Underground geologic structure imaging system based on azimuth gamma rays

Abstract

The invention relates to an underground geological structure imaging system based on azimuth gamma rays, belongs to the technical field of underground logging and geological exploration, and aims to enhance scarce or low-quality data through a multi-azimuth detector array and multi-modal data acquisition and fusion by utilizing a Transform architecture and a GAN (Gamma Area Network) and self-supervised learning, so that the imaging accuracy of the underground geological structure is improved. And a multi-head cross attention mechanism is combined to accurately identify a complex stratum. The system supports multiple tasks of lithology classification, fault prediction, abnormal stratum alarm and the like, realizes real-time and high-resolution imaging and safety early warning, and is suitable for scenes of oil and gas drilling, underground engineering and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of downhole well logging and geological exploration, and in particular relates to an downhole geological structure imaging system based on azimuth gamma rays. Background Art

[0002] With the increasing demand for oil and gas field exploration and exploitation, how to accurately obtain formation information and identify key structures such as faults and fractures in complex underground environments has become an important research direction for current drilling and logging technologies. Azimuth gamma ray technology has been widely used in measurement while drilling and formation evaluation because it can obtain radioactive signals within a 360° circumferential range. Through azimuth gamma detection, the gamma count value can be analyzed in a directional manner in combination with the sensor attitude information, thereby obtaining more detailed results in the characterization of geological structures.

[0003] Chinese invention patent CN106285632B discloses an azimuth gamma measurement device, which acquires gamma count values ​​and probe postures by focusing gamma sensors, three-axis accelerometers, and three-axis fluxgates, and integrates the collected gamma data in different directions into the same coordinate system, thereby realizing continuous measurement and multi-posture adaptation of azimuth gamma, which can effectively determine the properties of complex strata and stratum changes, and improve the drilling rate.

[0004] Chinese invention patent CN109581522B discloses a probe-tube azimuth gamma probe for a downhole inclinometer. By setting a gamma-ray probe composed of a NaI crystal and a photomultiplier tube in the probe tube and combining it with three gravity acceleration sensors, the gamma detection direction is determined to achieve directional measurement of drill collar-type gamma data, overcoming the problems of complex traditional structure, high cost and low precision, and having certain practical value.

[0005] The above design realizes directional gamma measurement through azimuth gamma probe and attitude sensor, which meets the basic identification requirements of underground formations, but there are still certain limitations, such as multimodal data fusion, real-time high-resolution imaging and dynamic prediction of complex formations. It is difficult to fully cope with the comprehensive analysis requirements of various geological anomalies (faults, cracks, lithology transitions, etc.). At the same time, the current azimuth gamma logging system still has room for improvement in data missing completion, noise suppression, remote transmission visualization, etc. If the technology can be upgraded in the above links, it can not only greatly improve the efficiency of underground surveys, but also significantly reduce the trial and error costs and potential risks caused by incomplete information during drilling. Summary of the invention

[0006] In response to the above-mentioned existing technologies, a downhole geological structure imaging system based on azimuthal gamma rays, by integrating advanced spatiotemporal modeling, multimodal data fusion and gamma completion algorithms, aims to achieve more accurate and dynamic intelligent identification and imaging of downhole stratigraphic structures, and provide efficient and reliable solutions for geological exploration and oil and gas development.

[0007] In order to solve the above problems, the present invention provides an underground geological structure imaging system based on azimuthal gamma rays.

[0008] In summary, this application has the following beneficial effects: 1. Through the annular multi-directional gamma-ray detector array, all-round monitoring of the underground formation is achieved, and changes in radioactivity intensity from any direction can be detected in time. The dynamic signal adjustment unit can adjust the detector sensitivity in real time according to environmental factors such as wellbore eccentricity and drilling fluid influence, further reducing the interference of noise on the detection results.

[0009] 2. With the help of multimodal equipment such as resistivity, acoustic wave, temperature and pressure sensors, the system can obtain key formation information such as conductivity, density, porosity, etc. while monitoring the radioactive characteristics of the formation; the data synchronization interface can ensure that the gamma-ray data and other modal data are highly consistent in time and space, providing a solid foundation for subsequent multimodal fusion and imaging, enhancing the perception of complex formation environments, and effectively reducing data omissions or time sequence disorders during real-time acquisition and transmission, making the detection results more reliable. For underground environments with high noise levels or unstable acquisition conditions, the present invention introduces a preprocessing and denoising mechanism to greatly improve the signal-to-noise ratio of the original gamma-ray signal and multimodal data. By analyzing historical data and real-time acquisition signals, the gamma-ray attenuation or abnormal fluctuation caused by drilling fluid absorption, wellbore eccentricity, etc. can be automatically identified and corrected; combined with advanced algorithm modules such as GAN and self-supervised learning, data reliability can be further improved under conditions where data is scarce or of low quality, ensuring the accuracy and consistency of data in subsequent modeling and fusion links, and providing stable support for obtaining high-quality signals in harsh underground environments.

[0010] 3. The Transformer-based spatiotemporal modeling unit can complete multiple tasks such as lithology classification, structural analysis and fault prediction of underground strata, improving the functional integration of the system; the highly automated deep learning mechanism reduces the reliance on manual intervention. When the underground sensor data is undersampled or the resolution is not high, GAN can generate high-resolution gamma-ray feature maps to achieve data enhancement and missing completion. Self-supervised learning uses pseudo-label generation and comparative learning to significantly reduce the demand for manually labeled data, so that the model can continue to optimize in large-scale unlabeled or weakly labeled data scenarios. , adapt to various sudden acquisition interferences, and can establish deep correlations between gamma rays and multimodal data such as resistivity, acoustic waves, temperature and pressure, identify previously imperceptible fine stratigraphic features, and reduce missed or false detections through interactions and comparisons between more modes, making stratigraphic imaging more realistic and details clearer. The system maintains high precision and high robustness in multimodal and multi-task scenarios, and is particularly sensitive to the identification of key geological structures such as faults, cracks, and lithologic transition zones, reducing reliance on large-scale, high-quality manual annotations, while improving feasibility when original data is insufficient or noise interference is severe.

