Pre-stored fracture identification method and system based on deep learning

Abstract

The invention relates to a pre-stored fracture identification method based on deep learning, and the method comprises the steps: (1), obtaining seismic data meeting a set signal-to-noise ratio condition according to original three-dimensional seismic data; step (2), based on the obtained seismic data meeting a set signal-to-noise ratio condition, preprocessing of data normalization, continuous splicing and resampling is carried out; step (3), according to the preprocessed data, selecting a set area to carry out fine interpretation and making a label data set; step (4), training a deep learning model of a set network architecture by using the label data set; and step (5), using the trained deep learning model to carry out identification of the pre-stored fracture of the to-be-identified area.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a method and system for identifying pre-existing fractures based on deep learning. Background Technology

[0002] Pre-existing faults are faults located below the basement formed by tectonic activity before the formation of rift basins. Identifying pre-existing faults helps to clarify the influence of early tectonic movements on later tectonic movements and also plays an important role in the exploration of deep oil and gas resources.

[0003] Due to the large burial depth and ambiguous quality of deep seismic data, the identification of pre-existing faults faces significant challenges with existing technologies. Currently, gravity and magnetic data are the primary guideline. While these data offer broad and comprehensive coverage, magnetic anomalies are significantly influenced by lithological differences and the superposition of field sources at different depths. Spatial gravity anomalies are closely related to seafloor topography, structural morphology, and interfaces, providing valuable guidance for the planar distribution of large regional faults. However, gravity and magnetic data lack sufficient resolution and accuracy for smaller, localized areas, making detailed and in-depth research impossible. For precise localized studies, the main approach currently relies on manual interpretation of seismic profile characteristics. Constrained by existing regional well data, pre-existing faults are interpreted regionally and spatially closed, and planar distribution maps of these faults are created. This method is labor-intensive, and manual interpretation of pre-existing regional faults is inefficient. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a deep learning-based method and system for identifying pre-existing faults. Compared to traditional gravity, magnetic properties, and seismic attributes, this method can more objectively and realistically reflect the planar distribution characteristics of pre-existing faults. It helps to analyze the mechanism of selective activation of pre-existing structures, enabling them to better serve exploration and production, guide the selection of advantageous exploration areas, and bring significant economic and social benefits to exploration.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, this application provides a pre-existing fracture identification method based on deep learning, comprising: Step (1): Obtain seismic data that meets the set signal-to-noise ratio conditions based on the original three-dimensional seismic data; Step (2): Based on the obtained seismic data that meets the set signal-to-noise ratio conditions, perform data normalization, patch stitching and resampling preprocessing; Step (3): Based on the preprocessed data, select a specific region for detailed interpretation and create a labeled dataset; Step (4): Train a deep learning model with a defined network architecture using the labeled dataset; Step (5): Using the trained deep learning model, identify the pre-existing fractures in the region to be identified.

[0006] In one implementation, in step (1), obtaining seismic data that satisfies the set signal-to-noise ratio condition includes: Compared to the original 3D seismic data, the signal-to-noise ratio of deep faults is improved.

[0007] In one implementation, methods to improve the signal-to-noise ratio of deep fractures include: Use frequency reduction and Q compensation; or Construction-guided filtering based on fracture property constraints.

[0008] In one implementation, in step (2), the resampling is set to 4ms resampling.

[0009] In one implementation, in step (3), the formation mechanism and seismic profile characteristics of pre-existing faults are analyzed to determine the existence patterns of pre-existing faults in different periods.

[0010] In one implementation, the labeled dataset uses 3D labels.

[0011] In one implementation, the deep learning model that defines the network architecture is a Transformer-based network model.

[0012] In one implementation, step (5) further includes: To address the identification noise in pre-existing fracture identification, a morphological operation of fine-line processing is employed.

[0013] The present invention has the following advantages due to the adoption of the above technical solutions: The method provided in this invention aims to enhance the signal-to-noise ratio of deep seismic data by highlighting deep fault imaging. Preprocessing, including seismic data normalization, patch stitching, and resampling, improves the continuity of pre-existing fault identification. By analyzing the formation mechanism and seismic profile characteristics of pre-existing faults, the existence patterns of pre-existing faults at different stages are determined. For areas with widespread pre-existing fault distribution, areas with dense pre-existing fault development are selected for detailed interpretation and labeled datasets are created. This invention applies the Transformer network model, which has stronger anti-interference capabilities compared to other network models. Based on 3D seismic data, this invention can identify regional pre-existing faults using the improved Transformer network model. Compared to traditional gravity, magnetic, and seismic attribute methods, it can more objectively and realistically reflect the planar distribution characteristics of pre-existing faults, helping to analyze the mechanism of selective activation of pre-existing structures, enabling them to better serve exploration production, guide the selection of advantageous exploration areas, and bring significant economic and social benefits to exploration. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall method flow in one embodiment of this application; Figures 2(a) and 2(b) are comparative schematic diagrams of the construction of a guided filter before and after one embodiment of this application; Figure 3 This is a schematic diagram illustrating the differences in pre-existing fractures in this application; Figure 4(a) and Figure 4(b) are before and after comparisons of a profile 1 for application scenario recognition; Figure 5(a) and Figure 5(b) are before and after comparisons of a profile 2 for application scenario recognition; Figure 6 This is a fault prediction planar distribution map for an application scenario. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0016] based on Figure 1 The above forms the overall framework of the present invention. In one embodiment, a pre-existing fracture identification method based on deep learning is described.

