaptive histogram equalization is used. The image is divided into multiple sub-blocks, and the histograms of each sub-block are calculated and equalized. Contrast gain is limited to avoid noise amplification and enhance image details. Finally, the image is uniformly scaled to 512×512 pixels to meet the input requirements of subsequent models. A bicubic interpolation algorithm is used to maintain smooth image edges while minimizing information loss during scaling.

[0043] S13. Apply morphological operations to the output pixel-level segmentation results to obtain the pixel-level layer boundary coordinates.

[0044] For the output pixel-level segmentation results, morphological erosion is first applied to shrink the segmented region using a structuring element, removing isolated noise points and small mis-segmented areas. Then, dilation is performed to expand the region using a structuring element, connecting fractured stratigraphic boundaries and making the segmentation results more continuous and complete. Finally, precise pixel-level stratigraphic boundary coordinates are extracted and output, providing an accurate data foundation for subsequent stratigraphic logging.

[0045] By employing a combined strategy of attention gating, residual connectivity, and morphological post-processing, this method specifically addresses the problems of blurred layer edges (such as weathering transition zones) and texture interference (such as joints) in core images. It replaces traditional manual visual identification, achieving automated and high-precision positioning of core layers and reducing subjective errors.

[0046] S2. Formation interfaces in well logging curves are extracted through wavelet transform and multi-scale analysis.

[0047] In the embodiment, the problem of being susceptible to noise interference (such as misjudgment of thin interlayers) under a single scale is solved by multi-scale modulus maximum consistency verification and depth correction, and the formation interface is objectively extracted from continuous logging data to make up for the limitations of discontinuous core sampling.

[0048] Specifically, the extraction of formation interfaces from well logging curves through wavelet transform and multi-scale analysis includes the following steps:

[0049] S21. Noise reduction processing is performed on the original logging curves.

[0050] The Savitzky-Golay filtering algorithm is used to denoise the original logging curves. By constructing a least-squares polynomial fitting window for the data points, random noise is effectively suppressed while preserving signal characteristics. Based on data quality and signal characteristics, the window size is set to 5-11 points to specifically remove spike outliers caused by factors such as instrument vibration and electromagnetic interference, ensuring the reliability of subsequent analysis data.

[0051] S22. After multi-scale decomposition processing, the logging curves are used to obtain high-frequency detail coefficients and low-frequency approximation coefficients at different scales.

[0052] Using the db4 wavelet basis function and considering the actual geological exploration requirements, the scale range is set to 5-20, corresponding to a depth resolution of 0.5-2m. This parameter range can accurately capture the thickness variations of common stratigraphic interfaces and meet the accuracy requirements of geological stratification.

[0053] Discrete wavelet transform (DWT) is performed on the denoised logging curves to decompose the original signal into high-frequency detail coefficients and low-frequency approximation coefficients at different scales. High-frequency detail coefficients refer to the coefficients of the corresponding high-frequency subbands after wavelet decomposition, reflecting local abrupt changes and detailed features in the signal. In the logging curves, these coefficients correspond to abrupt changes in physical parameters at formation interfaces, such as step changes in resistivity and natural gamma at lithological interfaces, containing local variation information and highlighting the abrupt changes in physical properties at formation interfaces. Low-frequency approximation coefficients refer to the coefficients of the corresponding low-frequency subbands after wavelet decomposition, reflecting the overall trend and background information of the signal. In the logging curves, these coefficients correspond to the macroscopic sedimentary characteristics and gradual changes in physical properties of the formation, preserving the overall trend of the signal and reflecting the macroscopic changes in formation physical properties. Through multi-scale analysis, multi-resolution characterization of formation information is achieved.

[0054] S23. Using the modulus maxima of the high-frequency detail coefficients as the criterion for identifying potential formation interfaces, extract formation interfaces from the logging curves.

[0055] In the interface identification process, the modulus maxima of high-frequency detail coefficients are used as the criterion for identifying potential stratigraphic interfaces. The modulus maxima points correspond to abrupt changes in the signal, which can effectively indicate changes in stratigraphic lithology and physical properties. To reduce the false positive rate, a consistency verification mechanism for modulus maxima at adjacent scales is introduced: when a point has modulus maxima at ≥3 consecutive scales, the point is determined to be a real stratigraphic interface, which significantly improves the accuracy and reliability of interface identification.

[0056] Furthermore, due to depth errors caused by factors such as instrument stretching and cable creep during the logging process, it is necessary to use the depth correction data provided by the logging instrument to correct the detected formation interface. By establishing a depth correction model, the interface detection results are converted into actual formation depth coordinates to ensure that the depth information of the formation logging strictly corresponds to the actual geological profile, thus providing an accurate data basis for subsequent geological analysis.

[0057] Wavelet transform is used to perform time-frequency decomposition on non-stationary signals of well logging curves. By selecting appropriate wavelet basis functions, a multi-resolution analysis framework is constructed. In the time-frequency domain, it can accurately capture abrupt changes in physical parameters at formation interfaces, such as step increases or decreases in resistivity curves at lithological change interfaces, and slope inflection phenomena in sonic transit curves caused by porosity changes. Through multi-scale analysis, fine features of thin-layer interfaces are identified from high-frequency detail components, and large-scale formation structure changes are grasped using low-frequency approximation components, achieving full coverage of interface features at different depth resolutions. Simultaneously, wavelet coefficient thresholding technology is combined to effectively suppress noise interference, enhance the accuracy and reliability of formation interface identification, and provide a high-quality data foundation for subsequent formation stratification and lithology determination.

