Fault detection semi-supervised learning method for three-dimensional seismic data sparse labeling scene

Abstract

The invention provides a fault detection semi-supervised learning method for a three-dimensional seismic data sparse labeling scene. The fault detection semi-supervised learning method is used for realizing stable and continuous three-dimensional fault identification under the condition that only a small number of two-dimensional slices are labeled and a large number of voxels are not labeled. According to the method, firstly, a supervision mask is generated in a three-dimensional space based on two-dimensional marking, and supervision learning is executed only in an effective area to adapt to marking scarcity characteristics; and then, constructing a double-student model, keeping prediction coordination in an unlabeled area through mutual generation of pseudo labels, probability consistency and confidence constraint, and applying consistency constraint to overlapped areas of adjacent three-dimensional sub-blocks in combination with a teacher model so as to maintain spatial continuity of a fault structure. Through joint optimization of synthetic data pre-training and field data migration training, the model can still obtain reliable fault structure expression under the field conditions of strong noise, weak reflection and obvious cross-work-area difference.

Description

Technical Field

[0001] This invention belongs to the fields of geophysics and artificial intelligence applications, specifically relating to a semi-supervised learning method for fault detection in sparsely labeled 3D seismic data scenarios. Background Technology

[0002] Three-dimensional fault identification is a core task in seismic interpretation and subsurface structural analysis, with significant application value in oil and gas exploration, reservoir prediction, deep resource evaluation, and geological hazard monitoring. Actual 3D seismic volumes are typically characterized by their massive size, strong noise interference, weak reflection signals, and complex stratigraphic structures. Fault structures also exhibit significant cross-scale variations, diverse geometric shapes, and blurred boundaries, making automatic fault detection in real-world work areas highly challenging to model. Currently, industrial workflows still heavily rely on interpreters manually delineating fault locations frame-by-frame from 2D profiles. However, due to the enormous scale of seismic volumes and the heavy workload of interpretation, in practice, only a limited number of 2D fault annotations are often obtained from a very small number of profiles, leaving the vast majority of voxels unannotated, creating a typical sparse annotation scenario that fails to meet the requirements for constructing complete 3D labeled data.

[0003] In recent years, with the development of deep learning, 3D convolutional networks, multi-scale modeling methods, and the Transformer architecture have shown great potential in 3D seismic fault detection. However, these models generally rely on large-scale, complete, and high-quality 3D labeled data in practical applications, while the most readily available high-quality labels in the industry mainly come from synthetic seismic bodies. Although synthetic data has advantages such as complete labels and regular structures, its reflection characteristics, noise distribution, and structural complexity differ significantly from real-world seismic data. This leads to a decrease in accuracy, instability, or misidentification of structures when models trained on synthetic data are directly transferred to the field. To alleviate the differences between synthetic and field data and improve the model's adaptability to real structures, practical applications usually require the introduction of a small number of reference-value 2D slice labels from the field seismic body to constrain the model. However, 2D labels are sparse and lack spatial continuity, making it difficult to support high-quality 3D fault detection with only a limited number of labels. Therefore, how to fully utilize a large amount of unlabeled 3D seismic data and a small number of 2D slice labels for collaborative modeling under conditions of extremely scarce field labels, and construct a stable semi-supervised fault detection framework with cross-field generalization capabilities, has become a key problem that urgently needs to be solved in the field of intelligent earthquake interpretation. Summary of the Invention

[0004] Given that existing 3D fault detection methods generally rely on a large amount of complete labeled data in field work areas, while real 3D seismic bodies typically only have a few 2D slice annotations, which are extremely sparse and discontinuous, traditional supervised models are prone to instability, poor cross-domain adaptability, and structural misjudgment in complex field environments. This invention aims to provide a semi-supervised learning method for fault detection in scenarios with sparse 3D seismic data annotation. This method, through the construction of a sparse supervised mask, a dual-student complementary learning mechanism, teacher-student consistency constraints, and cross-block structural continuity modeling, enables the model to learn stable, continuous, and geologically meaningful spatial structural representations under the combined influence of a small number of labeled and a large number of unlabeled 3D seismic bodies. This significantly improves the stability, accuracy, and cross-field generalization ability of fault detection under field conditions such as strong noise interference, weak reflection, and complex multi-scale structures, making it suitable for practical needs in seismic interpretation where labeled data is insufficient.

