AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system and method

Abstract

This invention discloses an AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system and method, relating to the interdisciplinary field of artificial intelligence and geophysical exploration. The system includes: multiple downhole artificial seismic sources, multiple inter-well seismic signal receiving devices, a distributed acquisition and control network, and an AI processing and interpretation platform. This invention supports four stereoscopic acquisition modes: single-well excitation of adjacent well reception, single-well excitation of multiple well reception, multiple-well excitation of single-well reception, and multiple-well excitation of multiple well reception. Through deep learning models, it achieves intelligent denoising, automatic first arrival picking, intelligent wavefield separation, AI-driven full-waveform inversion velocity modeling, intelligent amplitude-preserving migration imaging, and reservoir fluid prediction. Finally, it combines reinforcement learning to optimize development well deployment. This invention realizes intelligent management of the entire process of inter-well seismic data acquisition, processing, and interpretation, significantly improving the acquisition efficiency, imaging resolution, and reservoir description accuracy of inter-well seismic data.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of applied geophysics, artificial intelligence and oil and gas exploration and development, and specifically relates to an AI-based multi-source multi-receiver well seismic data acquisition, processing and interpretation system and method. Background Technology

[0002] Inter-well seismic surveying is a seismic exploration method conducted within wells, employing excitation and receiving systems adapted to the downhole environment. The downhole excitation system can consist of sources with various polarization excitation modes, characterized by high energy and wide bandwidth. The inter-well seismic data receiving system comprises multi-stage geophones or hydrophones, capable of receiving rich wavefield information. The dominant frequency of inter-well seismic signals is typically several times or even higher than that of surface earthquakes, and its imaging results can clearly reveal the geological structure between wells, used for detailed reservoir characterization.

[0003] Traditional inter-well seismic acquisition systems typically employ a single type of source and a single type of receiver, with acquisition primarily based on single-well excitation and adjacent-well reception. This approach is insufficient to meet the demands of multi-dimensional exploration of complex oil and gas reservoirs. As exploration targets become increasingly complex, higher demands are placed on the flexibility, efficiency, and accuracy of inter-well seismic data acquisition. Multi-source, multi-detector combined acquisition and multi-well multi-dimensional observation are emerging as technological trends.

[0004] In terms of data processing and interpretation, traditional inter-well seismic processing heavily relies on manual intervention and empirical parameter selection. Steps such as first arrival picking, wavefield separation, velocity modeling, migration imaging, and reservoir interpretation all require extensive manual operation, resulting in long processing cycles, high subjectivity, and limited accuracy. In recent years, artificial intelligence (AI) technology has made breakthroughs in computer vision, natural language processing, and scientific computing. Deep learning, reinforcement learning, and physical information neural networks have provided new technical pathways for the intelligent processing and interpretation of geophysical data. However, current technologies have not yet deeply integrated AI with multi-source, multi-detector inter-well seismic systems, lacking intelligent solutions covering the entire acquisition, processing, and interpretation process.

[0005] Therefore, there is an urgent need for an inter-well seismic system that can integrate multiple downhole seismic sources and multiple inter-well seismic data receiving devices, support flexible three-dimensional acquisition modes, and achieve fully automated processing and intelligent comprehensive interpretation based on artificial intelligence technology. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based multi-source, multi-receiver well seismic data acquisition, processing, and interpretation system and method to solve the technical problems of single acquisition mode, reliance on manual processing and interpretation, low efficiency, and insufficient accuracy in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides an AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system, comprising:

[0009] The source subsystem includes at least one inter-well seismic source well and a downhole source deployed in the inter-well seismic source well;

[0010] The receiving subsystem includes at least one inter-well seismic receiving well and inter-well seismic data receiving equipment deployed in the inter-well seismic receiving well;

[0011] The acquisition control and synchronization subsystem includes a distributed acquisition control node, a downhole seismic source excitation control device, a multi-channel modulation and demodulation instrument, and a synchronization cable, which are used to realize time synchronization and trigger control between the downhole seismic source and the inter-well seismic data receiving device;

[0012] An artificial intelligence processing and interpretation platform is used for deep learning-driven intelligent processing, inversion imaging, and comprehensive reservoir interpretation of acquired inter-well seismic data.

[0013] The system supports the following four stereoscopic acquisition modes:

[0014] Single-well excitation and adjacent-well reception mode: a single or multiple array-type downhole seismic sources in an inter-well seismic source well excite the seismic data, which is received by the inter-well seismic data receiving equipment in an adjacent inter-well seismic receiving well;

[0015] Single-well excitation and multi-well reception mode: a single or multiple array-type downhole seismic sources in a single well-to-well seismic source well excite the seismic data, which is simultaneously received by the inter-well seismic data receiving equipment in multiple surrounding inter-well seismic receiving wells.

[0016] Multi-well excitation and single-well reception mode: The downhole sources in the seismic source wells of multiple wells are excited sequentially according to a preset time sequence or coding strategy, and the data is received by the inter-well seismic data receiving equipment in the single-well seismic receiving well.

[0017] Multi-well excitation and multi-well reception mode: The downhole sources in the seismic source wells of multiple wells are excited sequentially according to a preset time sequence or coding strategy, and are simultaneously received by the inter-well seismic data receiving equipment in the inter-well receiving wells.

[0018] Furthermore, the downhole seismic source includes at least one of the following: a downhole controllable seismic source, a downhole eccentric wheel seismic source, a downhole piezoelectric crystal seismic source, a downhole electric spark seismic source, a downhole gas explosion seismic source, a downhole explosive or detonator seismic source, a downhole plasma seismic source, and a downhole air gun seismic source.

