Three-dimensional fault detection self-supervision pre-training method based on volume-level contrast learning

Abstract

The invention provides a three-dimensional fault self-supervision pre-training method based on volume-level contrast learning. The method is used for enhancing the fault structure characterization capacity under the label-free condition. The method comprises the following steps: firstly, carrying out volume partitioning on a seismic body to form a sub-block view with a fixed spatial position; and then randomly cutting target sub-blocks consistent with the size of the blocks, and inputting the target sub-blocks and the volume blocks into a shared coding network to obtain uniform feature representation. And constructing a position prediction task by calculating the real overlapping proportion of the target block and the sub-block, and guiding the model to deduce the geometric proximity relationship according to the feature similarity. Meanwhile, space smooth regularization based on the three-dimensional Manhattan distance is introduced, so that adjacent features are more continuous, remote features are more distinguished, and continuity and abrupt change features of the fault are fitted. Finally, through joint optimization of the position prediction loss and the spatial smoothing loss, the model learns robust volume-level spatial structure representation, and the accuracy and stability of downstream fault detection are effectively improved.

Description

Technical Field

[0001] This invention belongs to the fields of geophysics and artificial intelligence applications, specifically relating to a self-supervised pre-training method for three-dimensional fault detection based on volumetric comparative learning. Background Technology

[0002] Three-dimensional fault identification is one of the core tasks in seismic data interpretation, playing an irreplaceable role in engineering scenarios such as oil and gas exploration, deep resource evaluation, and earthquake disaster monitoring. By analyzing three-dimensional seismic reflection data, the location, morphology, and extension of underground faults can be inferred, providing crucial evidence for structural analysis, reservoir prediction, and risk assessment. However, three-dimensional seismic data often contains complex features such as strong noise, weak reflections, and irregular stratigraphic structures, posing significant challenges to the automatic fault extraction process. Furthermore, the significant cross-scale variations and complex spatial morphologies of fault structures further increase the difficulty for detection algorithms to maintain stability and accuracy in complex structural regions. Traditional fault detection relies on manual annotation by experienced interpreters, but the massive scale of three-dimensional data, high annotation costs, long processing times, and strong subjectivity make it difficult to obtain high-quality annotations on a large scale, thus limiting the application and performance of supervised learning-based methods in practical exploration scenarios.

[0003] With the development of deep learning, 3D feature modeling and image processing technologies have brought new ideas to fault detection. However, their reliance on large-scale, high-quality labeled data still falls short of meeting the reality of scarce labeling in the seismic industry. Against this backdrop, self-supervised learning, as a paradigm that learns expressive capabilities from unlabeled data without manual labeling, is gradually becoming an important direction for improving the intelligent processing capabilities of seismic data. By constructing proxy tasks that conform to the spatial structural characteristics of seismic bodies, self-supervised models can automatically extract effective features from the natural spatial organization within 3D seismic bodies, significantly reducing the reliance on manual labeling. In seismic application scenarios characterized by scarce labeling, complex structures, and massive data volumes, self-supervised pre-training can fully utilize massive amounts of unlabeled data to form geologically significant volumetric structural priors, providing a more efficient and robust feature foundation for subsequent fault detection. Summary of the Invention

[0004] Given that existing 3D fault detection methods heavily rely on labeled data and exhibit instability in complex seismic tectonic scenarios, this invention aims to provide a self-supervised 3D fault detection method based on volumetric comparative learning. By constructing a pre-training task that conforms to the spatial structural characteristics of seismic bodies, the model can automatically learn fault-related spatial representation capabilities on a large amount of unlabeled 3D seismic data. This method, through volumetric partitioning, spatial relationship inference, and distance-constrained modeling, enables the network to acquire geologically significant volumetric structural features under unlabeled conditions. This significantly improves the stability, accuracy, and generalization ability of downstream 3D fault detection models in complex noisy environments, weak reflection regions, and multi-scale fault scenarios, making it particularly suitable for seismic interpretation tasks where labeled data is scarce.