[0011] 4. Considering the factors of non-uniform sampling or dynamic sampling interval in time position coding makes the model more adaptable to the actual changes in downhole acquisition frequency. In the spatial position coding, adding information such as borehole inclination and eccentricity helps the system to model the data collected by the detector at different azimuths and depths in a more detailed manner. The system can predict the next moment or the next orientation based on the signal at the current azimuth or the current moment in the continuous time-space domain, and judge the fault activity or lithology boundary movement in advance. Combined with GAN data completion and self-supervision strategies, it greatly improves the prediction accuracy and real-time performance in dynamic scenarios, more completely captures the dynamic process of formation changes with time and orientation, and provides a scientific basis for real-time decision-making in the well. It can respond in time when sudden formation activities (such as micro-earthquakes and fluid mutations) may occur, and prevent safety hazards in downhole operations.

[0012] 5. The "abnormal formation alarm" function is integrated into the multi-task support module. Once abnormal signals such as faults, cracks or pressure surges are detected, the system can issue an alarm in real time. Multimodal cross-validation improves the credibility of the alarm and reduces false alarms caused by single-modal data. With 2D / 3D imaging and visualization interface, geological engineers can quickly interpret and make decisions. It supports historical data superposition analysis, which helps to track the long-term trend of underground environmental changes, form predictive and planned production strategies, improve the safety factor and real-time monitoring efficiency of underground operations, and effectively avoid construction delays or accidents caused by formation anomalies. The visualization display method meets the needs of modern operations and allows the team to intuitively grasp the overall situation underground. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is the system architecture diagram of this application; DETAILED DESCRIPTION

[0014] In order to make the objectives, technical solutions and beneficial effects of the present invention more clear, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings to facilitate understanding by technicians.

[0015] Example 1 like Figure 1 As shown, this embodiment focuses on how to build and operate a downhole geological structure imaging system based on azimuthal gamma rays, covering core units such as gamma ray detection module, multimodal data acquisition module, signal processing and correction module, intelligent modeling and multimodal fusion module, and data transmission and display module.

[0016] The system consists of several hardware and software modules, forming an integrated solution from underground data acquisition to surface visualization analysis. It can collect gamma ray and other modal sensor data in real time underground, and process, fuse and visualize them efficiently. The whole process includes: The gamma ray detection module is used to obtain the azimuthal gamma ray intensity and energy spectrum data of the underground formation to form the gamma ray original signal input. The detectors are distributed in 360° directions through a circular multi-directional gamma-ray detector array; each detector is responsible for collecting gamma-ray signals in a specific direction. Equipped with a dynamic signal conditioning unit, the detector sensitivity is adjusted in real time according to different environments (such as drilling fluid properties, wellbore eccentricity, etc.) to maintain high accuracy of signal acquisition.

[0017] In addition to gamma rays, the multimodal data acquisition module introduces multimodal sensors such as resistivity and acoustic logging to collect information such as conductivity, formation density, porosity, etc. of the downhole environment; at the same time, it monitors environmental parameters such as temperature and pressure. The data synchronization interface performs unified processing of timestamps during the acquisition process to ensure that the time dimensions of gamma-ray data and other modal data are aligned, providing a basis for subsequent multimodal fusion.

[0018] The signal processing and correction module performs preprocessing on the raw acquired data, such as normalization and denoising, to improve the signal-to-noise ratio of the gamma-ray signal. The dynamic correction mechanism is applied to reduce the errors caused by downhole environmental factors such as wellbore eccentricity and drilling fluid absorption by analyzing historical data and current real-time signals.

[0019] The intelligent modeling and multimodal fusion module adopts a core algorithm framework based on Transformer to generate high-resolution images of underground geological structures and realize multi-task functions such as lithology classification, geological structure analysis and fault prediction. It includes a multimodal feature fusion unit, a dynamic weight allocation unit and a spatiotemporal modeling unit, which respectively fuse different modal data, perform weighted scheduling and in-depth modeling at the time-space level.

[0020] The data transmission and display module uses high-speed wireless communication or fiber optic technology to transmit massive downhole collection and analysis results back to the surface in real time. It provides an intuitive visualization interface that can display 2D / 3D geological structure maps, lithology distribution maps, fault location maps, and various real-time signal trend maps; it supports overlaying historical data on the same layer for comparative analysis.

[0021] How it works The annular multi-azimuth detector array is arranged on the downhole instrument. Each detector corresponds to a different azimuth angle and collects the intensity and energy spectrum data of gamma rays in the formation in real time. The sensitivity of the detector is dynamically adjusted according to environmental disturbances to minimize the interference of noise on the results.

[0022] Simultaneously with gamma-ray detection, resistivity, acoustic logging, temperature and pressure sensors work together to obtain multi-modal signals that can reflect the characteristics of the formation. Through the data synchronization interface, the data of different time segments are aligned to achieve the synchronization of azimuthal gamma rays and other modal signals in the time and space dimensions.

[0023] In the preprocessing unit, the system normalizes, filters and denoises the collected gamma-ray data and other modal data to remove random noise caused by the external environment and the electrical equipment itself. The dynamic correction mechanism will correct the gamma-ray attenuation or abnormal fluctuation caused by factors such as wellbore eccentricity and changes in drilling fluid density based on historical and current data, making subsequent modeling and analysis more reliable.

[0024] The system uses the Transformer encoder-decoder structure to perform spatiotemporal modeling of the input gamma ray data (and other modal data): Data input: The gamma-ray signals of each azimuth detector are combined into an input sequence, supplemented with timestamp information, to form a spatiotemporal data set; Autoregressive modeling: Use the Transformer decoder part to predict the next signal based on the signal at the current time or position; Attention mechanism: The multi-head attention module automatically learns which points and time periods are more characteristic of stratigraphic changes, thereby deeply exploring key details such as lithology changes and fault fissures; Position encoding: Add position encoding to the azimuth and depth information of each detector to help the model better understand the differences in the spatial and temporal distribution of the strata; Multimodal Fusion Feature-level fusion: Map sensor data such as resistivity and acoustic waves to the same feature space as gamma rays, and form a multimodal fusion vector by splicing or weighting. Decision-level fusion: In some tasks (such as lithology classification or fault location prediction), each modality independently produces a prediction result, which is then voted or weighted averaged to finally generate the optimal prediction. Dynamic weight allocation: Based on the quality of downhole signals (e.g., the resistivity signal is weak at certain moments or the gamma ray is seriously interfered with), the system automatically adjusts the weights of different modal features in the final result; Spatiotemporal modeling unit: combines time position coding and spatial position coding to identify the dynamic change trend of the strata and output stratum structure prediction results with higher spatiotemporal resolution. Multitasking support Lithology classification: Based on gamma-ray energy spectrum and multi-modal fusion characteristics, the lithology type of the formation is determined. Geological structure analysis: Automatically identify formation dips, fractures and fault locations. Fault prediction: Through the fusion analysis of historical data and current azimuth signals, the activity trend and possible location of faults are predicted.