[0019] The downhole controllable vibration source is a frequency-controlled scanning vibration source; the downhole eccentric wheel vibration source is a mechanical eccentric rotating impact source; the downhole piezoelectric crystal vibration source is an electromechanical transducer pulse source; the downhole electric spark vibration source is a high-voltage discharge pulse source; the downhole gas explosion vibration source is a high-pressure gas explosion pulse source; the downhole explosive or detonator vibration source is a chemical explosion source; the downhole plasma vibration source is a high-energy plasma pulse source; and the downhole air gun vibration source is a high-pressure air release pulse source.

[0020] Furthermore, the inter-well seismic data receiving equipment includes at least one of the following: a downhole four-component data receiving array, a downhole three-component geophone array, a downhole hydrophone array, a downhole accelerometer array, a downhole armored optical cable, a downhole armored spiral optical cable, and a downhole armored 3C-DAS optical cable; wherein the downhole four-component data receiving array is composed of a combination of a three-component geophone and a hydrophone, and the downhole armored 3C-DAS optical cable is a three-component distributed optical fiber acoustic wave sensing optical cable;

[0021] The downhole four-component data receiving array includes a multi-stage push-to-push hydrophone and three orthogonally arranged detectors for simultaneously recording scalar pressure fields and vector particle vibration fields; the downhole three-component detector array includes a multi-stage push-to-push three-component detector; the downhole hydrophone array includes a multi-stage hydrophone receiving cable based on fluid coupling; the downhole accelerometer array includes a multi-stage receiving array based on MEMS or fiber optic accelerometers; the downhole armored optical cable includes an armored optical cable with built-in straight single-mode optical fiber, featuring high temperature resistance, high sensitivity, high reflection coefficient, and hydrogen loss resistance; the downhole armored spiral optical cable includes a cylindrical elastomer and a single-mode optical fiber wound spirally on the outer surface of the cylindrical elastomer; the downhole armored 3C-DAS optical cable includes three optical fiber cores arranged at specific spatial angles for sensing three-dimensional vector strain fields.

[0022] Furthermore, the artificial intelligence processing and interpretation platform includes a data preprocessing module, an intelligent imaging module, and a comprehensive interpretation module:

[0023] The data preprocessing module is equipped with a deep learning denoising model, an intelligent first arrival picking model, and an intelligent wavefield separation model, which are used to suppress noise, automatically pick up the first arrival of direct waves, and intelligently separate reflected waves / converted waves from the raw inter-well seismic data.

[0024] The intelligent imaging module is equipped with a physical information neural network velocity modeling model, an AI-driven full waveform inversion model, and an intelligent amplitude-preserving migration imaging model, which are used to construct a high-precision three-dimensional inter-well velocity model and a high-resolution migration imaging data volume.

[0025] The comprehensive interpretation module is equipped with a multi-attribute fusion reservoir prediction model, an intelligent fluid identification model, and a well location optimization decision model, which are used for inter-well geological structure interpretation, reservoir parameter prediction, fluid distribution identification, and intelligent optimization deployment of development well locations.

[0026] The deep learning denoising model employs a neural network based on U-Net, DnCNN, or Transformer architecture, and is trained using inter-well seismic data acquired from multiple sources and multiple detectors along with corresponding noise labels to achieve intelligent suppression of random noise, cable waves, pipe waves, and coherent noise. The intelligent first arrival picking model uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) to automatically identify the first arrival of direct P-waves and direct S-waves. The intelligent wavefield separation model uses a deep learning network based on FNO (Fourier Neural Operator) or UNet++ architecture to intelligently separate up / down reflected P-waves, up / down reflected S-waves, and converted waves.

[0027] The physical information neural network velocity model adopts the PINN architecture, embedding the wave equation as a physical constraint into the neural network loss function, and using the picked first arrival travel time data to invert the three-dimensional inter-well P-wave velocity distribution and the three-dimensional inter-well S-wave velocity distribution; the AI-driven full waveform inversion model uses the three-dimensional inter-well P-wave velocity distribution and the three-dimensional inter-well S-wave velocity distribution as the initial model, and combines the generative adversarial network (GAN) or diffusion model for regularization constraints. Through three-dimensional full waveform P-wave / S-wave forward and inversion, the three-dimensional inter-well P-wave FWI velocity model and the three-dimensional inter-well S-wave FWI velocity model are obtained.

[0028] The intelligent amplitude-preserving migration imaging model employs Q-RTM or Q-LSRTM algorithms, combined with a deep learning amplitude compensation network. It utilizes the three-dimensional inter-well P-wave / S-wave FWI velocity model, the separated up- and down-flow reflection data volumes, and AI-estimated subsurface medium Q... P and Q S The values ​​are used to perform amplitude-preserving offset imaging processing for longitudinal and transverse waves, resulting in amplitude-preserving offset imaging data volumes for longitudinal and transverse waves.

[0029] The multi-attribute fusion reservoir prediction model employs a multimodal Transformer or graph neural network (GNN) architecture, fusing P-wave / S-wave amplitude-preserving migration imaging data, P-wave / S-wave velocity models, and attenuation Q-factors. P and Q SThe system uses seismic attribute parameters to predict the distribution of porosity, permeability, and saturation in reservoirs between wells; the intelligent fluid identification model uses a deep learning classifier or semi-supervised learning framework based on an attention mechanism to identify the distribution patterns of oil, gas, and water three-phase fluids in oil and gas reservoirs between wells; the well location optimization decision model uses reinforcement learning DQN or PPO algorithms combined with genetic algorithms for global optimization, and optimizes the deployment of development wells, adjustment wells, and infill wells based on reservoir description results.