[0005] To achieve the above objectives, the technical solution of the present invention mainly includes the following steps:

[0006] A. Three-dimensional volume partitioning and basic spatial structure construction:

[0007] The input 3D seismic volume is spatially and uniformly divided according to a preset mesh size, dividing the complete volume into several non-overlapping 3D sub-blocks. Each sub-block corresponds to a specific spatial location within the original volume. Through this volume partitioning operation, the overall structural morphology, local fracture variations, and spatial geometric relationships of the seismic volume are discretized into structural units to construct a basic representation suitable for spatial relationship modeling.

[0008] B. Target sub-block selection and feature encoding:

[0009] A 3D sub-volume with the same size as the volume block is randomly cropped from the original 3D volume as the target block. The target block and its non-overlapping 3D sub-blocks are then input into a 3D feature extraction network with shared parameters to generate its high-dimensional feature representation. After projection mapping, all sub-block features are aligned to a unified embedding space, thus forming a comparable and measurable feature set that can be used for spatial relationship inference.

[0010] C. Target block location prediction based on spatial overlap ratio:

[0011] The normalized overlap ratio is calculated based on the true voxel overlap volume of the target block and each base block in the original 3D volume, serving as the spatial location label of the target block relative to the base blocks. Simultaneously, the cosine similarity between the target block and the base blocks in the feature space is used as the predicted value, and the difference between the true overlap label and the predicted similarity is used to construct the location prediction loss. This allows the network to automatically learn the spatial organization and fault-related geometric topological relationships between sub-blocks through a contrastive learning mechanism.

[0012] D. Spatial smoothing regularization based on three-dimensional Manhattan distance:

[0013] Distance weights are constructed based on the three-dimensional Manhattan distance between the center points of each sub-block, and these weights are used to constrain the similarity of features between any two sub-blocks, ensuring that neighboring sub-blocks maintain high consistency in the feature space and distant sub-blocks maintain low consistency. Through this distance-weighted regularization, the model learns spatial relationships while establishing a volumetric feature distribution that conforms to the continuity of seismic tectonic structures and the characteristics of fault jumps.

[0014] E. Pre-training loss construction and parameter update:

[0015] The location prediction loss and the spatial smoothing regularization loss are summed to form a self-supervised pre-training training objective. Iterative optimization is performed on a large number of unlabeled 3D seismic volumes. The feature extraction network and the learnable distance parameter are updated simultaneously through gradient backpropagation, so that the model gradually forms a robust volumetric spatial structure representation during continuous training. Attached Figure Description

[0016] Figure 1 This is a model structure diagram of the present invention. Detailed Implementation

[0017] The following is combined Figure 1 The present invention will be described in further detail as follows:

[0018] A. Data preprocessing and view generation:

[0019] First, the method of this invention discretizes the continuous volumetric data into a set of spatially fixed sub-blocks by performing volumetric block processing on the input 3D seismic data, which are then used to construct subsequent volumetric comparative learning tasks. The specific steps are as follows:

[0020] (1) Selecting a 3D data block: Assuming the original 3D data block ,in These are the height, width, and depth of the data, which are the basic parameters for volume partitioning and spatial relationship calculations.

[0021] (2) Fixed structure volume partitioning: In order to obtain a discrete space with consistent structure, the present invention divides the data volume into blocks. Divided equally along three dimensions as These non-overlapping cube sub-blocks form a set of sub-blocks. Each sub-block Spatial location is determined by its index in the 3D mesh. The only certainty is that among them , , These represent the grid numbers of the sub-blocks in the height, width, and depth directions, respectively. Due to the strict grid division, there is no spatial overlap between any two different sub-blocks, and their relative grid order can be used for subsequent geometric relationship analysis.

[0022] (3) Target block generation and overlap ratio label construction: In order to construct a supervision signal that can truly reflect the geometric proximity relationship, this invention starts from the original volume Randomly select a 3D sub-volume with the same volume as the non-overlapping sub-block as the target block. Since the target block may span one or more volume blocks, its relationship with any sub-block can be calculated. Overlapping volume:

[0023] Vol

[0024] Where Vol(·) represents the number of voxels in the 3D volume. Based on the overlapping volume, a normalized overlap ratio label can be further constructed:

[0025]

[0026] in , This describes the spatial distribution of the target block relative to the 64 sub-blocks. Although there is no overlap between volume blocks, the target block may span multiple block regions for cropping. Therefore, the scale label built based on overlapping volumes can accurately reflect their geometric proximity and has clear physical meaning.