[0025] For ease of explanation, the following formulas are expressed using general symbols, and specific parameters can be adjusted accordingly based on actual application scenarios.

[0026] Transformer encoder-decoder structure Attention Mechanism Multi-Head Attention Expanding the above single-head attention mechanism to h parallel attention heads can be expressed as: Autoregressive Modeling In the decoder part, the subsequent time signal can be predicted based on the autoregressive strategy. For time step t, the next signal is predicted It can be expressed as: Positional Encoding In order to make the Transformer model perceive the position information of the input sequence in time and space, the following sine and cosine functions are often used for position encoding. Let the position index be pos and the dimension index be i, then the common time position encoding can be defined as: If the information such as azimuth angle α and well depth δ is introduced into the spatial position encoding, pos can also be replaced by a composite position index related to α and δ, and then the sine-cosine mapping can be performed in the same way: By combining temporal position encoding with spatial position encoding, a more complete spatiotemporal distribution embedding can be constructed for modeling the dynamic characteristics of downhole formations.

[0027] Multimodal feature fusion Feature-level fusion The multi-modal vectors after the above mapping are concatenated or weighted to obtain the fusion vector F t Decision-level fusion Dynamic weight allocation In view of the possible quality fluctuations of gamma-ray and multimodal data, the signal quality score SQ can be used. i (t) Dynamically adjust the weight of each modality in the fusion process. Let i represent the modality index, t be the time step, and define the weight update function as: Spatiotemporal Modeling Unit Combining the position encoding, attention mechanism, autoregressive prediction and multimodal fusion mentioned above, the fused feature sequence F t Input to the Transformer encoder-decoder structure: The encoder processes the input sequence in depth and outputs a spatiotemporal context vector Z t .

[0028] The decoder, in autoregressive mode, combines historical information Z≤t and attention allocation to predict the next stratigraphic signal or geological structure information. This can output stratigraphic structure prediction results with higher temporal and spatial resolution, such as lithology classification, fault location or other real-time monitoring indicators.

[0029] Image output and visualization After completing the fusion modeling of gamma-ray and other modal data, the system will dynamically generate two-dimensional or three-dimensional geological structure maps to help geologists intuitively understand the actual situation underground. If data loss or severe interference occurs during the acquisition process, AI algorithms can be used to complete or interpolate the missing data to improve the integrity and accuracy of the image. With the help of high-speed data transmission units, massive amounts of formation information and imaging results are sent to the ground operating platform in real time, realizing linkage between the well and the upper and lower wells; the visualization interface provides geological engineers with multi-level and multi-angle graphical displays, including lithology distribution maps, fault location maps, etc.

[0030] Through the gamma-ray detector array arranged in a 360° azimuth, the system can monitor the radioactive characteristics of the formation in an all-round and continuous manner, and improve accuracy by cooperating with the dynamic sensitivity adjustment unit. Multimodal data synchronous acquisition and dynamic correction mechanism significantly reduce environmental interference and improve signal credibility.

[0031] Multimodal fusion of resistivity, acoustic waves, temperature, pressure and other information with gamma-ray data enables the system to have more comprehensive data support when facing complex formations. It realizes multiple functions such as lithology classification, geological structure analysis, fault prediction, etc. on the same platform, providing a "one-stop" solution for underground decision-making and geological assessment.

[0032] With the support of high-speed wireless or fiber-optic communication technology, massive amounts of underground data can be obtained and analyzed in real time on the ground; the visualization interface can not only display two-dimensional / three-dimensional geological maps, but also overlay and compare historical data to assist geological engineers in making scientific judgments.

[0033] The modular system architecture introduced in this embodiment can easily upgrade or integrate more algorithms and hardware units (such as new sensors, enhanced AI models), leaving enough room for subsequent technological development.

[0034] In summary, Example 1 demonstrates a complete technical chain from all-round data acquisition in the well to high-resolution geological imaging on the ground through a detailed description of the gamma-ray detection module, the multi-modal data acquisition module, the signal processing and correction module, the intelligent modeling and multi-modal fusion module, and the data transmission and display module. The system can not only significantly improve the monitoring, diagnosis and prediction capabilities of the well formation, but also has high flexibility and multi-task compatibility, and has extremely high application value in the fields of oil drilling, geological exploration and underground engineering.

[0035] Embodiment 2: like Figure 1 As shown in the figure, this embodiment is based on the aforementioned embodiment 1 (i.e., the overall structure of the system and the realization of basic functions), and additionally introduces advanced algorithm modules such as generative adversarial networks (GAN) and self-supervised learning, which further enhances the depth and breadth of intelligent modeling and multimodal fusion. GAN can generate high-resolution gamma-ray feature maps to make up for insufficient sampling in the case of scarce data; at the same time, the self-supervised learning module greatly reduces the need for manual annotation through means such as pseudo-label generation, and improves the accuracy and robustness of the model under low-quality or incomplete data.

[0036] The adversarial learning module (GAN) is designed to handle the problems of insufficient data or low data resolution that are common in underground working conditions. By using the generator and discriminator of the adversarial generative network to compete with each other, the realism and resolution of the synthetic gamma-ray feature data can be continuously "improved". It complements the basic models such as the spatiotemporal Transformer and attention mechanism in Example 1, providing more stable and richer input data for subsequent multimodal fusion and multi-task support.

[0037] The self-supervised learning module is designed to address the problems that may exist in underground environments, such as insufficient manually labeled data and difficulty in obtaining real labels. Through pseudo-label generation and self-supervised target optimization, the model can automatically learn stratigraphic characteristics and association patterns without the need for large-scale manual labeling. Combined with the self-attention mechanism in the Transformer framework, more implicit spatiotemporal features can be mined, making the system more flexible and adaptable when dealing with complex situations such as low signal-to-noise ratio and geological anomalies.