[0030] Secondly, this invention provides an AI-based method for acquiring, processing, and interpreting multi-source, multi-receiver inter-well seismic data, comprising the following steps:

[0031] S1. Multi-mode three-dimensional data acquisition: Based on geological task requirements, select one of the following acquisition modes: single-well excitation and adjacent-well reception, single-well excitation and multi-well reception, multi-well excitation and single-well reception, or multi-well excitation and multi-well reception; deploy the selected downhole seismic source in the inter-well seismic source well, and deploy the selected inter-well seismic data receiving equipment in the inter-well seismic receiving well; achieve time synchronization between the seismic source and the receiving equipment through the acquisition control and synchronization subsystem, and perform equal-interval or continuous excitation and reception according to the designed interval to acquire inter-well seismic wave data for the entire well section;

[0032] In step S1, the multi-well excitation single-well reception or multi-well excitation multi-well reception mode uses a pseudo-random coding excitation strategy or a time-delay coding excitation strategy for the downhole seismic source within the wells between multiple wells. The encoding and decoding algorithm of the artificial intelligence processing and interpretation platform is used to realize the intelligent separation and imaging of multi-source aliased data.

[0033] S2. AI-driven data preprocessing: The collected raw inter-well seismic data is input into the data preprocessing module of the artificial intelligence processing and interpretation platform; intelligent denoising is performed using a deep learning denoising model; the first arrival of direct P-waves and direct S-waves is automatically picked up using an intelligent first arrival picking model; intelligent separation of up-and-down reflected P-waves and up-and-down reflected S-waves is performed using an intelligent wavefield separation model; simultaneously, the Q-wave of the subsurface medium is calculated using the amplitude attenuation ratio method or the center frequency shift method. P Value and transverse wave Q S value;

[0034] The deep learning denoising model described in step S2 adopts a multi-task learning framework, simultaneously learning denoising and wave field separation tasks; the intelligent first arrival picking model adopts a semi-supervised learning strategy, using a small amount of manually labeled first arrivals and a large amount of unlabeled data for joint training; the intelligent wave field separation model adopts physical information constraints, adding wave equation conservation terms and energy conservation terms to the loss function to ensure the physical consistency of the separated wave fields.

[0035] S3. AI-driven 3D velocity model construction: Using a physical information neural network velocity model, 3D tomography inversion is performed based on the picked first arrival travel time to obtain the 3D inter-well P-wave velocity distribution and the 3D inter-well S-wave velocity distribution; using this as the initial model, the 3D full waveform inversion model is used to perform 3D full waveform P-wave / S-wave forward and inversion to obtain the 3D inter-well P-wave FWI velocity model and the 3D inter-well S-wave FWI velocity model.

[0036] The AI-driven full-waveform inversion model described in step S3 adopts a multi-scale inversion strategy, gradually inverting from low frequency to high frequency, and introducing a generative adversarial network (GAN) as a regularization term at each scale to suppress high-frequency noise and artifacts in the inversion process; at the same time, it uses transfer learning technology to transfer the model parameters trained by the inter-well seismic data of the existing work area to the new work area, accelerating convergence and improving generalization ability.

[0037] S4. AI-driven intelligent migration imaging: Utilizing an intelligent amplitude-preserving migration imaging model, combined with the aforementioned three-dimensional inter-well P-wave / S-wave FWI velocity model, the separated reflected wave data volume, and the subsurface medium Q... P and Q S The values ​​are processed by Q-RTM or Q-LSRTM, and amplitude consistency is corrected by a deep learning amplitude compensation network to obtain longitudinal wave amplitude-preserving offset imaging data volume and transverse wave amplitude-preserving offset imaging data volume.

[0038] S5. Multi-attribute intelligent extraction and reservoir prediction: Extract various inter-well seismic attribute parameters from the P-wave / S-wave amplitude-preserving migration imaging data volume; predict inter-well reservoir physical property parameters using a multi-attribute fusion reservoir prediction model; identify the distribution patterns of oil, gas and water three-phase fluids using an intelligent fluid identification model to obtain fine description results of inter-well reservoirs;

[0039] The various inter-well seismic attribute parameters mentioned in step S5 include: instantaneous amplitude, instantaneous frequency, instantaneous phase, reflection intensity, absorption attenuation attribute, P-wave / S-wave velocity ratio, Poisson's ratio, anisotropy parameter, AVO attribute, and frequency variation attribute; the multi-attribute fusion reservoir prediction model automatically assigns weights to different attributes through an attention mechanism to achieve adaptive fusion.

[0040] S6. Intelligent Integrated Interpretation and Well Location Optimization: High-resolution and detailed interpretation of inter-well geological structures based on P-wave / S-wave migration imaging data; Intelligent optimization deployment of development wells, adjustment wells, and infill wells based on reservoir physical parameters and fluid distribution results, using a well location optimization decision model to improve single-well production and oilfield recovery rate.

[0041] The well location optimization decision model described in step S6 uses the detailed description results of reservoirs between wells, the interpretation results of geological structures, the existing well network constraints, and economic cost parameters as the state space, the well location coordinates, well inclination angle, and well depth as the action space, and the maximization of net present value (NPV) or the maximization of recovery rate as the reward function. Through the interaction between the reinforcement learning agent and the environment, it outputs the optimal well location deployment scheme.