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

[0006] A. Construction of sparse labeled data and generation of supervised masks:

[0007] A fully annotated synthetic 3D data volume and a field data volume containing only a few 2D slice annotations are used together for training. A supervision mask is generated based on the position of the field annotations in 3D space, and the corresponding voxels are set as supervised regions, while the remaining voxels are treated as unsupervised regions. During training, supervision error is calculated only for supervised regions, thereby avoiding the unlabeled voxels from misleading the model and making the supervision mechanism adaptable to the sparse field annotations.

[0008] B. Structural Construction of the Dual-Student Model and Teacher Model:

[0009] A dual-branch structure is constructed, consisting of two types of student models with different feature focuses. One type is responsible for extracting local details and fault boundary features, while the other is used to capture overall tectonic and large-scale structural information. The same 3D seismic data is simultaneously input into both models to obtain complementary features. The teacher model is obtained by smoothly updating the parameters of one of the student networks, which is used to provide a more stable structural reference in subsequent training.

[0010] C. Cooperative consistency constraints in the dual-student model:

[0011] For unlabeled regions, pseudo-labels are generated from the predictions of one student model to constrain another student model, gradually aligning their class judgments at the voxel level. Simultaneously, soft pseudo-labels are created by flexibly adjusting the predicted probabilities to guide the other student model towards consistency in the probability space. Furthermore, a confidence relationship is constructed based on the class probability outputs of the two student models, ensuring coordination in their class distributions, thereby enhancing the discriminative reliability of unlabeled regions and the overall structural stability.

[0012] D. Cross-block information consistency constraints in the teacher-student model:

[0013] The three-dimensional seismic body is divided into several adjacent sub-blocks, which are then input into the teacher model and the student model respectively. By applying consistency constraints to the feature representation of the overlapping areas of the sub-blocks, the model maintains structural continuity in space, reducing the occurrence of discontinuous fault boundaries, local voids, or misjudged structural areas.

[0014] E. Pre-training and on-site transfer training:

[0015] First, basic training is performed using fully labeled synthetic data volumes to enable the model to recognize initial structures. Then, the model is transferred to sparsely labeled data volumes in the field, and supervised updates are performed only in limited labeled areas based on supervised masks. A unified semi-supervised objective is constructed by combining student consistency and teacher-student cross-block consistency to achieve end-to-end optimized training. Attached Figure Description

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

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

[0018] A. Construction of sparse labeled data and generation of supervised masks:

[0019] This implementation method unifies the organization of synthetic 3D data and on-site sparse labeled data, and constructs a 3D supervision mask based on 2D slice annotations, enabling the model to obtain stable supervision information under conditions of limited annotations and a large number of unlabeled voxels. The specific steps are as follows:

[0020] (1) Construction of hybrid dataset: Let the synthetic 3D seismic body be ,in , and These represent the height, width, and depth of the data, respectively, and are fundamental parameters for volumetric partitioning and spatial relationship calculations. The corresponding complete fault labeling is as follows: ,in This indicates that the voxel belongs to a fault. This indicates that the voxel does not belong to a fault, among which These represent the position index or coordinates in the height, width, and depth directions of the volume data, respectively.

[0021] The on-site three-dimensional data volume is recorded as Only in several two-dimensional sections These cross-sections, which are manually labeled, form a set:

[0022]

[0023] To balance complete annotation information with actual seismic reflection characteristics, this implementation method uses both types of data to form a hybrid training set:

[0024]

[0025] (2) Constructing a 3D supervised mask based on slice annotations: Slice annotations from the field data are mapped to their corresponding spatial locations in the 3D volume. For any slice... Its set of labeled voxels in three-dimensional space can be represented as:

[0026]

[0027] in This indicates the position index or coordinates of the slice in the height, width, and depth directions. The effective 3D supervised region formed by all slice annotations is... Based on this, a three-dimensional supervised mask is constructed. Among them, belonging to voxel correspondence The remaining voxels correspond to .