[0038] Compared with Example 1, Example 2 focuses on a detailed discussion of how the above two new modules are embedded in the existing system and their impacts on improving data quality and model robustness.

[0039] Working mechanism of the adversarial learning module (GAN) The introduction of GAN usually contains two sub-networks: Generator: Based on the initial gamma-ray data or multimodal fusion vector as input, a "fake" high-resolution gamma-ray feature map is generated through a multi-layer convolution / deconvolution network or Transformer encoder. Discriminator: It is trained in opposition to the generator and is responsible for determining whether the input gamma-ray feature map is authentic from actual collected data or forged by the generator. The discriminator scores the "realism" of the generator's output, forcing the generator to continuously optimize itself and output more realistic and higher-resolution gamma-ray feature maps.

[0040] Specific process Some lower-resolution or incomplete gamma-ray data from downhole acquisition are selected as input to the generator; multimodal data (such as resistivity and acoustic logging) can be regarded as conditional vectors to help the generator better understand the formation background. The output of the generator generates a gamma-ray feature map that is "completed" or "enhanced" in the space-time dimension. This feature map surpasses traditional interpolation algorithms in terms of spatial resolution and detail fidelity. The role of the discriminator is to compare the generator output with the actual collected "real" gamma ray images and learn to distinguish between the real and the fake. If the discriminator recognition rate is high, it forces the generator to further improve the authenticity of the forged image; if the discriminator recognition rate decreases, it means that the performance of the generator has been significantly improved. Loss Function and Training Adversarial Loss is usually used, and Perceptual Loss or Reconstruction Loss can also be added to constrain the feature layer. In the iteration of mini-batch stochastic gradient descent (Mini-batch SGD) or other optimizers, the generator and discriminator continuously improve each other until the training converges, obtaining a high-quality gamma-ray feature map that can fill in missing data and enhance resolution. Adversarial Loss Let the generator be G, the discriminator be D, and the noise or conditional input distribution (such as sparse gamma-ray data, multimodal features, etc.) be z. Assume that the true distribution is p data , using adversarial objectives: Discriminator loss: Generator loss: Perceptual Loss Perceptual loss is often used to measure the difference between the generated image (or feature map) and the real image in the high-level feature space. Assuming ϕ represents a pre-trained feature extraction network (such as some convolutional layers of VGG and ResNet), then: By comparing in a high-level semantic space, the model is better able to preserve details and structural features.

[0041] Reconstruction Loss In data completion or super-resolution scenarios, reconstruction errors with the true resolution data are usually added, for example: Or using mean square error: In scenarios with low quality or a lot of missing data, the reconstruction loss can ensure that the output of the generator has sufficient pixel-level consistency with the true gamma-ray image, reducing overfitting or blurring problems.

[0042] Comprehensive generator loss In order to consider the adversarial loss, perceptual loss and reconstruction loss at the same time, they can be weighted and summed to obtain the final generator objective function: Mini-batch SGD and training process In the iteration of mini-batch stochastic gradient descent (Mini-batch SGD) or optimizers such as Adam, the discriminator and generator are updated alternately: Discriminator update: Generator Updates: where θ D ,θ G are the discriminator and generator parameters, respectively, and η is the learning rate. Through the game between the two, the realism and resolution of the generated image are continuously improved until the loss converges.

[0043] Synergy with Transformer Architecture The high-resolution gamma-ray feature map generated by GAN will be sent again to the Transformer spatiotemporal modeling unit in Example 1 for further analysis and fusion. When performing multimodal feature fusion or attention allocation, due to the "data enhancement" effect of GAN, the originally scarce or noisy data is supplemented to a certain extent, allowing the Transformer's self-attention mechanism to capture richer stratigraphic details.

[0044] GAN Data Augmentation The trained generator G will produce a completed or enhanced high-resolution gamma-ray feature map, denoted as .

[0045] Will is regarded as a new data sample, and the original multimodal data r t ,s t , and are input into the Transformer spatiotemporal modeling unit to capture richer stratigraphic details in self-attention, position encoding, and autoregressive prediction.

[0046] Role in multimodal feature fusion It can be spliced ​​with other modal features (feature-level fusion) or independently output predicted (decision-level fusion), and under the dynamic weight allocation mechanism, if the GAN output quality is high, its weight in the final stratigraphic structure imaging or fault prediction task will be appropriately increased; otherwise, its influence will be automatically reduced to ensure the robustness of the overall system.

[0047] Overall pipeline Step 1: Train GAN to obtain high-quality gamma-ray image generation capabilities; Step 2: Use the GAN output and the actual collected data to feed into the Transformer module to perform multi-head attention, position encoding and spatiotemporal modeling; Step 3: Under the multimodal fusion and self-attention mechanism, the system performs multi-task analysis on the underground formation (such as lithology classification, fault prediction, abnormal alarm, etc.) to improve accuracy and interpretability.

[0048] Through the above loss function and training process, the present invention can successfully generate reliable high-resolution feature maps in an environment with missing or noisy gamma-ray data, and is closely integrated with the Transformer spatiotemporal modeling unit, making the intelligent analysis and prediction of complex stratigraphic structures more accurate, real-time and robust.

[0049] High-resolution imaging: Greatly improves the detail and visualization of stratigraphic imaging, allowing for better identification of microstructures such as cracks and faults. Dealing with scarce data: Reduce reliance on high-density sampling to maintain system performance as much as possible under conditions such as tight drilling time, sensor failure, and complex downhole terrain. Share with other modalities: The fake data output by the generator can be fused with resistivity and acoustic wave data at the same time, so that multimodal learning can obtain more dimensional and clearer features during modeling, further improving the recognition and prediction capabilities of complex formations.

[0050] Self-supervised learning module When applied to the massive amounts of data collected underground, effective labels are often very limited; in addition, the labeling process is time-consuming and requires in-depth participation of professional geologists. There is a large amount of low-quality data (noisy, incomplete, etc.), and using it directly will weaken the performance of the model.

[0051] In unsupervised or weakly supervised scenarios, the model uses its own structure to mine “learnable” clues in the data; by completing self-prediction tasks or generating pseudo labels, it gradually approaches the learning effect of real annotations. Compared with traditional supervised learning, self-supervised learning requires no or only a small amount of manually labeled data, which greatly improves the feasibility of large-scale data processing in underground scenarios.