[0017] The method of this application embodiment includes the following steps: Step (1): Obtain seismic data that meets the set signal-to-noise ratio conditions based on the original three-dimensional seismic data; Specifically, addressing the low signal-to-noise ratio (SNR) characteristic of deep seismic data, this study employs a fault-guided approach, using frequency reduction and Q-compensation to lower the resolution and improve the SNR of deep data. A fault-controlled structural-guided smoothing filter is applied to preserve fault information while reducing noise interference. This method utilizes multiple attributes, including stratigraphic dip angle, azimuth angle scanning, and maximum likelihood properties, to perform directional filtering, thereby preserving fault information and improving the SNR. Furthermore, it enhances the continuity of phase axes while emphasizing cross-sectional imaging.

[0018] Figures 2(a) and 2(b) illustrate the comparison before and after structural guidance filtering. Figure 2(a) shows the original seismic profile, and Figure 2(b) represents the effect of structural guidance filtering based on stratigraphic and fault attribute constraints.

[0019] Step (2): Based on the obtained seismic data that meets the set signal-to-noise ratio conditions, perform data normalization, patch stitching and resampling preprocessing; The fracture identification process predicts fracture information in seismic data by moving a fixed-unit cube at equal intervals. However, the small receptive field of fracture identification can cause the network model to become overly sensitive to seismic information, misidentifying noise as faults. Furthermore, limitations imposed by physical memory and algorithm model parameters prevent the window from being directly expanded.

[0020] This step indirectly expands the identification window through resampling to improve the continuity of deep fault identification. A comparison between the identification results of the original data and the identification results of the resampled 4ms data reveals that the identification of faults in earthquakes without resampling is easily affected by noise, while the continuity of fault identification in earthquake data after 4ms resampling is significantly enhanced.

[0021] Step (3): Based on the preprocessed data, select a specific region for detailed interpretation and create a labeled dataset; There are two main categories of methods for creating fault identification label datasets: the first is based on synthetic data. Models trained on synthetic data labels perform well on high-quality seismic data with clear fault lines. However, their performance is poor for identifying deep faults with low signal-to-noise ratios. In contrast, manually interpreted labels fully utilize expert experience and perform well in identifying actual seismic data profiles. However, two-dimensional labels do not utilize three-dimensional spatial information. Experimental results show that network models trained solely on two-dimensional manual labels identify discontinuous fault plane distributions. To address these shortcomings, this invention combines the aforementioned equally spaced 2D data labels into 3D labels for fault prediction. The 3D labels created by this method are discontinuous. Compared to models trained on 2D data labels, models trained on these 3D labels show a significant improvement in the planar continuity of fault identification.

[0022] Specifically, this step involves creating tags for pre-existing faults based on the characteristics of seismic data and guided by geological understanding. Pre-existing faults are categorized into two main types based on whether they cut through the basement: late-stage activated faults and late-stage inactive faults. Compared to late-stage faults developed in rift basins, early-stage pre-existing faults primarily exhibit "fault-wave" characteristics in seismic data, distinguishing them from late-stage faults that mainly exhibit "phase axis faulting" characteristics. The tag types show significant differences. Due to multiple periods of regional tectonic activity in the early stages, different tectonic stress fields resulted in faults of varying attitudes and scales. On the cross-section, they can be classified as high-angle and low-angle pre-existing faults, such as... Figure 3 The meaning is as shown.

[0023] Pre-existing faults are widely distributed in the study area of ​​the northern South China Sea. To improve identification efficiency, the method of this invention selects a 500km stretch of pre-existing faults that are densely developed and have a comprehensive pattern distribution. 2 Seismic data was used as a fault labeling test area. It was interpreted in detail (40 equally spaced channels) and labeled. The generated fault label data volume was used to train a model and predict the distribution of pre-existing faults in the entire study area.

[0024] Step (4): Train a deep learning model with a defined network architecture using the labeled dataset; With the continuous iteration of artificial intelligence deep learning network models, fault identification algorithms have achieved significant application results. Existing technologies employ algorithms such as the U-net network model. In areas with good seismic data quality and uncomplicated faults, the identification results of the above models are not significantly different. However, when the geological structure is relatively complex, the faults are fragmented, and the seismic data quality is poor, the U-net network model is easily affected by noise, and the identification accuracy decreases significantly.