[0028] For synthetic data Because the annotations are complete, the supervisory mask is always full 1s, while the field data has annotations only on a few sections, so the vast majority of voxels in the mask are 0s.

[0029] (3) Sparse supervision loss: Suppose the model's prediction probability for a 3D voxel is... Then the monitoring loss is calculated only in the effective area of ​​the mask, and its form is:

[0030]

[0031] in Represents the model's position The predicted probability of voxels, ranging from This refers to the prediction confidence that the voxel belongs to the fault region. The true label of the voxel is usually binary, with 1 indicating a tomographic voxel and 0 indicating a non-tomographic voxel. This represents the supervisory mask, indicating which voxels are valid labeled regions. If If so, the voxel is a valid labeled region, and the model will calculate the loss; if If the voxel is unlabeled, the model will not calculate the loss. Describe a very small constant (e.g.) This is used to avoid division by zero errors, ensure that the denominator is always positive, and avoid instability in calculations. The value represents the voxel-level error function, and the Dice loss metric is used to measure the consistency between the predicted volume and the actual fault volume.

[0032] B. Structural Construction of the Dual-Student Model and Teacher Model:

[0033] By constructing a structural system consisting of two types of student models and a teacher model, the local details and overall structural features of a 3D seismic body can be simultaneously acquired. Specifically, the following steps are included:

[0034] (1) Construction and feature extraction of the dual-student model: Let the input three-dimensional seismic body be... ,in , and These represent the height, width, and depth of the input data, respectively. To simultaneously depict both the local details and the overall structural trend of the fault, this implementation constructs two student models with complementary structural characteristics. Each model includes a feature extraction module and a prediction module, responsible for extracting intermediate feature representations and outputting voxel-level class probability predictions, respectively. First, structural feature representations at different scales are obtained through the feature extraction modules of the two models:

[0035]

[0036] Among them, the model The UNet network, based on a 3D convolutional structure, is more suitable for extracting high-resolution structural features such as fault pinch-outs, local curvatures, and rupture surface details. More emphasis is placed on the expression of local structures; model Employing a 3D Swing Transformer network based on a self-attention mechanism, it can model long-range spatial dependencies and large-scale tectonic trends, and its output features... This better reflects the overall structural morphology and fault extension pattern. Based on this, the prediction modules of the two student models further generate voxel-level category probability distributions:

[0037]

[0038] in This represents the number of categories (e.g., faults and non-faults). and These represent the prediction modules of student model A and student model B, respectively, used to represent the corresponding features. , Mapped to a voxel-level class probability distribution, thus obtaining and The predicted output provides the probabilistic information needed for subsequent sparse supervised learning and consistency constraints.

[0039] (2) Construction and parameter updating of the teacher model: based on the student model Parameters for constructing a teacher model And updated using an exponential moving average method:

[0040]

[0041] in and They represent teachers respectively. With students Model parameters, The smoothing factor is used. The teacher model provides a more stable structural reference with stationary parameters, which can avoid feature drift of the student model in sparse labeling and long-term training environments, and provide a reliable benchmark for subsequent cross-block consistency constraints.

[0042] C. Cooperative consistency constraints in the dual-student model:

[0043] Voxel-level probability predictions for two types of student models were obtained. Subsequently, to ensure that the two student models with different structural characteristics maintain consistent structural representations in unlabeled regions, this implementation method constructs collaborative constraints based on prediction results in three aspects: category consistency, probability consistency, and confidence consistency. Specifically, this includes the following steps:

[0044] (1) Category consistency constraint: for the student model Output voxel-level class probabilities Perform along the category dimension Discretization (taking the category index corresponding to the highest probability) generates hard pseudo-labels, where... This represents a category index (such as in binary tomography). Indicates non-fault voxels. (Representing tomographic voxels)