[0058] S3. Based on the attention mechanism and autoencoder multi-source feature fusion strategy, the comprehensive features of the stratigraphic boundary are analyzed by combining the stratigraphic boundary, the stratigraphic interface and the seismic facies feature vector.

[0059] In this embodiment, a multi-layer convolutional pooling architecture is constructed using a convolutional neural network (CNN), and the weights are initialized through transfer learning from a pre-trained ImageNet model. First, the seismic profile is preprocessed by normalization and denoising. Then, in the shallow network, 3×3 convolutional kernels are used to capture local texture details and extract basic texture features such as parallelism, divergence, and pre-product. The deep network uses dilated convolutions to expand the receptive field, identify seismic facies boundaries and sedimentary geometry, and compress the learned features through global average pooling. A Softmax classifier is then used to classify typical seismic facies such as channels, deltas, and turbidite fans. Finally, a seismic facies feature vector containing seismic facies category probabilities, texture complexity indices, and structural morphology parameters is generated, providing core seismic attribute descriptions for subsequent multi-source data fusion.

[0060] Specifically, the analysis of the seismic phase feature vector includes the following steps:

[0061] S311. Perform data augmentation processing on the seismic reflection profile data.

[0062] At the spatial domain level, a ±15° rotation transformation was performed to simulate the characteristics of seismic data under different acquisition azimuth angles; a scaling factor of 0.8–1.2 was used to adapt to the scale differences in stratigraphic structures. In terms of signal amplitude processing, a ±10% amplitude perturbation was introduced to simulate signal fluctuations caused by factors such as the heterogeneity of the geological medium and instrument noise during actual exploration. After these processing steps, the training sample size was expanded to more than three times that of the original data, significantly enhancing the diversity and richness of the dataset.

[0063] S312. Construct a deep neural network model for seismic data feature extraction, and combine the deep neural network model with the enhanced data to analyze the seismic phase feature vector.

[0064] A deep neural network model for seismic data feature extraction is constructed based on an improved ResNet-18 network architecture. The specific design is as follows:

[0065] Input layer: Receives seismic profile slice data with a resolution of 512×512 pixels. Through standardization, the pixel values ​​are mapped to the range of [-1,1], effectively eliminating the impact of data scale differences on model training.

[0066] Multi-scale convolution module: Innovatively employs a parallel structure of 1×1, 3×3, and 5×5 convolution kernels. The 1×1 kernel primarily achieves cross-channel information fusion and feature dimensionality reduction; the 3×3 kernel focuses on capturing local stratigraphic texture features; and the 5×5 kernel is used to extract larger-scale geological structural features. Through parallel computation and feature concatenation operations, efficient extraction and fusion of geological structural features at different scales are achieved.

[0067] Global average pooling layer: Replacing the traditional fully connected layer with a global average pooling layer, each feature map is converted into a single value through global averaging. This operation significantly reduces the number of model parameters and the risk of overfitting while fully preserving the overall statistical properties of the feature maps.

[0068] Furthermore, a training dataset was constructed using typical seismic facies categories (including parallel facies, lenticular facies, and chaotic facies) manually labeled by geological experts as supervised learning labels. The cross-entropy loss function was used to quantify the difference between the model's predicted probability distribution and the true labels, and the Adam optimizer was selected for updating network parameters. The initial learning rate was set to 0.0001, and a cosine annealing learning rate decay strategy was configured to dynamically adjust the learning rate during training, thereby accelerating model convergence and effectively avoiding getting trapped in local optima.

[0069] The output of the penultimate layer of the model is selected as the seismic facies feature representation, generating a 128-dimensional feature vector. This feature vector is abstracted layer by layer through the network, encoding not only seismic facies type information but also containing key texture feature weights. It can effectively characterize the geological structural features in seismic data, providing high-quality input for subsequent multi-source data feature fusion.

[0070] By using a CNN architecture with multi-scale convolution and global pooling, structural features of different scales (such as small-scale sand bodies and large-scale faults) in seismic data are captured in a targeted manner. The feature vectors are bound to the facies type to identify stratigraphic sequences and structures (such as faults and sand body distribution) at the macro scale, thus making up for the local limitations of core and well logging.