[0027] (4) Construction of 3D Manhattan distance labels: In order to further characterize the spatial topological relationship between sub-blocks, the distance labels are constructed based on the grid center coordinates of each sub-block. (in , , Define the three-dimensional Manhattan distance between any two sub-blocks as:

[0028]

[0029] This distance is used to measure spatial proximity, where a smaller distance indicates closer proximity and a larger distance indicates greater distance. It can be used in subsequent spatial smoothing regularization terms to enhance the model's ability to express earthquake fault continuity and structural abrupt changes.

[0030] B. Target sub-block and volume block feature encoding:

[0031] After completing the three-dimensional volume segmentation and spatial label construction, this invention further refines the target block. With all volume blocks A unified feature encoding process is performed to establish a high-dimensional representation space that can be used for spatial relation learning. The specific steps are as follows:

[0032] (1) Three-dimensional feature extraction network: a three-dimensional feature extraction network with shared parameters The input volume is encoded. The network takes a 3D volume as input and outputs a high-dimensional feature vector describing local structural patterns and potential fracture features. A shared weight design ensures that the encodings of different volumes reside in a consistent feature representation domain, providing a foundation for subsequent spatial comparison and correlation modeling.

[0033] (2) Target block and sub-block feature encoding: Encoding the target block Input the feature extraction network to obtain its high-dimensional encoding vector. At the same time, 64 integral blocks The results of inputting each data point into the same network are as follows:

[0034]

[0035] in This represents the encoded representation of the target block in the feature space. This represents the encoded representation of each sub-block, with all encoded vectors having a dimension of 1. .

[0036] (3) Feature projection and unified embedding space construction: To further enhance the separability of spatial geometric relations in the feature space, this invention uses a feature mapping head. Projecting the encoded vectors onto a unified embedding space yields:

[0037]

[0038] in and This is the final feature representation used for subsequent spatial relationship inference; the feature dimension after projection is denoted as... .

[0039] (4) Constructing a spatial relationship inference feature set: embedding vector of the target block With the set of embedding vectors of 64 sub-blocks Together, they constitute a feature set for characterizing the three-dimensional spatial organization structure. This feature set, along with the spatial landing point labels obtained in the previous stage, constitutes a feature set. Manhattan distance between sub-blocks Combined, these techniques can be used for subsequent target location prediction and spatial smoothing regularization learning.

[0040] C. Target block location prediction based on spatial overlap ratio:

[0041] Obtain the target block Volume partitioning Feature representation and Subsequently, this invention utilizes spatial overlap ratio labels. As a supervisory signal, a task is constructed to predict the position of the target block in 3D space, so that the similarity in the feature space can accurately reflect its true geometric proximity relationship. The specific steps are as follows:

[0042] (1) Construction of prediction based on cosine similarity: by calculating the target block features With each sub-block feature The cosine similarity is used to obtain the predicted value of the target block's landing point on each sub-block, in the form of:

[0043]

[0044] Among the predicted values This reflects the model's view of the target block and sub-blocks in the feature space. The degree of relative closeness.

[0045] (2) Location Prediction Loss Design: To ensure that the predicted similarity distribution remains consistent with the proportion of actual spatial landing points, this invention employs an element-wise regression approach, comparing... With tags The difference is used to construct the supervision signal, and the location prediction loss is defined as:

[0046]

[0047] This loss forces the network to learn characteristic relationships that conform to the geometry of a three-dimensional seismic body.

[0048] D. Spatial smoothing regularization based on three-dimensional Manhattan distance

[0049] Obtain the three-dimensional Manhattan distance between sub-blocks and their corresponding feature representations Subsequently, this invention constructs distance-based feature smoothing constraints, enabling the relative relationships in the feature space to reflect the proximity structure of the real three-dimensional space. The specific steps are as follows:

[0050] (1) Weight allocation based on spatial distance: In order to maintain a stronger correlation between spatially adjacent regions in the feature space, this invention assigns weights based on the distance between sub-blocks. Construct distance decay weights:

[0051]

[0052] The weight is a function that decreases with distance, ensuring that the closer the sub-blocks are, the higher their weight and the stronger their constraint; while the weight of the more distant sub-blocks is close to zero, and the constraint effect is naturally weakened.