[0042] The beneficial effects of this invention are:

[0043] 1. Flexible and diverse acquisition modes: Supports four modes: single-well excitation and adjacent-well reception, single-well excitation and multi-well reception, multi-well excitation and single-well reception, and multi-well excitation and multi-well reception. It can be flexibly configured according to geological targets and well network conditions to achieve high-density three-dimensional observation and significantly improve acquisition efficiency and coverage times.

[0044] 2. Multi-source and multi-detector combination: It integrates eight downhole seismic sources and seven inter-well seismic data receiving devices, covering traditional electronic detectors, hydrophones, accelerometers and new distributed fiber optic sensing (DAS) technology. It can achieve joint acquisition of multi-physics field, multi-wavelength, and multi-scale information according to the optimal combination based on target depth, frequency band requirements and well conditions.

[0045] 3. AI-driven throughout the entire process: Deep learning, physical information neural networks, reinforcement learning and other artificial intelligence technologies are embedded into the entire process of data preprocessing, velocity modeling, migration imaging, reservoir prediction and well location optimization, so as to realize automated and intelligent processing from raw data to development decisions, significantly shorten the processing cycle and reduce the reliance on human experience.

[0046] 4. Improved imaging accuracy and resolution: Through AI-FWI, Q-compensated intelligent offset and physical information-constrained wavefield separation, high-resolution, high-fidelity inter-well P-wave / S-wave imaging data volumes are obtained, which can finely characterize the micro-structures and reservoir heterogeneity between wells.

[0047] 5. Intelligent Decision Support: Based on multi-attribute fusion and reinforcement learning, well location optimization can comprehensively consider geological, engineering and economic factors, output scientific and reasonable development well location deployment plans, and improve single well production and oilfield ultimate recovery rate. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the system architecture of the AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system provided in this embodiment of the invention;

[0049] Figure 2 This is a flowchart of the AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system method provided in this embodiment of the invention;

[0050] Figure 3This is a schematic diagram of multi-mode inter-well seismic data acquisition provided in an embodiment of the present invention;

[0051] Figure 4 These are schematic diagrams of four stereoscopic acquisition modes provided in the embodiments of the present invention;

[0052] Figure 5 This is a schematic diagram of the cross-section of the downhole armored 3C-DAS optical cable provided in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the downhole seismic source type and intelligent optimization system provided in an embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of the types of downhole receiving equipment and the collaborative acquisition system provided in the embodiments of the present invention;

[0055] Figure 8 This is a flowchart of AI-driven data preprocessing provided in an embodiment of the present invention;

[0056] Figure 9 This is a flowchart of AI-FWI full waveform inversion and intelligent migration imaging provided in an embodiment of the present invention.

[0057] Figure 10 This is a flowchart of intelligent well location optimization decision-making based on reinforcement learning provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0059] like Figure 1 As shown in the system architecture diagram, the AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system of this embodiment includes:

[0060] Seismic source subsystem: A controllable downhole seismic source 8a or an eccentric wheel seismic source 8b is deployed in the inter-well seismic source well 1. The controllable downhole seismic source 8a is connected to the downhole seismic source excitation and control equipment 7 inside the inter-well seismic source vehicle 6 via an armored photoelectric composite cable 81; the eccentric wheel seismic source 8b is also connected via a photoelectric composite cable. The downhole seismic source excitation and control equipment 7 is connected to the multi-channel modulation and demodulation instrument 5 inside the seismic data acquisition vehicle 4 in the well via a synchronization cable 9 (synchronization cable or optical cable) to achieve nanosecond-level time synchronization.

[0061] Receiving Subsystem: A downhole four-component data receiving array 3a (three-component geophone + hydrophone) or a downhole armored 3C-DAS optical cable 3b is deployed in the inter-well seismic receiving well 2. When using an electronic geophone / hydrophone array, it is connected to a multi-channel modem 5 via a multi-core cable; when using a DAS optical cable, it is connected to the DAS modem 5 via a single-mode optical fiber. The inter-well seismic receiving well 2 is equipped with casing and oil / gas tubing. The armored optical cable can be fixed to the outer wall of the casing or oil / gas tubing via equally spaced metal clips, or adsorbed onto the inner wall of the metal casing via a high-magnetic ring / high-magnetic steel pipe.

[0062] Artificial intelligence processing and interpretation platform 10: Deployed at a ground computing center or in the cloud, it connects to the seismic data acquisition vehicle 4 in the well in real time or near real time via a high-speed network, receives raw data and outputs processing and interpretation results.

[0063] AI-based methods for multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation, such as Figure 2 As shown, it includes the following steps:

[0064] S1. Multi-mode stereo data acquisition:

[0065] Four stereoscopic acquisition modes are implemented:

[0066] like Figure 3 Schematic diagram of multi-mode inter-well seismic data acquisition and Figure 4 The four stereoscopic acquisition modes are illustrated in the planar diagram. This embodiment describes the specific implementation methods of the four acquisition modes:

[0067] Mode A: Single well triggers adjacent well reception ( Figure 2 a) Select a central well as source well 1, and deploy receiving equipment in an adjacent receiving well 2. The downhole source generates seismic signals at equal intervals from the bottom to the top of the well, and the receiving well receives signals synchronously throughout its entire length. This mode is suitable for fine structural imaging between two wells.