[0052] Based on the small amount of existing real labels or historical statistical information, preliminary "label inference" is first performed on large quantities of unlabeled gamma-ray / multimodal data. For example, the system can use the spatiotemporal features learned in the Transformer to automatically label similar signal segments with the same or similar labels. The generated pseudo labels are regarded as “soft labels”, and the model prediction results are compared with the pseudo labels to minimize a certain loss function (such as cross entropy or mean square error) to continuously update the model parameters. Through cyclic iteration, the pseudo-labels are continuously "improved" under the feedback of the model, and the model gradually "corrects" itself under the guidance of the pseudo-labels, forming a positive cycle. In some cases, we can combine the idea of ​​comparative learning to perturb the same batch of data (such as adding noise, randomly rotating the orientation, etc.) and train the model to distinguish between "homologous" and "heterogeneous" data, thereby learning more robust feature representations. This process assists GAN in data enhancement: GAN provides synthesized or completed high-quality data, and the self-supervision module generates more transformed samples for this data, making the model more tolerant to noise or interference.

[0053] The multimodal fusion unit mentioned in Example 1 can work in conjunction with self-supervised learning: when a certain modality data lacks manual labels, the corresponding pseudo-labels can be inferred using the existing "weak supervision information" of other modalities. The dynamic weight allocation mechanism (described in Example 1) can determine the weight allocation in the fusion stage according to the credibility of the pseudo-labels of each modality; if the reliability of the pseudo-labels of a certain modality is enhanced, the system will automatically increase its contribution to the final decision.

[0054] The self-supervised strategy can be pre-trained or trained in parallel on a large amount of unlabeled data, significantly reducing the need for long-term labeling by professional geological experts. Even in the face of harsh underground environment, severe noise or incomplete data collection by some sensors, the system can continue to optimize through pseudo-labeling and self-learning mechanisms, maintaining a high tolerance and recognition accuracy for abnormal situations. In actual surveys, as new data continue to enter, the system can continuously "learn-correct-relearn" through self-supervised learning strategies to achieve long-term evolutionary intelligent geological imaging.

[0055] This embodiment completely follows the hardware and core software architecture of Embodiment 1: basic functions such as the annular multi-directional gamma-ray detector, multi-modal sensor integration, signal processing and correction process, and Transformer-based intelligent modeling remain unchanged. Two new sub-modules, adversarial learning and self-supervised learning, have been added to the "Intelligent Modeling and Multimodal Fusion Module" as an upgrade to the existing algorithm framework.

[0056] Example 1 provides a robust spatiotemporal modeling mechanism, providing powerful "control data" and "structured embedding" for GAN and self-supervised learning. The high-resolution gamma-ray feature map generated by GAN can not only improve multi-task support such as fault prediction and lithology classification, but also can be deeply coupled with the dynamic weight allocation mechanism in Example 1. The pseudo-label generation and comparative learning mechanism of self-supervised learning can produce higher learning efficiency on originally sparse or noisy data sets, realizing the "adaptive evolution" of intelligent stratigraphic imaging.

[0057] This embodiment is particularly suitable for scenarios where the underground environment is dynamic and changeable, and where there are data gaps or unstable quality in sampling. It can continuously improve data availability and model reasoning effects at each stage. As the types of downhole acquisition sensors increase, GAN and self-supervision modules can also be extended to more modal data, thereby further improving the overall imaging accuracy and geological analysis level.

[0058] The Generative Adversarial Network (GAN) module enables the system to obtain “high-simulation, high-resolution” gamma-ray feature maps even when data is scarce or resolution is insufficient, greatly improving the imaging details of the formation and the ability to detect abnormal structures.

[0059] The self-supervised learning module uses pseudo-labels and contrastive learning mechanisms to greatly alleviate the pressure of manual labeling, enabling the system to have strong self-iteration capabilities and robustness to environmental noise.

[0060] Compared with Example 1, Example 2 provides a leap-forward "intelligent upgrade" for the downhole geological structure imaging system through the introduction of these two modules, truly realizing high-precision stratum identification and dynamic prediction under complex geological environments and a large amount of unlabeled data.

[0061] In summary, while inheriting the complete functions of Example 1, Example 2 focuses on the deep integration and application of GAN and self-supervised learning, provides an effective solution to the challenges of scarce data and noise faced by downhole geological imaging, and further promotes the application value of this system in actual oil and gas field drilling, geological exploration and other fields.

[0062] Embodiment 3: like Figure 1 As shown in the figure, based on the first two examples, the system already has the capabilities of gamma ray detection, multimodal fusion (including advanced methods such as GAN and self-supervised learning), and spatiotemporal Transformer modeling. However, for more complex underground environments, we need to further strengthen the following three aspects: Multi-head cross attention mechanism The gamma-ray signal is deeply correlated with other modes (resistivity, acoustic waves, temperature, pressure, etc.) to identify implicit features that exist between different modes but have not been explicitly captured.

[0063] More sophisticated spatiotemporal position encoding Further refine the temporal position encoding and spatial position encoding to ensure that under the spatiotemporal Transformer architecture, the details of the temporal changes of the strata and the distribution characteristics of different orientations can be fully captured.

[0064] Abnormal formation alarm module On the basis of multi-task support, a new real-time early warning function has been added to conduct online analysis of gamma rays and multimodal data collected underground. Once abnormal formation signals such as faults and cracks are identified, an alarm message will be issued to assist underground operations and safety decisions.

[0065] In general, Example 3 will "integrate and upgrade" the technical modules of the two previous examples, and through multi-head cross-attention and position coding mechanisms, it will enable more accurate and faster detection and alarm of formation anomalies.

[0066] Multi-head Cross-Attention is similar to the "Self-Attention" we mentioned earlier, but it specifically emphasizes the interaction and matching between different modalities or different data sources. Traditional multi-head attention often learns structural associations within the same modality or the same sequence, while multi-head "cross" attention can produce richer coupling between gamma-ray features and resistivity features, acoustic wave features and even GAN-completed data.