[0045]

[0046] And use it as a supervision signal to constrain the student model. Output The reverse can also be achieved:

[0047]

[0048] Used to guide student models This constraint ensures consistency between the two models in class classification, avoiding significant class conflicts in unlabeled regions and improving the stability of class prediction. Ultimately, this is achieved through the class consistency constraint. The loss function is constrained by minimizing the mean squared error between the hard pseudo-labels of the two models. The loss function is defined as follows:

[0049]

[0050] in It is an unlabeled set of voxels. For the number of unlabeled voxels, For voxel index (position index).

[0051] (2) Probabilistic consistency constraint: To flexibly guide the learning process of the model in the uncertain region, the student model is subject to the following constraints: The probability prediction performs a temperature scaling Sharpen operation to generate soft pseudo-labels:

[0052]

[0053] in The temperature parameter is used to enhance high-confidence regions and suppress noisy regions. This soft pseudo-label is further used to guide the student model. Probability prediction can be reversed using:

[0054]

[0055] Guided student model Therefore, the probabilistic consistency constraint loss The calculation is as follows:

[0056]

[0057] in Indicates voxel index, This indicates an unlabeled set of voxels. and The two student models are located at the following positions. Probability prediction, , The corresponding temperature-sharpened soft pseudo-labels are used. This constraint ensures that the two models maintain consistency in probability distribution, enabling them to obtain smooth and stable prediction results in areas of low reflection and high noise.

[0058] (3) Confidence Consistency Constraint: To avoid inconsistent confidence distributions between the two student models in complex structural regions, this implementation method imposes a confidence consistency constraint on the probability predictions of the two student models. and Global normalization is performed to obtain the global confidence vector at the category level:

[0059]

[0060] Among them, Normalize( This represents normalizing the class probabilities of the model output across all voxels, so that each class obtains a global confidence level that reflects the overall confidence level. Then, by minimizing... and The difference between them ensures consistency in the overall representation of class confidence between the two models. This difference can be achieved by minimizing the mean squared error loss function, specifically in the form:

[0061]

[0062] in, It is the number of categories. and Represent student model A and student model B respectively for the first... The confidence level of the class. Specifically, This corresponds to a non-fault voxel, i.e., the probability or confidence level that the model predicts a voxel as a non-fault region; while The corresponding tomographic voxel represents the probability or confidence that the model predicts a voxel belongs to a tomographic region. By minimizing the mean squared error between the class confidence scores of the two models, we ensure that they maintain consistency in their class confidence score distributions.

[0063] D. Cross-block information consistency constraints in the teacher-student model

[0064] To maintain the cross-block continuity of fault structures in unsupervised regions, this implementation utilizes a teacher model to provide stable structural priors, aligning adjacent 3D sub-blocks in shared regions based on feature and probability consistency. The specific steps are as follows:

[0065] (1) Division of adjacent sub-blocks and determination of overlapping areas: from the input three-dimensional seismic volume Two sub-blocks of the same size are randomly selected from the sample, and denoted as . ,in and Adjacent in three-dimensional space, and sharing a portion of the overlapping region composed of voxels. The proportion of spatial voxels in overlapping regions typically accounts for 25%-50% of the total number of voxels in the sub-blocks.

[0066] (2) Feature and probability output of teacher and student models: Two overlapping sub-blocks are input into the feature extraction modules of the teacher and student models respectively to obtain local feature representations. The outputs of the teacher and student models are:

[0067] ,

[0068] in and These represent the feature extraction modules for the teacher model and the student model, respectively. and belong , where represents the intermediate feature representations extracted by the two models on corresponding sub-blocks, For feature dimensions.

[0069] (3) Feature consistency constraint of overlapping regions: in the overlapping regions of two sub-blocks Above, construct the feature consistency loss:

[0070]

[0071] This constraint forces the student model to maintain consistency with the teacher model in the feature representation of spatially adjacent regions, so that structural information such as fault boundaries, bedding extension, and dip changes can be continuously transmitted between sub-blocks.