[0068] Mode B: Single-well excitation and multi-well reception ( Figure 2 (b) Using a central well as the seismic source well 1, receiving equipment is simultaneously deployed in 3-4 surrounding receiving wells 2. A single excitation is used, with multiple wells receiving simultaneously, avoiding inconsistencies in energy, spectrum, and coupling caused by multiple excitations, thus significantly improving acquisition efficiency.

[0069] Mode C: Multi-well excitation, single-well reception ( Figure 2 c). Two or more source wells 1a and 1b are sequentially excited using a pseudo-random coding or time-delay coding strategy, while a single receiving well 2 receives the aliased data. High-precision separation of the aliased data is achieved through the coding / decoding algorithm and intelligent wavefield separation model in the artificial intelligence processing and interpretation platform 10.

[0070] Mode D: Multi-well excitation and multi-well reception ( Figure 2 d). A three-dimensional observation system is formed by multiple source wells and multiple receiving wells. By adopting a coding excitation strategy, it can realize ultra-high density, all-angle, and multi-offset inter-well seismic data acquisition, which is suitable for monitoring complex fractured reservoirs or heavy oil thermal recovery.

[0071] Figure 5 This is a schematic diagram of the cross-section of the downhole armored 3C-DAS optical cable provided in an embodiment of the present invention.

[0072] Figure 6 This is a schematic diagram of the downhole seismic source type and intelligent optimization system provided in an embodiment of the present invention.

[0073] Figure 7 This is a schematic diagram of the types of downhole receiving equipment and the collaborative acquisition system provided in the embodiments of the present invention.

[0074] S2, AI-driven data preprocessing:

[0075] like Figure 8 The AI-driven data preprocessing flowchart is shown in this embodiment. This embodiment details the data preprocessing process:

[0076] Step S2-1: Deep Learning Intelligent Denoising. A deep learning denoising model based on the Transformer architecture is constructed. The input is raw inter-well seismic data acquired from multiple sources and detectors, and the output is the denoised clean data. During model training, a large amount of noisy data (including random noise, pipe waves, cable waves, and coherent interference) is generated using numerical simulation, and data augmentation strategies (such as random pruning, amplitude perturbation, and time flipping) are used to expand the training set. The loss function uses a combination of L1 loss and perceptual loss. For DAS data, a physical consistency loss is specifically introduced for strain-displacement transformation.

[0077] Step S2-2: Intelligent First Arrival Picking. A first arrival picking model based on a bidirectional LSTM-CNN hybrid architecture is constructed. The model input is denoised single-channel or multi-channel inter-well seismic records, and the output is the first arrival time of the direct P-wave and the first arrival time of the direct S-wave. During the training phase, a semi-supervised strategy is adopted: first, the model is pre-trained using a small amount of manually labeled data; then, the model is used to perform pseudo-label prediction on a large amount of unlabeled data; and finally, the model performance is improved through confidence filtering and iterative optimization. For low signal-to-noise ratio channels, the model combines the spatial correlation between adjacent channels for constrained picking.

[0078] Step S2-3: Intelligent Wavefield Separation. An intelligent wavefield separation model based on the Fourier Neural Operator (FNO) is constructed. The input is full-wavefield inter-well seismic data, and the output is the separated up-reflected P-wave, down-reflected P-wave, up-reflected S-wave, down-reflected S-wave, and converted wave. During training and inference, the model enforces the wave equation and energy conservation relationship through a physical information constraint layer, ensuring the physical rationality of the separated wavefield. For multi-component data (four-component, three-component, 3C-DAS), the model utilizes multi-channel input to learn the polarization and phase relationships between different components, improving separation accuracy.

[0079] Step S2-4: Smart Q P and Q S Value estimation. The time-frequency characteristics of the downlink direct wave are automatically extracted using a deep learning time-frequency analysis network (such as STFT-CNN). The P-wave Qp and S-wave Qs are calculated using the amplitude attenuation ratio method or the center frequency shift method, and a Q value estimation mechanism is established. P and Q S The value is a continuous profile along the well section.

[0080] S3, AI-driven 3D velocity model construction:

[0081] like Figure 9 As shown, this embodiment illustrates AI-driven velocity modeling and offset imaging:

[0082] Step S3-1: PINN Intelligent Tomography. A Physical Information Neural Network (PINN) is constructed. The network input is the three-dimensional coordinates of the source and receiver points, and the output is the inter-well P-wave velocity V. P With transverse wave velocity V S During network training, in addition to data-driven loss (first-arrival travel time residuals), a physical constraint loss is also incorporated: the wave equation for the velocity field predicted by the network, solved through automatic differentiation, should be consistent with the observed travel time. A multi-scale training strategy is used to first fit the low-frequency background velocity and then gradually recover the high-frequency details.

[0083] Step S3-2: AI-FWI Full Waveform Inversion. Using the PINN tomography results as the initial model, AI-driven full waveform inversion is initiated. The inversion process employs a multi-scale strategy, gradually transitioning from low frequencies (e.g., 10Hz) to high frequencies (e.g., 200Hz). At each frequency scale, a pre-trained Generative Adversarial Network (GAN) is introduced as a regularizer: the generator attempts to generate a "realistic" geological model from the current velocity model, and the discriminator distinguishes between the real geological model and the generated model. Adversarial training suppresses high-frequency noise and outliers in the inversion. Simultaneously, transfer learning is used to transfer the network parameters from adjacent work areas or similar geological conditions to the current work area, reducing the number of iterations by more than 50%.