[0067] Specific process Input feature alignment: First, perform basic preprocessing (normalization, embedding, position encoding, etc.) on the gamma-ray data and other modality data to ensure that they can interact with each other in the same depth dimension. Cross-attention calculation, in a certain Attention layer in Transformer, the input comes from gamma rays and other modalities respectively, forming three sets of vectors: Query, Key, and Value; Multiple attention heads will "observe" the interaction between these two feature sources from different angles and output attention distributions in different dimensions; Finally, the outputs of all attention heads are concatenated, weighted, or linearly transformed in the channel dimension or feature dimension to generate a high-dimensional fusion feature. Implicit feature mining: thanks to multimodal cross-attention, previously imperceptible weak correlation information (for example, when the gamma-ray signal in a certain direction suddenly increases, and the sound wave data at the corresponding position also changes in certain frequency bands) can be better captured and given a higher feature weight; This helps identify subtle stratigraphic changes such as lithology transitions, potential fault boundaries, and may even provide real-time indication of complex factors such as drilling fluid disturbance.

[0068] Compared with self-attention, cross-attention is more targeted in the source of query, key, and value, and can establish deep connections between gamma-ray data and other modalities (resistivity, acoustic waves, etc.).

[0069] Cross-Attention Suppose we have two sets of inputs: In the cross-attention layer, we usually let: Single-head cross attention The calculation of single-head cross attention is similar to self-attention, except that the input source is divided into different modes: Multi-Head Cross-Attention If h parallel cross-attention heads are used, it can be expanded to: After concatenating the outputs of all attention heads in the channel dimension or feature dimension, they are then passed through W O Linear transformation finally generates high-dimensional fusion features.

[0070] Hidden Feature Mining Through the above cross-attention mechanism, the system can dynamically calculate the correlation weight between "gamma rays" and "other modalities" and extract those weak correlation phenomena that were originally difficult to detect. For example: Weak correlation enhancement When the gamma-ray signal x i When the Query suddenly increases at a certain moment, and the corresponding sound wave or resistivity in Y (Key & Value) also changes slightly in a specific frequency band, by comparing QK ⊤ The dot product result of the attention module will be these (x i ,y j ) are assigned higher weights, indicating that they have potential coupling relationships in stratigraphic characteristics.

[0071] Focus from multiple angles The multi-head mechanism ensures that each attention head can "observe" the data from different angles: For example, head 1 focuses on temporal patterns (short-term fluctuations), head 2 focuses on spatial proximity or filtering of noise signals, and head 3 focuses on high-frequency band features.

[0072] During the final fusion, the outputs of these different attention heads are concatenated and linearly transformed to form a more complete high-dimensional cross feature.

[0073] Identification of formation anomalies By capturing the multi-head cross-attention distribution of gamma rays and other modal signals, anomalies can be detected in time in terms of direction or time: If the attention weight remains high in certain directions and more than one mode is abnormal, it indicates that there may be more significant stratigraphic changes such as faults, cracks or well wall collapse.

[0074] In practical applications, these implicit features are not only conducive to the detailed characterization of rock transitions and potential fault boundaries, but can also predict and alarm downhole drilling fluid disturbances and sudden geological conditions through continuous tracking and subsequent model analysis, helping engineers to accurately adjust drilling plans or take protective measures in a timely manner.

[0075] Realizing a deep coupling of multimodal features such as gamma rays, resistivity, and acoustic waves allows the system to have a more delicate expression between "feature-level fusion" and "decision-level fusion". Enhanced sensitivity to abnormal strata: When faced with anomalies such as faults, cracks or fluid activity, only one modal signal may show abnormal movement. Through multi-head cross-attention, not only can this abnormal movement be captured, but also correlation analysis can be performed to see if other modal signals also show corresponding phenomena, thereby improving the confidence of anomaly detection. Adaptation to GAN-generated data: The high-resolution gamma-ray feature map generated by GAN mentioned in Example 2 can also be compared and integrated with the actual multimodal data through the cross-attention mechanism, so as to avoid the negative impact of "false data" on the overall judgment.

[0076] Transformer-based spatiotemporal modeling: Temporal Positional Encoding In Examples 1 and 2, we have discussed using timestamp information to characterize the dynamic changes of gamma-ray signals. In Example 3, the time position coding not only includes the traditional sine / cosine function to encode different time steps, but also introduces the concept of dynamic time interval: That is, for certain non-uniform sampling or "jump" sampling (such as downhole equipment pausing, accelerating or changing position, etc.), the system can adaptively adjust the encoding method to allow the model to more accurately reflect the actual time distribution of data acquisition. This more flexible time encoding strategy can further reduce the model's reliance on a "fixed sampling frequency" and maintain good temporal consistency modeling even when temporary fluctuations occur in the downhole environment.

[0077] Spatial Positional Encoding is similar to temporal encoding, and spatial positional encoding has also been extended accordingly: Considering that the azimuth layout of downhole detectors is not always 360° uniformly distributed (there may be a certain degree of inclination or offset in some cases), this embodiment can additionally record the actual azimuth coordinates, well diameter, well wall roughness and other information of the detectors, and convert them into learnable embedding vectors. At the same time, in multi-head cross-attention, these spatial position encodings will also be used to guide the attention matching between different modalities, so that the high-dimensional fusion features can accurately reflect the real geographical orientation and depth distribution of the well.

[0078] Construction of spatial position coordinates Assume that for the jth detector or acquisition position, we record the following spatial information: This information can be mapped with some learnable or fixed function to obtain a scalar or low-dimensional vector pos j ,For example: Where f(.) is a function that can be used to normalize or nonlinearly transform the roughness of the wellbore, etc., or it can be directly spliced {α j ,δ j ,ρ j} Encoding in high-dimensional space.

[0079] Spatial position encoding formula Similar to the time position encoding, the position index pos j Use sine and cosine functions for embedding. Let the position encoding vector dimension be d model , index i∈{0,1,2,…,d model −1}. Then the spatial position encoding of the jth acquisition point can be defined as: If we consider α j ,δ j ,ρ j They have different effects on the position encoding, and they can also be substituted or weighted separately during embedding, for example: Then map to d according to the above sine and cosine functions model dimensional space.

[0080] Application of multi-head cross attention In the multimodal cross attention layer, the spatial position encoding vector PE space (pos j ) After being spliced ​​or added with the original feature vectors (such as gamma intensity, resistivity, acoustic waves, etc.) collected by the detector, it can be used as the input of Query, Key, and Value to achieve more accurate orientation matching: Precise position matching Different detectors have different spatial position codes due to different inclinations, eccentricities or formation roughness; Cross-attention takes these positional encodings into account when computing similarity, ensuring that the mapping between multi-modalities is more physically meaningful.