[0072] E. Pre-training and on-site transfer training

[0073] The training process in this implementation consists of two stages: pre-training with synthetic data and transfer training with field data. The aim is to gradually obtain a 3D fault detection model with cross-work area generalization capabilities, maintaining a stable and continuous fault structure representation even under conditions of sparse annotation and a large number of unannotated voxels. The specific steps are as follows:

[0074] (1) Model pre-training on synthetic data: First, on the fully labeled synthetic 3D seismic body The two student models are independently supervised and pre-trained, enabling the UNet based on the 3D convolutional structure and the dual student model based on the 3D SwinTransformer to initially acquire the ability to learn local details of faults and large-scale structural trends, and to form complementary local-global structural representation capabilities in the feature space.

[0075] (2) Field Data Transfer Training: Based on pre-training, the model is transferred to field data with at least a minimum number of labeled slices. In this stage, a teacher model updated based on exponential moving average is introduced. Simultaneously apply all supervision and consistency constraints, including sparse supervision loss. Category consistency constraints Probabilistic consistency constraints Confidence consistency constraint And the cross-block feature consistency constraint of the teacher-student model in the overlapping region. Based on the above monitoring and consistency signals, the total loss during the migration phase is constructed as follows:

[0076]

[0077] The weights for each loss are set empirically. Through this joint optimization process, the model can simultaneously take into account explicit supervision and implicit structural priors, under the condition that sparse annotations and a large number of unlabeled voxels are involved, thereby obtaining the ability to represent three-dimensional faults with stability and spatial continuity in complex construction scenarios.

Claims

1. A semi-supervised learning method for fault detection in scenarios with sparse annotation of 3D seismic data, characterized in that, Includes the following steps: A. Construction of sparse labeled data and generation of supervised masks: The training set is constructed by combining a synthetic 3D data volume with complete annotations with a field data volume with only a few slice annotations. A supervision mask is generated in 3D space based on the limited annotations of the field data. The annotated area is set as the supervised area and the unannotated area is set as the unsupervised area. During training, the supervision error signal is calculated only for the supervised area and no gradient contribution is generated for the unannotated area. This forms a supervision mechanism suitable for sparse annotation conditions and avoids the bias caused by unknown voxels being misclassified as background. B. Structural Construction of the Dual-Student Model and Teacher Model: A dual-branch system is constructed, consisting of two types of student models with different structural characteristics. One type of model focuses on extracting local structural information, while the other type focuses on modeling the overall spatial structure and long-range dependencies. The same 3D data volume is input into the two types of student models to obtain complementary feature representations. A teacher model is constructed by smoothly updating the parameters of one of the student models, so that it can provide a stable structural reference in subsequent stages. C. Cooperative consistency constraints in the dual-student model: The prediction results of one student model are discretized and used as pseudo-labels to constrain another student model, so that the two are consistent in class judgment. At the same time, the probability prediction is nonlinearly transformed to generate soft pseudo-labels and used to guide the other student model, so that it maintains flexible consistency in the probability space. The confidence relationship is constructed based on the probability output of the two student models and consistency constraints are applied to keep them coordinated in class confidence, thereby improving prediction stability and overall structural consistency in unlabeled areas. D. Cross-block information consistency constraints in the teacher-student model: Spatially adjacent 3D sub-blocks are input into the teacher model and student model respectively, and consistency constraints are applied to the feature representation in their overlapping areas to maintain the construction continuity of the unsupervised region and avoid breakage, empty response or prediction degradation. E. Pre-training and on-site transfer training: First, the model is pre-trained using the complete annotations of the synthetic data volume to obtain basic structural representation capabilities. Then, the model is transferred to the field sparsely labeled data volume and transfer training is performed based on the supervised mask to adapt the model to the structural characteristics of the field data. Finally, the sparse supervision mechanism, the two-student consistency constraint, and the teacher-student cross-block consistency are all incorporated into a unified semi-supervised optimization framework.

Images