[0084] S4, AI-driven intelligent offset imaging:

[0085] Using the FWI velocity model obtained in step S3 and the reflected wave data volume separated in step S2, Q-RTM or Q-LSRTM migration is performed. During the migration process, a deep learning amplitude compensation network (such as a Q-compensation network based on U-Net) is introduced, which learns Q... P and Q S The value absorption attenuation compensation operator performs real-time intelligent compensation for amplitude loss during the migration process, resulting in high-resolution imaging data volume with amplitude preservation.

[0086] Example 5: Intelligent Reservoir Prediction and Well Location Optimization

[0087] like Figure 10 The flowchart of the intelligent well location optimization decision-making based on reinforcement learning is shown in this embodiment. This embodiment illustrates the comprehensive explanation and intelligent decision-making:

[0088] S5. Multi-attribute intelligent extraction and reservoir prediction:

[0089] More than 20 seismic attributes were extracted from the P-wave / S-wave amplitude-preserving migration data to construct a multimodal feature set. Using a multimodal Transformer model, different attributes were treated as different modes. A self-attention mechanism was employed to learn the nonlinear relationships and weight allocations between attributes, outputting a three-dimensional volume of well porosity, permeability, and saturation. For fluid identification, a graph neural network (GNN) based on an attention mechanism was used. Well logging data were used as node labels, seismic attributes as node features, and spatial adjacency relationships as edges. Graph convolution and attention aggregation were used to predict fluid type and distribution probability.

[0090] S6. Intelligent Integrated Interpretation and Well Location Optimization:

[0091] A well location optimization environment simulator is constructed. The state space includes: current reservoir description results (porosity, saturation, and permeability in three dimensions), geological structure interpretation (fault and fracture distribution), existing well network location, and economic cost parameters (drilling cost, oil price, and recovery coefficient). The action space consists of the three-dimensional coordinates of candidate well locations and wellbore parameters. The reward function is designed as follows: ;

[0092] Where ΔNPV is the net present value increment, and ΔEUR is the projected recoverable reserves increment. For drilling costs, For geological risk penalty items, α, β, γ, and δ are weighting coefficients.

[0093] The agent is trained using the Proximal Policy Optimization (PPO) algorithm. The agent learns the optimal policy through multiple interactions with an environmental simulator. The algorithm outputs a well deployment scheme that maximizes long-term cumulative rewards. To enhance global optimization capabilities, a genetic algorithm (GA) is combined with the PPO: the strategy output by the PPO is used as the initial population of the GA, and optimization is performed in the global solution space through selection, crossover, and mutation operations. Finally, the algorithm outputs a deployment scheme for development wells, adjustment wells, and densification wells that balances local refinement and global optimality.

[0094] Multi-source joint acquisition example:

[0095] In a fractured-vuggy carbonate reservoir, a multi-well excitation and multi-well reception mode was adopted. Two source wells were equipped with downhole spark sources (high frequency, suitable for fracture and vuggy identification) and downhole controllable sources (low frequency, suitable for background velocity modeling), respectively. Three receiving wells were equipped with downhole four-component data receiving arrays (two wells for vector wave field recording) and downhole armored 3C-DAS optical cables (one well for high-density full-well section reception), respectively. Alternating encoding excitation from the two source wells and synchronous reception from the three receiving wells were achieved through distributed acquisition and control nodes. After processing by an artificial intelligence processing and interpretation platform, AI-FWI imaging clearly identified fractures and vuggy bodies between wells, intelligent reservoir prediction accurately identified oil and gas enrichment areas, and reinforcement learning-optimized infill wells encountered high-yield layers, increasing single-well production by 35%.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system, characterized in that: include: The source subsystem includes at least one inter-well seismic source well and a downhole source deployed in the inter-well seismic source well; The receiving subsystem includes at least one inter-well seismic receiving well and inter-well seismic data receiving equipment deployed in the inter-well seismic receiving well; The acquisition control and synchronization subsystem includes a distributed acquisition control node, a downhole seismic source excitation control device, a multi-channel modulation and demodulation instrument, and a synchronization cable, which are used to realize time synchronization and trigger control between the downhole seismic source and the inter-well seismic data receiving device; An artificial intelligence processing and interpretation platform is used for deep learning-driven intelligent processing, inversion imaging, and comprehensive reservoir interpretation of acquired inter-well seismic data.

2. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system according to claim 1, characterized in that, The system supports the following four stereoscopic acquisition modes: Single-well excitation and adjacent-well reception mode: a single or multiple array-type downhole seismic sources in an inter-well seismic source well excite the seismic data, which is received by the inter-well seismic data receiving equipment in an adjacent inter-well seismic receiving well; Single-well excitation and multi-well reception mode: a single or multiple array-type downhole seismic sources in a single well-to-well seismic source well excite the seismic data, which is simultaneously received by the inter-well seismic data receiving equipment in multiple surrounding inter-well seismic receiving wells. Multi-well excitation and single-well reception mode: The downhole sources in the seismic source wells of multiple wells are excited sequentially according to a preset time sequence or coding strategy, and the data is received by the inter-well seismic data receiving equipment in the single-well seismic receiving well. Multi-well excitation and multi-well reception mode: The downhole sources in the seismic source wells of multiple wells are excited sequentially according to a preset time sequence or coding strategy, and are simultaneously received by the inter-well seismic data receiving equipment in the inter-well receiving wells.

3. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system according to claim 1, characterized in that, The downhole seismic source includes at least one of the following: a downhole controllable seismic source, a downhole eccentric wheel seismic source, a downhole piezoelectric crystal seismic source, a downhole electric spark seismic source, a downhole gas explosion seismic source, a downhole explosive or detonator seismic source, a downhole plasma seismic source, and a downhole air gun seismic source. The downhole controllable vibration source is a frequency-controlled scanning vibration source; the downhole eccentric wheel vibration source is a mechanical eccentric rotating impact source; the downhole piezoelectric crystal vibration source is an electromechanical transducer pulse source; the downhole electric spark vibration source is a high-voltage discharge pulse source; the downhole gas explosion vibration source is a high-pressure gas explosion pulse source; the downhole explosive or detonator vibration source is a chemical explosion source; the downhole plasma vibration source is a high-energy plasma pulse source; and the downhole air gun vibration source is a high-pressure air release pulse source.

4. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system according to claim 1, characterized in that, The well-to-well seismic data receiving equipment includes at least one of the following: a downhole four-component data receiving array, a downhole three-component geophone array, a downhole hydrophone array, a downhole accelerometer array, a downhole armored optical cable, a downhole armored spiral optical cable, and a downhole armored 3C-DAS optical cable; wherein the downhole four-component data receiving array is composed of a combination of a three-component geophone and a hydrophone, and the downhole armored 3C-DAS optical cable is a three-component distributed optical fiber acoustic wave sensing optical cable. The downhole four-component data receiving array includes a multi-stage push-to-push hydrophone and three orthogonally arranged detectors for simultaneously recording scalar pressure fields and vector particle vibration fields; the downhole three-component detector array includes a multi-stage push-to-push three-component detector; the downhole hydrophone array includes a multi-stage hydrophone receiving array based on fluid coupling; the downhole accelerometer array includes a multi-stage receiving array based on MEMS or fiber optic accelerometers; the downhole armored optical cable includes an armored optical cable with built-in straight single-mode optical fiber, featuring high temperature resistance, high sensitivity, high reflection coefficient, and hydrogen loss resistance; the downhole armored spiral optical cable includes a cylindrical elastomer and a single-mode optical fiber wound spirally on the outer surface of the cylindrical elastomer; the downhole armored 3C-DAS optical cable includes three optical fiber cores arranged at specific spatial angles for sensing three-dimensional vector strain fields.

5. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system according to claim 1, characterized in that, The artificial intelligence processing and interpretation platform includes a data preprocessing module, an intelligent imaging module, and a comprehensive interpretation module: The data preprocessing module is equipped with a deep learning denoising model, an intelligent first arrival picking model, and an intelligent wavefield separation model, which are used to suppress noise, automatically pick up the first arrival of direct waves, and intelligently separate reflected waves / converted waves from the raw inter-well seismic data. The intelligent imaging module is equipped with a physical information neural network velocity modeling model, an AI-driven full waveform inversion model, and an intelligent amplitude-preserving migration imaging model, which are used to construct a high-precision three-dimensional inter-well velocity model and a high-resolution migration imaging data volume. The comprehensive interpretation module is equipped with a multi-attribute fusion reservoir prediction model, an intelligent fluid identification model, and a well location optimization decision model, which are used for inter-well geological structure interpretation, reservoir parameter prediction, fluid distribution identification, and intelligent optimization deployment of development well locations.

6. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation system according to claim 5, characterized in that, The deep learning denoising model employs a neural network based on U-Net, DnCNN, or Transformer architecture, and is trained using inter-well seismic data acquired from multiple sources and multiple detectors along with corresponding noise labels to achieve intelligent suppression of random noise, cable waves, pipe waves, and coherent noise. The intelligent first arrival picking model uses a convolutional neural network (CNN) or a long short-term memory network (LSTM) to automatically identify the first arrival of direct P-waves and direct S-waves. The intelligent wavefield separation model uses a deep learning network based on FNO or UNet++ architecture to intelligently separate up / down reflected P-waves, up / down reflected S-waves, and converted waves. The physical information neural network velocity model adopts the PINN architecture, embedding the wave equation as a physical constraint into the neural network loss function, and using the picked first arrival travel time data to invert the three-dimensional inter-well P-wave velocity distribution and the three-dimensional inter-well S-wave velocity distribution; the AI-driven full waveform inversion model uses the three-dimensional inter-well P-wave velocity distribution and the three-dimensional inter-well S-wave velocity distribution as the initial model, and combines the generative adversarial network (GAN) or diffusion model for regularization constraints, and obtains the three-dimensional inter-well P-wave FWI velocity model and the three-dimensional inter-well S-wave FWI velocity model through three-dimensional full waveform P-wave / S-wave forward and inversion; The intelligent amplitude-preserving migration imaging model employs the Q-RTM or Q-LSRTM algorithm, combined with a deep learning amplitude compensation network. It utilizes the three-dimensional inter-well P-wave / S-wave FWI velocity model, the separated up- and down-flow reflection data volumes, and AI-estimated subsurface medium Q... P and Q S The values ​​are used to perform amplitude-preserving offset imaging processing for longitudinal and transverse waves, resulting in amplitude-preserving offset imaging data volumes for longitudinal and transverse waves. The multi-attribute fusion reservoir prediction model adopts a multimodal Transformer or graph neural network (GNN) architecture, fusing P-wave / S-wave amplitude-preserving migration imaging data, P-wave / S-wave velocity models, and attenuation Q-factors. P and Q S The system uses seismic attribute parameters to predict the distribution of porosity, permeability, and saturation in reservoirs between wells; the intelligent fluid identification model uses a deep learning classifier or semi-supervised learning framework based on an attention mechanism to identify the distribution patterns of oil, gas, and water three-phase fluids in oil and gas reservoirs between wells; the well location optimization decision model uses reinforcement learning DQN or PPO algorithms combined with genetic algorithms for global optimization, and optimizes the deployment of development wells, adjustment wells, and infill wells based on reservoir description results.