[0071] Furthermore, a feature weight allocation network is constructed using a multi-head attention mechanism. By calculating the spatiotemporal correlation between the texture features of the high-resolution core image and the resistivity and gamma values ​​of the logging curves, the weights of multi-source data are dynamically allocated. The focus is on capturing the correspondence between abrupt changes in mineral composition at core stratigraphic interfaces and inflection points in logging curve morphology, such as the color difference at the sandstone-mudstone interface and the stepwise changes in the acoustic transit time curve of the logging curve. Combining the dimensionality reduction characteristics of the autoencoder, the core scanning electron microscope data, logging array data, and ground-penetrating radar reflection wave data are compressed and encoded. Effective features are selected based on the principle of minimizing reconstruction error, and finally fused to form a comprehensive stratigraphic boundary feature vector containing multi-dimensional information such as lithology, physical properties, and electrical properties, providing a high-dimensional representation for subsequent stratigraphic interface identification. Specifically, the multi-source feature fusion strategy based on the attention mechanism and autoencoder, combined with the stratigraphic boundaries, stratigraphic interfaces, and seismic facies feature vectors, analyzes the comprehensive stratigraphic boundary features, including the following steps:

[0072] S321. Construct a feature association matrix to transform the dependencies between multi-source data into numerical weights.

[0073] To eliminate differences in dimensions and scales among different data sources, a phased standardization strategy is adopted. For spatial coordinate data (such as core sample location and radar scanning range), linear transformation is used to normalize them to the [0,1] interval, making the geographic information collected by different devices comparable. For numerical data such as well logging curves and seismic wave intensity, the Z-score standardization method is used to map them to [-1,1].

[0074] A feature correlation matrix is ​​constructed, and the dependencies between multi-source data are transformed into numerical weights by calculating indicators such as the overlap between core layer depth and well logging interface depth, and the correlation between seismic facies distribution and radar reflection signal.

[0075] For example, point mutual information is used to measure the correlation strength between features x and y, forming an initial weight matrix.

[0076] S322, based on a multilayer perceptron, adaptively adjusts weights.

[0077] An adaptive weight adjustment module is constructed based on a multilayer perceptron. The input layer receives the standardized feature vector, which undergoes a nonlinear transformation through two hidden layers (128 and 64 neurons respectively). The output layer is normalized using the Softmax function to obtain the attention weight matrix. This process minimizes the cross-entropy loss function through backpropagation, dynamically enhancing the contribution of key features.

[0078] S323. Design encoders and decoders to analyze the comprehensive characteristics of stratigraphic boundaries.

[0079] A deep neural network consisting of three convolutional layers and two fully connected layers is constructed as the encoder. The input is a weighted multi-source feature tensor (core texture feature map, well logging curve sequence, seismic facies slice). Local features are extracted through convolutional layers, and fully connected layers compress them into a 64-dimensional latent space. Dropout (probability 0.2) and Batch Normalization techniques are used to prevent overfitting and improve the model's generalization ability.

[0080] The decoder consists of fully connected layers with a symmetric structure and transposed convolutional layers, reconstructing the 64-dimensional vector into the original feature dimension. During training, mean squared error (MSE) is used as the loss function, and the network parameters are iteratively updated through the Adam optimizer (learning rate 0.001) until the reconstruction error converges to the optimum.

[0081] The output 64-dimensional feature vector represents the comprehensive features of the stratigraphic boundaries. The first 20 dimensions encode the spatial location information of the stratigraphy (such as depth and horizontal coordinates), the middle 20 dimensions store the physical properties of the rocks (density, resistivity, etc.), and the last 24 dimensions represent the tectonic background features (fault strike, fold morphology). Through mutual information initial weights + MLP dynamic adjustment + autoencoder compression, adaptive optimization of feature weights is achieved (e.g., when there is a conflict between the core and logging interfaces, the weights of features with low reliability are automatically reduced), eliminating scale differences and redundant information from multi-source data and strengthening the consistency of stratigraphic boundaries.

[0082] S4. Based on the stratigraphic attribute map construction strategy of graph neural network, associate the stratigraphic reflection layer and lithological classification feature vector in the radar image to obtain stratigraphic attribute map features.

[0083] In another embodiment, identifying the formation reflector layer includes the following steps:

[0084] S411. Construct a velocity model for underground media.

[0085] The initial velocity model is constructed based on known borehole data, which contains actual velocity information of the strata at different depths. These discrete borehole velocity data are extended to the entire exploration area using interpolation algorithms (such as Kriging interpolation) to form an initial subsurface medium velocity model. Subsequently, the model is adjusted through iterative optimization. Actual seismic data is compared with data simulated based on the initial model, and optimization algorithms (such as the conjugate gradient method) are used to continuously update the velocity model parameters, minimizing the error between the simulated and actual data, thereby obtaining a more accurate subsurface medium velocity model.

[0086] S412. Based on the underground medium velocity model, the received wave field is propagated in reverse using the acoustic wave equation to generate an offset profile.

[0087] During seismic data acquisition, direct waves are the signals that arrive at the receiving point first along the high-velocity medium at or near the Earth's surface after the seismic source excitation, and they can interfere with the identification of subsequent effective reflected waves. By analyzing the wave propagation time characteristics and setting a reasonable time threshold, direct signals arriving earlier than a specific time are separated out, thereby reducing their impact on subsequent processing. Furthermore, an adaptive Wiener filtering algorithm is used to suppress background noise. Based on the statistical characteristics of noise and signal, the filter coefficients are dynamically adjusted by calculating the minimum mean square error, which can effectively suppress random noise, improve the signal-to-noise ratio of the data while preserving the characteristics of the effective signal, and make the data more suitable for subsequent processing.