[0025] Compared to the U-Net network, this step utilizes the Transformer network with a self-attention mechanism. Its dynamically adaptive global receptive field is more advantageous, allowing it to better focus on fault information and effectively improve the accuracy of fault identification. Through application comparison, the Transformer network model demonstrates stronger anti-interference capabilities and superior results for identifying pre-existing faults. To fully leverage the advantages of the Transformer architecture in global context modeling, a U-shaped encoder-decoder structure is introduced into the 3D fault prediction network, and the self-attention mechanism is further subdivided into more efficient spatial attention and channel attention.

[0026] Step (5): Using the trained deep learning model, identify the pre-existing fractures in the region to be identified.

[0027] After obtaining the identification results in this step, post-processing of the fault prediction results can be performed.

[0028] Since the signal-to-noise ratio varies among different earthquakes, the identification results will exhibit varying degrees of identification noise depending on the quality of the seismic data. To address this issue, this invention employs a fine-line processing method to reduce "clump-like" noise, making fine-line faults easier to interpret.

[0029] The following describes the actual effects achieved by the above-described methods in a specific application scenario.

[0030] Application scenarios introduction: This study is applied to the identification and research of pre-existing faults in the Pearl River Estuary in the northern South China Sea. This region was mainly controlled by the strong compression of the pre-Cenozoic Indosinian and Yanshanian orogenies, which formed pre-existing thrust faults in the Pearl River Estuary Basin, mainly in the NE and NW directions. During the Indosinian period, under the SW-NE compressive stress field, large-scale NW-trending thrust-nappe structures were formed. During the Yanshanian period, under the NW-SE compressive stress field, the early large-scale high-angle NW-trending thrust faults gradually evolved into deep strike-slip faults.

[0031] This study focuses on the Xijiang 36 Depression and its surrounding faults, and explores methods and applications for identifying pre-existing faults. Firstly, labeling data was created from the 3D seismic data of the Xijiang 36 Depression's periphery, including early-stage low-angle faults trending NW, early-stage high-angle faults trending NW, and northeast-trending faults. Labeling data was then created for each type of fault, and model training was completed.

[0032] Figures 4(a), 4(b), 5(a), and 5(b) show schematic diagrams of the recognition results. The technical solution before optimization was: no improvement in deep signal-to-noise ratio, no resampling, and prediction using a U-net network structure.

[0033] Based on the above identification results, the basal interface layers were flattened and shifted downwards to create layer-by-layer attribute slices of the fracture identification body. The planar distribution map of the layer-by-layer slices reveals the development scale and orientation of pre-existing fractures. Through digital identification of the pre-existing fracture images, it was determined that the dominant pre-existing fracture in the Xijiang 36 depression trends NE-NEE, the dominant pre-existing fracture in the Huixi low uplift trends NWW, the dominant pre-existing fracture in the Xijiang medium-low uplift trends NE, and the dominant pre-existing fracture in the Dongsha uplift trends NWW. This provides support for subsequent analysis and research on pre-existing fractures. (See schematic diagram below.) Figure 6 .

[0034] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0035] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A pre-existing fracture identification method based on deep learning, characterized in that, The method comprises the following steps: Step (1), obtaining seismic data satisfying a set signal-to-noise ratio condition according to original three-dimensional seismic data; Step (2), based on the obtained seismic data satisfying the set signal-to-noise ratio condition, performing data normalization, splicing and re-sampling preprocessing; Step (3), according to the preprocessed data, selecting a set region for fine interpretation and making a label data set; Step (4), training a deep learning model with a set network architecture by using the label data set; Step (5), using the trained deep learning model to identify pre-existing faults in a to-be-identified region.

2. The method of claim 1, wherein, In the step (1), the obtained seismic data satisfying the set signal-to-noise ratio condition comprises: Compared with the original three-dimensional seismic data, the signal-to-noise ratio of deep faults is improved.

3. The method of claim 2, wherein, The method for improving the signal-to-noise ratio of deep faults comprises: Downshifting and Q compensation; or Structure-oriented filtering based on fault attribute constraints.

4. The method of claim 1, wherein, In the step (2), the re-sampling is set to 4ms re-sampling.

5. The method of claim 1, wherein, In the step (3), the formation mechanism and seismic profile characteristics of pre-existing faults are analyzed to determine the existing patterns of pre-existing faults of different periods.

6. The method of claim 5, wherein, The label data set adopts a 3D label.

7. The method of claim 1, wherein, The deep learning model with a set network architecture is a network model based on Transformer.

8. The method of claim 1, wherein, In the step (5), the method further comprises: For the identification noise of pre-existing fault identification, a morphological operation of thinning processing is used for post-processing of the prediction result.

9. A computer storage medium, characterized in that The computer program is stored in the memory, and the computer program is executed by the processor to realize the method of any one of claims 1 to 8.

10. A pre-existing fault identification system based on deep learning, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is executed by the processor to realize the method of any one of claims 1 to 8.

Images