[0081] Introducing implicit associations When the azimuth angles are similar or the wellbore roughness is similar, the corresponding pos j Probably closer; The attention distribution will classify such locations as "potentially relevant" or "neighborhood" areas, making subtle stratigraphic changes (such as the extension of cracks) easier to capture.

[0082] Through the above-mentioned spatial position encoding, this embodiment can flexibly adapt to the detector layout that is not strictly uniformly distributed, while fully integrating multi-source geological environment information such as well diameter, roughness, inclination, etc., effectively improving the performance of multi-head cross-attention in underground formation structure identification and dynamic analysis.

[0083] Richer and more flexible spatiotemporal encoding enables Transformer to accurately capture subtle changes in the temporal and azimuthal distribution of gamma-ray signals. Cooperation with GAN and self-supervised learning: When the completed data generated by GAN or obtained by self-supervised learning are integrated into the spatiotemporal modeling unit, these position codes can better constrain the consistency of the completed data with the real stratigraphic environment, thereby preventing overfitting problems such as "false faults" or "misplaced lithology".

[0084] Abnormal formation alarm function in multi-task support module The aforementioned embodiments have included multi-task capabilities, such as lithology classification, geological structure analysis, fault prediction, etc. However, in actual underground operations, timely detection and early warning of faults, cracks or other abnormal formations are often crucial and may directly affect operational safety and construction efficiency. Therefore, the abnormal formation alarm module aims at “real-time detection—rapid alarm” and integrates the modeling results of gamma rays and multimodal signals.

[0085] Real-time detection process Signal monitoring: The system continuously detects the predicted values ​​of multimodal fusion features and spatiotemporal Transformer at each time step or each well depth interval. Trigger mechanism: When the gamma ray intensity in a certain direction or other modal signals exceed the threshold (such as obvious abnormal mutations, spikes or continuous fluctuations), the cross-attention unit will synchronously evaluate: Is it just a single-mode random noise? Do other modal data also show corresponding abnormal signs? Have similar situations in historical data caused significant changes in the formation structure? Alarm generation: If the anomaly is supported by multi-modal cross-validation or historical model retrospective analysis with high confidence, the system will send an alarm to the ground platform in real time, including: abnormal category judgment (for example: suspected faults, cracks, well collapse, etc.); the depth and azimuth of the abnormal formation; the properties of the surrounding formations (such as possible lithology and porosity characteristics).

[0086] Decision support: Once an early warning is issued, ground geological engineers can respond based on the severity of the alarm and related information, such as adjusting the drilling path, strengthening well wall support, and checking for potential risks.

[0087] At each time step t or each well depth interval δ, the system will perform multimodal fusion feature F t And the predicted value of Transformer output We can define a multimodal monitoring vector: M t =[F t , ,…],in: F t Including fused features such as gamma rays, resistivity, and acoustic waves; is the prediction value of the spatiotemporal Transformer (or multi-task prediction, such as fault risk scoring, etc.); Other contextual information (such as wellbore inclination, real-time pressure, etc.) can be added as needed.

[0088] Triggering Mechanism Anomaly detection threshold In order to determine whether there is formation anomaly, a threshold τ can be set for each modal feature or predicted value. i . Let M t =[mt,1,m t ,2 ,…,m t ,K ], where m t ,i represents the value of the i-th mode (or indicator) at time t, then the “preliminary anomaly detection” condition is defined as: m t ,i>τ i or |m t ,i − ∣>η i in represents the mean (or median) of the feature in historical data, η i Represents a threshold value set based on standard deviation or other robust indicators. If the feature meets this condition for Δt time steps or well sections, it is considered a “suspected anomaly”.

[0089] Multimodal Cross Validation Once a suspected anomaly occurs, the cross-attention unit will perform a synchronous evaluation to determine whether anomalies occur simultaneously in the same direction or adjacent depth between different modalities. We can define the cross-attention distribution A t (From multi-head cross attention calculation): Q i For the gamma ray mode, K j , V j For other modes (resistivity, acoustic wave, etc.); If A t (i, j) is greater than a certain threshold λ, indicating a strong correlation between mode i and mode j at this position / moment. When the suspected anomaly mainly comes from a single mode and A t (i, j) is low, it may be just random noise; if significant anomalies appear in multiple modes simultaneously and the attention score is high, it may be a real formation change.

[0090] Historical comparison Before making a final judgment, the current moment M t will also be compared with the similar patterns of M t′ in historical data (t′ < t). For example: If it is found that a similar characteristic pattern has occurred in history and is confirmed as an actual situation such as "severe fault" or "wellbore collapse", the confidence level of this anomaly detection will be increased; If the similarity is very low or the corresponding situation has been determined as "equipment noise", the alarm level can be lowered.

[0091] Alarm Generation If the above multi-modal verification and historical comparison results show a high confidence level of anomaly, an alarm can be sent in real time. Let the anomaly confidence level evaluation function be: Decision Support Once the alarm information is sent, the on-site geological engineer will make corresponding decisions according to the following forms: Adjust the drilling path: If the risk of Fault or Fracture is high; Strengthen the wellbore support: Such as when Caving or the wellbore roughness is too large; Investigate potential risks: If the alarm level is not high, but historical data shows that such anomalies may evolve into serious faults; Ignore or mark: If on-site inspection and background data confirm that this alarm is a "false alarm", it can be marked in the man-machine interaction system.

[0092] This process can be summarized as: The multimodal fusion features and Transformer prediction output are "real-time sensed", and cross-attention and historical comparison are combined to perform "trigger judgment", and finally an alarm is generated and decision-making is assisted after the abnormality is confirmed. Through this method, major risks such as formation faults, cracks or well wall collapse can be discovered in time, significantly enhancing the safety and efficiency of downhole logging and geological exploration.

[0093] The model enhanced by GAN and self-supervised learning introduced in Example 2 enables the abnormal formation alarm module to maintain a high detection accuracy even in scenarios with missing data or noisy data. The multi-head cross-attention mechanism allows the alarm system to quickly screen "pseudo-anomalies" (noise in one mode while other modes are normal), reducing false alarms; it also allows real anomalies to be supported by more modes, thereby improving the confidence of the alarm.