[0088] Based on the acoustic wave equation ,in, Represents a wave field. Representing the medium velocity, this involves the reverse propagation of the received wavefield. In seismic exploration, the wavefield generated by the seismic source propagates through the subsurface medium. When it encounters interfaces between different media, it undergoes reflection and refraction, and the receiving instrument records these reflected wavefields. Reverse propagation involves extrapolating the wavefield from the moment of reception back to the moment of seismic source excitation. The reverse-propagated received wavefield and the forward-propagating source wavefield are processed under imaging conditions (such as cross-correlation imaging, where the cross-correlation value of the two wavefields at each spatial location is calculated; a larger cross-correlation value indicates a higher probability of the presence of a reflecting interface at that location) to generate an offset profile. This profile more accurately reflects the true location of the subsurface reflecting interface.

[0089] In this embodiment, coherent enhancement processing is performed on the migrated profile to strengthen continuous reflection layers by calculating the similarity of signals from adjacent traces. Specifically, coherence coefficient calculation methods are employed, such as normalized coherence coefficient calculation based on a sliding time window. At each time sampling point and spatial location, the coherence coefficient of adjacent seismic trace signals within a certain time window is calculated. Regions with high coherence coefficients correspond to continuous reflection layers; by enhancing the signal intensity in these high coherence coefficient regions, the continuity and stability of the stratigraphic reflection interface are highlighted.

[0090] Furthermore, an adaptive threshold filtering method is used to retain effective reflected signals. The threshold is calculated based on the local statistical characteristics of the data. First, the offset profile is divided into multiple local regions, and the mean and standard deviation of the signal in each local region are calculated. The method is set to retain reflected signals with an amplitude higher than twice the local mean standard deviation. Signals with an amplitude lower than this threshold are considered noise or weak reflected signals and are filtered out. This enhances the effective reflected signals while further suppressing residual noise.

[0091] S413. The Canny edge detection algorithm is used to extract the contour of the reflection layer of the profile;

[0092] The Canny edge detection algorithm was used to extract the reflective layer contour of the enhanced profile. Edge extraction was performed by setting dual thresholds (0.15 and 0.35 in this embodiment). First, Gaussian filtering was applied to the enhanced profile to smooth the data and reduce the influence of noise. Then, the gradient magnitude and direction were calculated to determine possible edge points. Next, non-maximum suppression was applied to retain local maxima points along the gradient direction and remove non-edge points. Finally, dual thresholds were used for edge connection. The low threshold of 0.15 was used to detect weak edges, and the high threshold of 0.35 was used to determine strong edges. Weak edges connected to strong edges were retained, thus obtaining a complete reflective layer contour and outputting its spatial coordinates, providing accurate reflective interface information for subsequent stratigraphic logging.

[0093] This invention combines iterative optimization of velocity models with coherence enhancement to specifically address the weak reflection problem caused by radar wave attenuation in highly conductive strata (such as clay), and clarifies the stratigraphic reflection layer in complex structures (such as faults and folds), providing a basis for the analysis of underground stratigraphic continuity.

[0094] Further, the lithological classification feature vector is extracted, including the following steps:

[0095] S421. Principal component analysis is used to reduce the dimensionality of lithological characteristics.

[0096] In geological data processing, raw data often exhibits dimensional differences due to factors such as measurement methods and instrument precision. For example, core permeability and well logging resistivity have completely different numerical ranges and units, which can interfere with subsequent algorithm analysis. To standardize data scale, the Z-score standardization method is adopted. By converting the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, the influence of dimensions is effectively eliminated, making different types of data comparable and providing standardized input for subsequent model training.

[0097] It is understandable that multi-source data acquired in geological exploration often suffers from high-dimensional redundancy, such as multiple waveform parameters in seismic data and various physical property indicators in well logging curves. To reduce computational complexity while retaining key information, Principal Component Analysis (PCA) is employed. First, the eigenvalue covariance matrix is ​​calculated. Then, eigenvalues ​​and eigenvectors are obtained through eigenvalue decomposition. These eigenvalues ​​reflect the variance of the data in the corresponding direction, i.e., the information content. The top k principal components with a cumulative explained variance ≥ 90% are selected, mapping the original high-dimensional features to a low-dimensional space. This removes redundant information while retaining the main features of the data, improving operational efficiency and stability.

[0098] S422. Perform cluster analysis based on the dimensionality-reduced feature vectors.

[0099] Optimization is achieved by incorporating prior geological knowledge. For example, the principal component mean values ​​of known lithological samples (such as sandstone, shale, and limestone) are obtained from regional geological data and used as initial cluster centers. This approach fully utilizes existing geological knowledge, making the clustering process more closely resemble actual geological patterns and significantly reducing classification bias caused by random initialization.

[0100] The optimal number of clusters is determined by maximizing the silhouette coefficient: In actual strata, lithology types usually exhibit certain regularities, with 3-8 common types corresponding to typical lithologies such as sandstone, shale, limestone, and mudstone. By calculating the silhouette coefficient of data points under different cluster numbers, the degree of compactness between the sample and its own cluster and its separation from other clusters is comprehensively measured; the closer the value is to 1, the better the clustering effect. The possible number of clusters (3-8) is traversed, and the cluster number corresponding to the maximum silhouette coefficient is selected as the final result, thus achieving a more accurate lithology classification.