[0053] (2) Spatial smoothing loss construction: Based on the distance weight, this invention weights and accumulates the difference between the features of any two sub-blocks according to the spatial decay weight to obtain the spatial smoothing loss:

[0054]

[0055] in and They represent the first The and the first The feature representation of each sub-block in the embedding space, the loss is obtained through distance weights. Stronger feature consistency constraints are imposed on spatially adjacent sub-blocks, while weaker constraints are imposed on spatially distant sub-blocks, thereby maintaining continuity between adjacent regions in the feature space, while retaining necessary feature differences in distant regions.

[0056] E. Pre-training loss construction and parameter update

[0057] After completing spatial location prediction and smoothing regularization, this invention combines these two objectives for self-supervised training and iteratively optimizes the feature network parameters to form a stable spatial structure representation. The specific steps are as follows:

[0058] (1) Construction of comprehensive pre-training loss: This invention integrates the location prediction loss With spatial smoothing loss By summing them up, a complete self-supervised learning objective is formed:

[0059]

[0060] The two losses respectively constrain the model's ability to represent the spatial location of the target block in the feature space and the continuity of the spatial topological relationship between sub-blocks, thereby enabling the model to simultaneously learn local geometric alignment characteristics and overall spatial structure rules.

[0061] (2) Parameter update and spatial structure representation: The three-dimensional feature extraction network is updated uniformly by backpropagating the gradient of the comprehensive loss. With feature mapping head The network parameters enable it to gradually learn stable and reliable spatial structure representations on a large number of unlabeled 3D seismic volumes. During continuous training, the model can automatically construct a volumetric feature distribution that conforms to the seismic tectonic laws, providing more accurate and robust feature priors for subsequent 3D fault detection.

Claims

1. A self-supervised pre-training method for three-dimensional tomographic detection based on volumetric contrastive learning, characterized in that, Includes the following steps: A. Three-dimensional volume partitioning and basic spatial structure construction: The input 3D seismic volume is uniformly divided into several non-overlapping sub-blocks according to a fixed grid. Each sub-block corresponds to a unique spatial location in the original volume, so as to ensure that the geometry and fault distribution of the entire volume can be completely preserved in the discrete block-level representation. B. Target sub-block selection and feature encoding: A three-dimensional sub-volume with the same size as the volume block is randomly cut from the original three-dimensional volume as the target block, and the remaining sub-blocks are used as the base block set. All sub-blocks are input into a shared three-dimensional feature extraction network to obtain high-dimensional features. Then, a sub-block representation with a unified dimension is obtained through projection mapping, so that the target block features and the base block features together constitute an expression space that can be used to infer spatial relationships. C. Target block location prediction based on spatial overlap ratio: The normalized overlap ratio of the target block and each base block in the original 3D volume is calculated as the location label. At the same time, the cosine similarity between the target block features and the features of each base block is calculated as the prediction value. The difference between the two is used to construct the location prediction loss, thereby driving the network to learn the geometric organization and fault spatial distribution pattern between sub-blocks. D. Spatial smoothing regularization based on 3D Manhattan distance: To maintain the continuity of fault plane along space and enhance the separability of distant regions, a distance weight is constructed based on the three-dimensional Manhattan distance of the sub-block center point, and it is used to constrain the similarity of features between any two sub-blocks, so that adjacent sub-blocks are closer in feature space and distant sub-blocks are easier to distinguish, thereby establishing a spatial smoothing regularity that conforms to the continuity of seismic structure and the characteristics of fault jump. E. Pre-training loss construction and parameter update: The location prediction loss and the spatial smoothing regularization loss are summed to form the final volumetric block spatial relationship learning objective. This comprehensive loss is then used for iterative optimization on a large number of unlabeled 3D seismic volumes. The feature extraction network and the learnable distance parameter are updated simultaneously through gradient backpropagation, so that the model gradually obtains a stable volumetric spatial structure representation capability during continuous training.

Images