7. An AI-based method for multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation, applied to the system described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Multi-mode three-dimensional data acquisition: Select one of the following acquisition modes according to the geological task requirements: single well excitation and adjacent well reception, single well excitation and multiple well reception, multiple well excitation and single well reception, or multiple well excitation and multiple well reception. Selected downhole seismic sources are deployed in the inter-well seismic source wells, and selected inter-well seismic data receiving equipment is deployed in the inter-well seismic receiving wells. The time synchronization between the seismic sources and the inter-well seismic data receiving equipment is achieved through the acquisition control and synchronization subsystem. Seismic data is acquired from the entire well section by performing equal-interval or continuous excitation and reception according to the design interval. S2, AI-driven data preprocessing: The collected raw inter-well seismic data is input into the data preprocessing module of the artificial intelligence processing and interpretation platform; Intelligent denoising is performed using a deep learning denoising model; the first arrival of direct P-waves and direct S-waves is automatically picked up using an intelligent first arrival picking model; intelligent separation of up / down reflected P-waves and up / down reflected S-waves is performed using an intelligent wavefield separation model; and the Q-wave of the subsurface medium is calculated using the amplitude attenuation ratio method or the center frequency shift method. P Value and transverse wave Q S value; S3. AI-driven 3D velocity model construction: Utilizing a physical information neural network velocity model, 3D tomographic inversion is performed based on the picked first arrival travel time to obtain the 3D inter-well P-wave velocity V. P Distribution and three-dimensional inter-well shear wave V S Velocity distribution; using this as the initial model, the three-dimensional full-waveform P-wave / S-wave forward and inverse model is performed using the AI-driven full-waveform inversion model to obtain the three-dimensional inter-well P-wave FWI velocity model and the three-dimensional inter-well S-wave FWI velocity model. S4. AI-driven intelligent migration imaging: Utilizing an intelligent amplitude-preserving migration imaging model, combined with the aforementioned three-dimensional inter-well P-wave / S-wave FWI velocity model, the separated reflected wave data volume, and the subsurface medium Q... P and Q S The values ​​are processed by Q-RTM or Q-LSRTM, and amplitude consistency is corrected by a deep learning amplitude compensation network to obtain longitudinal wave amplitude-preserving offset imaging data volume and transverse wave amplitude-preserving offset imaging data volume. S5. Multi-attribute intelligent extraction and reservoir prediction: Extract various inter-well seismic attribute parameters from the P-wave / S-wave amplitude-preserving migration imaging data volume; predict inter-well reservoir physical property parameters using a multi-attribute fusion reservoir prediction model; identify the distribution patterns of oil, gas and water three-phase fluids using an intelligent fluid identification model to obtain fine description results of inter-well reservoirs; S6. Intelligent Integrated Interpretation and Well Location Optimization: High-resolution and detailed interpretation of inter-well geological structures based on P-wave / S-wave migration imaging data; Intelligent optimization deployment of development wells, adjustment wells, and infill wells based on reservoir physical parameters and fluid distribution results, using a well location optimization decision model to improve single-well production and oilfield recovery rate.

8. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation method according to claim 7, characterized in that, In step S1, the multi-well excitation single-well reception or multi-well excitation multi-well reception mode uses a pseudo-random coding excitation strategy or a time-delay coding excitation strategy for the downhole seismic source within the wells between multiple wells. The encoding and decoding algorithm of the artificial intelligence processing and interpretation platform is used to realize the intelligent separation and imaging of multi-source aliased data.

9. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation method according to claim 7, characterized in that, The deep learning denoising model described in step S2 adopts a multi-task learning framework, simultaneously learning denoising and wave field separation tasks; the intelligent first arrival picking model adopts a semi-supervised learning strategy, using a small amount of manually labeled first arrivals and a large amount of unlabeled data for joint training; the intelligent wave field separation model adopts physical information constraints, adding wave equation conservation terms and energy conservation terms to the loss function to ensure the physical consistency of the separated wave fields. The AI-driven full-waveform inversion model described in step S3 adopts a multi-scale inversion strategy, gradually inverting from low frequency to high frequency, and introducing a generative adversarial network (GAN) as a regularization term at each scale to suppress high-frequency noise and artifacts in the inversion process; at the same time, it uses transfer learning technology to transfer the model parameters trained by the inter-well seismic data of the existing work area to the new work area, accelerating convergence and improving generalization ability.

10. The AI-based multi-source, multi-receiver inter-well seismic data acquisition, processing, and interpretation method according to claim 7, characterized in that, The various inter-well seismic attribute parameters mentioned in step S5 include: instantaneous amplitude, instantaneous frequency, instantaneous phase, reflection intensity, absorption attenuation attribute, P-wave / S-wave velocity ratio, Poisson's ratio, anisotropy parameter, AVO attribute, and frequency variation attribute; the multi-attribute fusion reservoir prediction model automatically assigns weights to different attributes through an attention mechanism to achieve adaptive fusion; The well location optimization decision model described in step S6 uses the detailed description results of reservoirs between wells, the interpretation results of geological structures, the existing well network constraints, and economic cost parameters as the state space, the well location coordinates, well inclination angle, and well depth as the action space, and the maximization of net present value (NPV) or the maximization of recovery rate as the reward function. Through the interaction between the reinforcement learning agent and the environment, it outputs the optimal well location deployment scheme.

Images