[0101] S423. Combine the center coordinates of each cluster with the principal component loading matrix to generate lithological classification feature vectors.

[0102] After clustering, the center coordinates of each cluster (located in the principal component space) are combined with the principal component loading matrix. The principal component loading matrix records the contribution of the original features to each principal component, i.e., the original feature weights. By combining these two, a lithology classification feature vector containing lithology type and key feature weights is generated. For example, if a cluster corresponds to sandstone lithology, the feature vector can not only identify the lithology but also reflect the weights of key geological features such as porosity and permeability, providing structured and interpretable input data for subsequent machine learning-based lithology identification and stratigraphic logging.

[0103] This invention utilizes principal component analysis (PCA) to reduce the dimensionality of lithological features. By solving for the eigenvalues ​​and eigenvectors of the covariance matrix, high-dimensional lithological data containing redundant information is mapped to a low-dimensional space. This effectively removes redundant components while retaining key features, improving subsequent processing efficiency. Subsequently, a clustering algorithm is used to calculate the similarity between samples based on the reduced eigenvectors, aggregating samples with similar lithological features into the same category. Finally, for each lithological category, its statistical features (such as mean, variance, and mode) are extracted to construct a quantitative feature vector containing lithological category identifiers and feature statistics. This transforms qualitative lithological descriptions into quantitative feature vectors, providing a standardized lithological classification data foundation for subsequent multi-source data fusion and stratigraphic logging.

[0104] Furthermore, the step of constructing a stratigraphic attribute map based on a graph neural network, and associating stratigraphic reflectance layers and lithological classification feature vectors in radar images to obtain stratigraphic attribute map features, includes the following steps:

[0105] S431. The radar reflector layer and lithological classification characteristics are structurally processed and abstracted into a graph structure model.

[0106] Based on the depth dimension, a dynamic layering strategy is employed to process the radar reflector layer. Considering the differences in strata variations under different geological environments, the radar reflector layer is layered at intervals of 1-5 meters. For each layer, a reflector layer node is constructed, whose attribute parameters not only cover basic depth information but also include key indicators such as reflection intensity, which reflects reflection characteristics, and extension length, which reflects the lateral extent of the reflector layer. These parameters are obtained by preprocessing the radar data through filtering and noise reduction, followed by wavelet transform technology to extract feature values, ensuring the accuracy and reliability of the data.

[0107] Based on spatial distribution characteristics, deep learning algorithms were used to accurately divide the lithology classification results into units. During the division process, a comprehensive judgment was made by combining high-resolution image analysis of core samples and well logging data to form lithology nodes. Each lithology node's attributes include lithology type (e.g., sandstone, shale), feature vectors obtained through spectral analysis and mineral composition detection, and precise spatial coordinates determined by a high-precision positioning system. This information provides rich and accurate foundational data for subsequent graph analysis.

[0108] S432. Based on graph neural networks, an information transmission mechanism is constructed. Through multi-layer message aggregation and update operations, the spatial distribution patterns and attribute evolution trends of complex geological bodies are captured, and the stratigraphic attribute map features are obtained.

[0109] Regarding spatial relationships, spatial association edges are established when the spatial distance between reflective layer nodes and lithological nodes is less than 2 meters. The weight of the edge is quantified based on the reciprocal of the distance, meaning the closer the distance, the greater the weight, thus highlighting the close connection between spatially adjacent nodes. In the actual calculation process, the spatial analysis function of a Geographic Information System (GIS) is used to accurately calculate the three-dimensional spatial distance between nodes, ensuring the accuracy of the spatial edge construction.

[0110] Furthermore, regarding attribute association, if the similarity between two node attributes (e.g., the correlation coefficient between reflection intensity and lithological density) exceeds a threshold of 0.6, an attribute association edge is established. The edge weight is determined by using attribute similarity as a metric. By calculating multiple metrics such as Pearson correlation coefficient and cosine similarity, the degree of similarity between node attributes is comprehensively evaluated, thereby determining the weight of the attribute edge and effectively reflecting the intrinsic connection between node attributes.

[0111] The GraphSAGE model was used for training. This model is based on inductive learning and iteratively updates the features of a node by sampling information from its neighboring nodes, satisfying the following conditions:

[0112]

[0113] in, Represents the node at the k-th iteration. The node feature vectors, This represents the weight matrix at the k-th iteration. This represents the node at the (k-1)th iteration. The node feature vectors, Represents a node The set of neighboring nodes, This represents the activation function.

[0114] During training, appropriate hyperparameters, such as the number of samples and the learning rate, are set and optimized using the stochastic gradient descent (SGD) algorithm. To improve training efficiency and the model's generalization ability, large-scale historical stratigraphic data is used for training, and model parameters are adjusted through cross-validation to ensure that the model can accurately learn the feature relationships in the stratigraphic attribute map.

[0115] After GNN training, the final feature vectors of all nodes are extracted. Simultaneously, adjacency matrix features, such as average degree and clustering coefficients, are combined with these features to reflect the graph's topological structure. Dimensionality reduction techniques, such as Principal Component Analysis (PCA), are used to fuse the node feature vectors with the adjacency matrix features, removing redundant information to ultimately form a complete stratigraphic attribute map feature vector. This feature vector not only contains stratigraphic attribute information but also implies the spatial and attribute relationships between nodes, providing comprehensive and effective data support for subsequent stratigraphic logging report generation.