[0094] With the help of multi-head cross-attention, the system can combine multimodal features more fully and capture more key information related to stratigraphic fractures, rock transitions, pressure anomalies, etc. at the difficult-to-observe "implicit level". Time position encoding and spatial position encoding make the spatiotemporal Transformer more "close to reality" when processing data, especially in cases of non-uniform sampling or eccentric detector layout, it can still maintain a high level of spatiotemporal modeling accuracy.

[0095] It makes up for the real-time and safety requirements of multi-task support in the previous embodiments, provides a rapid alarm channel for actual underground operations, greatly reduces operational risks, and provides a clearer and more actionable basis for subsequent decision-making.

[0096] The GAN and self-supervised learning results introduced in Example 2 are not separated at this stage. Instead, through further refinement of the cross-attention mechanism and position encoding, the quality and accuracy of formation anomaly alarms are simultaneously improved, making the system more flexible and practical in real oil and gas fields or geological exploration scenarios.

[0097] On the premise that the first two embodiments have completed the basic functions, GAN data enhancement and self-supervised learning, Example 3 further highlights the important value of multi-head cross attention, refined spatiotemporal position encoding and abnormal formation alarm for underground geological structure imaging system. Through this series of improvements: The system can establish deeper implicit associations between multimodal signals and generate more accurate and robust high-dimensional fusion features; With the help of spatiotemporal Transformer, the time and space distribution of heights can be modeled to timely detect and warn of abnormal areas in complex formations; Real-time performance, automation and collaborative gains are significantly improved, providing geologists and underground operation teams with powerful "prophetic" capabilities, greatly reducing construction risks and uncertainties.

[0098] Therefore, Example 3 and the previous two examples form a complete multi-modal, multi-stage, highly intelligent geological structure imaging system, which has extremely important technical support and practical value for oil and gas field drilling, geological disaster warning and underground engineering construction.

[0099] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A downhole geological structure imaging system based on azimuthal gamma rays, characterized by: Includes the following modules: Gamma ray detection module, used to collect azimuthal gamma ray intensity and energy spectrum data of underground formations; Multimodal data acquisition module, used to collect downhole multimodal data other than gamma rays; Signal processing and correction module, used for preprocessing and dynamic correction of gamma-ray data and multi-modal data; The intelligent modeling and multimodal fusion module, based on the Transformer architecture and multimodal data fusion algorithm, is used to generate high-resolution images of underground geological structures and realize multi-task functions, including: Data input: The gamma-ray intensity data of each azimuth detector constitutes an input sequence; the input sequence is divided by timestamp to form a spatiotemporal data set; Intelligent Modeling: Processing Gamma-Ray Data Using the Transformer Architecture: Autoregressive modeling: predicting the signal change at the next position or time point based on the current detector signal; Attention mechanism: Capture the correlation between gamma-ray intensity in direction and time through multi-head attention to analyze the heterogeneity of the formation; Use positional encoding to represent the azimuth and depth position of each gamma-ray detection point; The multimodal feature fusion unit uses feature-level fusion and decision-level fusion to generate unified multimodal fusion features using deep features extracted from modal data such as gamma rays, resistivity, and acoustic waves; Dynamic weight allocation unit, which uses modality perception mechanism to dynamically adjust the weights of different modal features according to signal quality and task requirements; The spatiotemporal modeling unit, based on the spatiotemporal Transformer architecture, captures the spatiotemporal correlation of formation signals through time position coding and spatial position coding, which is used to identify and predict dynamic changes of formations; Multi-tasking support module: Lithology classification: Use gamma-ray spectrum data to predict formation lithology; Geological structure analysis: identification of formation dips, cracks, and fault locations; Fault prediction: combining historical data and current azimuth signals to predict future signal trends and infer the location of faults or abnormal formations; Imaging output: Compare the predicted results with the actual data and dynamically generate two-dimensional or three-dimensional geological structure maps; Supports the use of AI algorithms to complete missing data and improve the integrity of imaging; The data transmission and display module is used to transmit the collected and analyzed data to the ground in real time and display them visually.

2. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The gamma-ray detection module includes: an annular multi-directional gamma-ray detector array, distributed in 360° directions, each detector is responsible for collecting gamma-ray signals in a specific direction; a dynamic signal adjustment unit, which adjusts the detector sensitivity in real time to adapt to different downhole environments (such as drilling fluid interference and wellbore eccentricity).

3. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The multimodal data acquisition module includes: a resistivity sensor for acquiring conductivity information of the formation; an acoustic logging sensor for acquiring density and porosity of the formation; a temperature and pressure sensor for monitoring downhole environmental parameters; and a data synchronization interface for time synchronization and integration of gamma ray data and multimodal data.

4. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The signal processing and correction module includes: a preprocessing unit that normalizes and denoises the gamma-ray data to improve the signal-to-noise ratio; a dynamic correction mechanism that automatically corrects signal errors caused by environmental changes such as drilling fluid absorption and wellbore eccentricity by analyzing historical data and real-time signals.

5. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The data transmission and display module includes: a data transmission unit, which adopts high-speed wireless communication or optical fiber technology to support low-latency transmission of large amounts of data; a visualization interface for displaying two-dimensional or three-dimensional geological structure maps, lithology distribution maps, fault location maps and real-time signal trend maps, and supports historical data overlay analysis.

6. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The intelligent modeling and multimodal fusion module also includes an adversarial learning module, which generates high-resolution gamma-ray feature maps through a generative adversarial network (GAN), thereby enhancing the utilization of scarce data.

7. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The intelligent modeling and multimodal fusion module includes a self-supervised learning module, which reduces the dependence on manually labeled data and improves the robustness of the model to low-quality data through pseudo-label generation and self-supervised target optimization.

8. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The intelligent modeling and multimodal fusion module captures the implicit correlation between gamma rays and other modal data through a multi-head cross-attention mechanism, generating high-dimensional fusion features for characteristic identification of complex formations.

9. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The spatiotemporal modeling unit is based on the Transformer architecture and uses time position coding to capture the dynamic characteristics of the gamma-ray signal changing over time, and spatial position coding to characterize the spatial distribution characteristics of the gamma-ray signal in different directions.

10. The underground geological structure imaging system based on azimuthal gamma ray according to claim 1, characterized in that: The multi-task support module further includes an abnormal formation alarm module, which detects abnormal formation signals (such as faults or cracks) in real time and generates alarms to assist downhole decision-making.

Images