[0116] S5. Construct a comprehensive decision model combining a random forest classifier and a rule engine, and generate a stratigraphic logging report by combining the comprehensive features of the geomorphic boundary, the stratigraphic boundary, and the stratigraphic attribute map features.

[0117] In this embodiment, obtaining the terrain boundary includes the following steps:

[0118] S511, Calculate curvature and slope.

[0119] A priority queue-based flooding algorithm is employed to gradually fill in elevation values ​​from the bottom of depressions outwards, effectively eliminating closed depressions. This step ensures the continuity of water flow paths in subsequent hydrological analyses and avoids calculation errors caused by terrain depressions. For example, in mountainous terrain, closed depressions may lead to water flow simulation results that are significantly inconsistent with reality; filling depressions makes the simulation results closer to the actual scenario.

[0120] A 3×3 window Gaussian filter kernel is used for smoothing. By assigning different weights to each pixel within the window, noise generated during the measurement process is suppressed. The weight allocation follows the Gaussian distribution principle, with the center pixel having the highest weight and the weights of edge pixels gradually decreasing. This effectively reduces noise interference and improves data quality while preserving terrain features.

[0121] Plane curvature reflects the degree of curvature of the terrain along contour lines, and its calculation is based on the analysis of the rate of change of slope along contour lines. In areas with greater plane curvature, the contour lines are more curved, which usually indicates more complex terrain changes and may include special landforms, such as the intersection of valleys or ridges.

[0122] Profile curvature focuses on reflecting the bending of the terrain along the direction of maximum slope and is determined by analyzing the second derivative of elevation in the direction of maximum slope. Profile curvature is of great value for studying water erosion and deposition processes and is one of the key parameters in the construction of soil erosion models.

[0123] Furthermore, based on the maximum gradient method, the slope value is obtained by calculating the partial derivatives of the elevation value in the x and y directions.

[0124] S512. Set dual threshold conditions for curvature and slope to initially identify the boundary.

[0125] Set dual threshold conditions: the slope is greater than 25° and the absolute value of the plane curvature is greater than 0.05m. -1 This process filters out potential geomorphic boundary areas. The slope threshold is set based on a large amount of measured data and experience. A slope of 25° or more usually corresponds to a relatively steep terrain change and may be the location of a geomorphic boundary. The absolute value threshold of plane curvature is used to further exclude areas with gentle terrain changes but slightly steep slopes, ensuring that the selected areas better match the characteristics of geomorphic boundaries.

[0126] S513. Obtain the landform boundary by optimizing the preliminary identification results through morphological optimization.

[0127] Morphological closure was performed on the initially identified boundaries using 3×3 structuring elements. This operation, through a process of dilation followed by erosion, effectively connects broken boundary segments, enhancing the continuity and integrity of the boundaries. When dealing with boundaries in complex terrain, the closure operation can integrate discontinuous boundary fragments caused by data noise or local terrain variations into complete boundary lines.

[0128] After the closure operation is completed, the skeleton of the boundary is extracted, simplifying it into lines that are only one pixel wide. This step preserves the core topological structure and geometric shape information of the boundary, removes redundant pixel information, facilitates subsequent boundary analysis and feature extraction, and also reduces the cost of data storage and processing.

[0129] In this embodiment, high-resolution satellite remote sensing imagery is also used to verify and correct the initially determined boundaries using image recognition technology. By comparing the features of land features in the remote sensing imagery, boundaries that are not geologically formed, such as man-made roads and buildings, are identified and eliminated. For example, man-made roads in remote sensing imagery typically have regular geometric shapes and obvious texture features, which are significantly different from natural landform boundaries. These features can be used to accurately distinguish non-geological boundaries, ensuring that the final determined landform boundaries better conform to the actual geological conditions.

[0130] Furthermore, the construction of the integrated decision model of random forest classifier and rule engine, combining the comprehensive features of landform boundaries, stratigraphic limits, and stratigraphic attribute map features, to generate a stratigraphic logging report includes the following steps:

[0131] S521. Integrate the geomorphic boundary, the comprehensive features of the stratigraphic boundary, and the features of the stratigraphic attribute map.

[0132] A feature stitching technique based on principal component analysis (PCA) for dimensionality reduction optimization was employed to deeply integrate geomorphic boundary features (40-dimensional) extracted from high-resolution remote sensing images, comprehensive boundary features (64-dimensional) generated from multi-scale segmentation, and attribute map features (24-dimensional) from a 3D geological model. Standardization was used to eliminate dimensional differences, constructing a 128-dimensional comprehensive feature vector to achieve preliminary fusion of multi-source data. Simultaneously, mutual information entropy was used to calculate feature redundancy and dynamically adjust feature weights to enhance the representativeness of the fused features.

[0133] S522. Construct a rule engine. Based on the classification probability matrix and the rule engine, construct an uncertainty reasoning framework according to the DS evidence theory, introduce Monte Carlo simulation to evaluate the stability of decision-making, and then generate a stratigraphic logging report.

[0134] The training set was constructed using stratigraphic logging data from 3,000 known boreholes in the National Geological Survey Database. The data annotations covered 28 detailed stratigraphic unit types, including Quaternary loess, Neogene sandstone, and Cambrian limestone. The dataset was divided into training and test sets in a 7:3 ratio using cross-validation.

[0135] A random forest classification model consisting of 100 decision trees was constructed, using the Gini coefficient as the node splitting criterion and an adaptive learning rate optimization algorithm. During training, a Bagging sampling strategy was used to randomly select 80% of the samples and 60% of the features for single-tree training. The model outputs the probability distribution matrix of each layer unit and simultaneously calculates the feature importance ranking to screen key influencing factors.

[0136] Furthermore, by integrating the International Union of Geological Sciences (IUGS) standard geological rule library and combining it with domestic regional geological standards, a rule library of more than 200 basic geological theories in five major categories was established, including stratigraphic sedimentary laws (such as the principle that stratigraphic sedimentation follows the lower old and upper new), geological structural characteristics (such as faults causing discontinuities in stratigraphic boundaries), and lithological combination laws.

[0137] Production rules are used to formally represent geological knowledge, constructing an IF-THEN rule network. For example, when "the comprehensive characteristics of stratigraphic boundaries show a mutation rate > 0.8 at three consecutive sampling points" and "the attribute map characteristics show a lithological density difference > 500 kg / m³", the rule network is constructed. 3When a non-integration surface determination rule is triggered, a rule priority and conflict resolution mechanism are set to ensure the accuracy of rule reasoning.

[0138] The classification probability matrix output by the random forest model is input into the rule engine, and an uncertainty reasoning framework is constructed based on DS evidence theory. For classification results that conflict with the geological rule base (such as the appearance of reverse stratigraphic sequences), the posterior probability is recalculated using a Bayesian network and corrected by combining expert experience knowledge base. Monte Carlo simulation is introduced to evaluate the stability of the decision, and finally, a stratigraphic unit division scheme with a confidence level of ≥90% is output.

[0139] Furthermore, the final stratigraphic unit division results are integrated with high-precision three-dimensional spatial coordinate information, lithological thin section analysis data, and geophysical inversion results through multi-source correlation. Using a template-driven automated workflow, a structured stratigraphic logging report is generated, including stratigraphic columnar sections, contour profiles, and lithological probability cloud maps. Simultaneously, an interactive visualization module is embedded, supporting users to dynamically query and spatially analyze stratigraphic information via a web interface. The report content complies with the requirements of the "General Rules for Solid Mineral Exploration" (GB / T13908-2020) standard.

[0140] This invention utilizes a dual-mechanism decision-making approach combining random forest data-driven and rule engine knowledge constraints. By integrating data-driven approaches with geological knowledge, it generates objective and geologically logical stratigraphic logging reports, effectively improving logging efficiency.

[0141] Please see Figure 2 In an embodiment, to efficiently execute the field exploration stratigraphic logging method provided by the present invention, the present invention also provides a field exploration stratigraphic logging system, including: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory contains program instructions for the steps of the field exploration stratigraphic logging method. The field exploration stratigraphic logging system of the present invention has a compact structure and stable performance, and can stably execute the field exploration stratigraphic logging method of the present invention, further improving the overall applicability and practical application capability of the present invention.

[0142] In this embodiment, the processor may be a central processing unit, but it can also be other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. Input devices can be used to acquire data. Output devices can be used to output the results obtained by storing program instructions contained in a computer program in the memory provided by this invention. The memory may include read-only memory and random access memory (RAM), and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory (RAM).

[0143] In one possible implementation, the memory may include a stored program area and a stored data area. The stored program area may store the operating system and applications required for at least one function; the stored data area may store data created during use. Furthermore, the memory may include read-only memory and random access memory, and provides instructions and data to the processor. The memory stores the operating system and operating instructions, executable modules, or data structures, or subsets thereof, or extended sets thereof. The operating instructions may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and handling hardware-based tasks.

[0144] The embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described field exploration stratigraphic logging method.

[0145] The storage medium can include various media that can store program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0146] In summary, this invention replaces manual identification of core layers with deep learning, avoiding subjective errors caused by differences in human experience and visual fatigue, thus improving the accuracy and efficiency of layering. Wavelet transform is used to process logging curves, compensating for discontinuities in core points and enabling continuous extraction of stratigraphic interfaces throughout the borehole, improving data continuity. A multi-source feature fusion strategy integrates layer boundaries, stratigraphic interfaces, and seismic facies features, overcoming the limitations of single data sources, highlighting high-reliability features, and improving the accuracy of comprehensive stratigraphic boundary judgment. Based on graph neural networks linking radar reflectance layers and lithological features, a stratigraphic attribute map is constructed, clearly presenting the spatial distribution patterns of strata. A comprehensive decision model combines multi-dimensional features such as geomorphic boundaries, balancing data-driven approaches with geological rules, ensuring that the logging results conform to geological understanding. Overall, this invention overcomes the problems of strong subjectivity and data fragmentation inherent in traditional logging, significantly improving the objectivity, efficiency, and reliability of logging, providing scientific and accurate stratigraphic evidence for mineral exploration and engineering surveys.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A method of logging a formation under exploration in situ, characterized in that, Includes the following steps: Using a deep learning image segmentation model, we can identify the stratification boundaries in core images. Formation interfaces in well logging curves are extracted using wavelet transform and multi-scale analysis. Based on the attention mechanism and autoencoder-based multi-source feature fusion strategy, the comprehensive features of the stratigraphic boundaries are analyzed by combining the stratigraphic boundaries, the stratigraphic interfaces, and the seismic facies feature vectors. Specifically, a feature weight allocation network is constructed using a multi-head attention mechanism. By calculating the spatiotemporal correlation between the texture features of the high-resolution core image and the resistivity and gamma values ​​of the logging curve, the weights of the multi-source data are dynamically allocated. The comprehensive feature vector of the stratigraphic boundaries contains multi-dimensional information on lithology, physical properties, and electrical properties. Based on the stratigraphic attribute map construction strategy of graph neural network, stratigraphic reflection layer and lithological classification feature vector are associated to obtain stratigraphic attribute map features; This includes the following steps: structuring the stratigraphic reflection layer and lithological classification feature vectors, abstracting them into a graph structure model; constructing an information transmission mechanism based on a graph neural network, capturing the spatial distribution patterns and attribute evolution trends of complex geological bodies through multi-layer message aggregation and update operations, and obtaining stratigraphic attribute map features; specifically, in terms of spatial association, when the spatial distance between stratigraphic reflection layer nodes and lithological nodes is less than 2 meters, a spatial association edge is established, and the edge weight is quantified based on the reciprocal of the distance; in terms of attribute association, if the similarity between the attributes of two nodes exceeds a threshold of 0.6, an attribute association edge is established, and the edge weight is measured by attribute similarity, comprehensively evaluating the similarity between node attributes, thereby... The weights of attribute edges are determined; the stratigraphic attribute map features contain stratigraphic attribute information, as well as spatial and attribute relationships between nodes; identifying the stratigraphic reflection layer includes the following steps: constructing a subsurface medium velocity model; based on the subsurface medium velocity model, using the acoustic wave equation to perform reverse propagation of the received wave field to generate an offset profile; using the Canny edge detection algorithm to extract the stratigraphic reflection layer contour of the profile; extracting the lithological classification feature vector includes the following steps: using principal component analysis to perform dimensionality reduction on the lithological features; performing cluster analysis based on the dimensionality-reduced feature vector; combining the center coordinates of each cluster with the principal component loading matrix to generate the lithological classification feature vector; A comprehensive decision-making model combining a random forest classifier and a rule engine is constructed. By combining the comprehensive features of the geomorphic boundary, the stratigraphic boundary, and the stratigraphic attribute map features, a stratigraphic logging report is generated.

2. The method of claim 1, wherein, The method of using a deep learning image segmentation model to identify the layer boundaries in core images includes the following steps: Construct a semantic segmentation network based on U-Net using deep learning; By combining Dice loss and cross-entropy loss, a hybrid loss function is formed; Morphological operations are applied to the output pixel-level segmentation results to obtain the pixel-level layer boundary coordinates.

3. The method of claim 1, wherein: The extraction of formation interfaces from well logging curves through wavelet transform and multi-scale analysis includes the following steps: The original logging curves are subjected to noise reduction processing; After multi-scale decomposition processing, the logging curves are used to obtain high-frequency detail coefficients and low-frequency approximation coefficients at different scales. The modulus maxima of high-frequency detail coefficients are used as the criteria for identifying potential formation interfaces to extract formation interfaces from well logging curves.

4. The method of claim 1, wherein: The analysis of the seismic facies feature vectors includes the following steps: Data augmentation processing is performed on seismic reflection profile data; A deep neural network model for seismic data feature extraction is constructed, and the seismic phase feature vector is analyzed by combining the deep neural network model with the enhanced data.

5. The method of claim 1, wherein: The construction of the integrated decision model combining a random forest classifier and a rule engine, which integrates the geomorphic boundary, the comprehensive features of the stratigraphic boundary, and the features of the stratigraphic attribute map, to generate a stratigraphic logging report includes the following steps: The features of the geomorphic boundary, the comprehensive features of the stratigraphic boundary, and the features of the stratigraphic attribute map are integrated; A rule engine is constructed, and based on the classification probability matrix and the rule engine, an uncertainty reasoning framework is built according to the DS evidence theory. Monte Carlo simulation is introduced to evaluate the stability of decision-making and then generate a stratigraphic logging report.

6. The method of claim 1, wherein: Obtaining the terrain boundary includes the following steps: Calculate curvature and slope; By setting dual threshold conditions for curvature and slope, the boundary is initially identified; The landform boundary is obtained by optimizing the preliminary identification results through morphological analysis.

7. A field exploration formation logging system characterized by, The field exploration stratigraphic logging system includes: an input device, an output device, a processor, and a memory, wherein the input device, the output device, the processor, and the memory are interconnected, and the memory includes program instructions for executing the field exploration stratigraphic logging method according to any one of claims 1-6.

Images