Semi-supervised 3D fault identification method based on sparse labeling

Abstract

The invention discloses a sparse annotation-based semi-supervised 3D fault identification method, which is characterized in that accurate identification and generalization of a three-dimensional fault under a complex geological background are realized by using two-dimensional slice annotation of a small amount of field data through synthetic data full-supervised learning, sparse annotation noise learning and unannotated data semi-supervised learning. Carrying out fully supervised learning on synthetic data through a three-dimensional HRNet model and Dice loss; generating small cubes from sparse labeling data through random cutting, inputting the small cubes into a student model and an isomorphic teacher model through strong and weak enhancement processing, executing slight noise refining processing on output of the teacher model, and calculating anti-noise sparse loss; and generating an enhanced small cube by the unlabeled data, respectively inputting the small cube into a student model and a double-teacher model, calculating stability through double-teacher output to obtain a pseudo label, and calculating consistency loss. The three tasks are executed in sequence, the student network is updated by adopting the cumulative gradient, the teacher network is updated through the EMA, and finally the fault position is deduced.

Description

Technical Field

[0001] The present invention belongs to the field of geophysics and artificial intelligence, and in particular relates to a semi-supervised 3D fault recognition method based on sparse annotation. Background Art

[0002] In seismic exploration and geological research, accurate identification of underground fault structures is of great significance for establishing geological models and improving the success rate of resource exploration. Traditional fault identification methods usually rely on the experience of geologists and manual interpretation, which is not only time-consuming and easily affected by human factors, but also easily interfered by non-fault factors when dealing with complex geological backgrounds and noisy environments, making it difficult to achieve efficient fault identification. In recent years, the introduction of deep learning methods has provided new possibilities for fault identification, providing more robust and accurate detection in variable or noisy environments. Rich annotations are crucial to the accuracy of deep learning methods, but the complex three-dimensional distribution of faults makes complete manual annotation almost impossible to achieve. The introduction of synthetic data is a major breakthrough in fault detection based on deep learning. However, there are significant differences between synthetic data and real field data, especially in terms of features and fault structure, which makes it difficult to generalize models trained only on synthetic data to real-world scenarios. Summary of the invention

[0003] In order to overcome the shortcomings of the prior art, the present invention provides a semi-supervised 3D fault recognition method based on sparse annotation. The method can be based on a small amount of sparsely annotated two-dimensional slices of field data, through fully supervised learning of synthetic data, sparsely annotated noise learning and semi-supervised learning of unlabeled data, to achieve accurate recognition and efficient generalization of three-dimensional seismic faults under complex geological background.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] Step 1: Obtain synthetic seismic data with fault annotations and unlabeled field seismic data as training data, and annotate a small amount of two-dimensional slices of some field data, and divide the training data into three categories: fully annotated data, sparsely annotated data, and unlabeled data;

[0006] Step 2: Perform strong enhancement on the fully annotated data, use HRNet as the basic architecture of the student model, expand HRNet in three dimensions to adapt to 3D seismic data, and input the enhanced fully annotated data into the student model to calculate the Dice loss;

[0007] Step 3: Randomly crop the sparse labeled data to form several small cubes, perform strong and weak enhancement processing on the small cubes, and input them into the student model and the teacher model with the same structure as the student model respectively. The output of the student model remains unchanged, and the output of the teacher model is slightly noise-refined. The loss of the two outputs is calculated according to the noise-resistant sparse loss function;

[0008] Step 4: Randomly crop the unlabeled data to form several small cubes. The small cubes are subjected to strong and weak enhancement respectively and input into the student model and the dual-teacher model respectively. The output of the student model remains unchanged. Stability calculation is performed on the two outputs of the dual-teacher model to obtain pseudo labels, and then the consistency loss is calculated.

[0009] Step 5: Repeat steps 2-4 and use their accumulated gradients to update the student network, and then update the teacher network through EMA until the specified number of training times is reached;

[0010] Step 6: Use the trained model to judge the unknown earthquake data and obtain the fault identification result.

[0011] Furthermore, the slight noise refinement process takes the predicted value of the student model, the predicted value of the teacher model and the average of the original sparse labels as new labels.

[0012] Furthermore, the noise-resistant sparse loss function effectively suppresses the overfitting of the model by noise data by combining the MAE loss function and the Dice loss function, and introduces a specially designed mask matrix M to solve the problem of sparse labels. The positions of unlabeled voxels in the matrix are marked as -1, and the positions of sparsely labeled voxels are marked as 1. When calculating the loss, only the loss value of the voxel corresponding to the position marked as 1 in the mask matrix is ​​considered. The loss function expression is:

[0013]

[0014] Among them, y i ′,y i are the values ​​of the predicted label and the true label at each voxel i, n represents the total number of voxels in the sample, and ∈ is a small constant to avoid the denominator being zero.

[0015] Furthermore, the stability calculation is based on two teacher networks with different parameters and the same structure to calculate the stability of two output results, and the intersection over union (IOU) is used as a criterion for measuring stability, and the formula is:

[0016]

[0017] Among them, w represents weak enhancement processing, They represent the probability outputs of the two teacher models t1 and t2 respectively. With the help of this indicator, we can judge the degree of overlap of the two prediction results. The larger the score, the higher the stability. We sort the stability scores of all unlabeled samples, take only the first 50% of the sample outputs as pseudo labels, and only calculate the consistency loss between the pseudo labels and the student model outputs.

[0018] The beneficial effects of the present invention are: it cleverly utilizes fully annotated synthetic data, sparsely annotated field data, and unannotated field data, breaking through the limitation of the existing deep learning fault identification method that only relies on synthetic data, and effectively overcoming the problems of difficult generalization of synthetic data, large workload of field annotation, and errors in manual annotation. Through this comprehensive method, not only the accuracy of fault identification is greatly improved, but also the generalization ability of the model to different work areas is greatly enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of training with sparsely annotated field data for step 3

[0020] Figure 2 Schematic diagram of training without labeled field data in step 4

[0021] Figure 3 Flowchart of the present invention DETAILED DESCRIPTION

[0022] Combine the following Figure 1 , 2 3. The present invention is further described in detail:

[0023] Step 1: Obtain an open-source synthetic seismic data set and a large amount of field data as training data, where each field data contains 1-3 complete fault-labeled 2D slices. In the present invention, the labeled slices are preliminarily predicted with the help of the fault recognition model and annotated with the assistance of expert experience, and there is slight noise data (wrong labeling).

[0024] Step 2: Since HRNet has multi-resolution parallel branches and information interaction mechanisms, it can effectively integrate semantic and position information and maintain high output resolution. In order to adapt to 3D tomographic data, we have expanded HRNet in three dimensions as the basic architecture of the student model. The fully annotated data after strong enhancement is used as input, the fault probability distribution is obtained through the student model, and Dice is used as the loss function to calculate the loss between the output and the synthetic data label. The Dice loss not only considers the predicted value of the current voxel when calculating, but also combines the predicted values ​​of other positions, thereby alleviating the impact of the imbalance in the number of fault and background voxels to a certain extent. The calculation method of Dice loss is shown in the following formula:

[0025]

[0026] Among them, y i ′,y i are the values ​​of the predicted label and the true label at each voxel i, n represents the total number of voxels in the sample, and ∈ is a small constant to avoid the denominator being zero.

[0027] Step 3: Randomly crop the sparsely labeled field data to form several small cubes. To ensure that the labeled slices are included in the cropped cubes, we use the field seismic data D fie,i ∈l t × x × i For example, l t ,l x ,l i They represent the size along the Timeline, Crossline, and Inline directions respectively. Assuming that the coordinate of a labeled slice in the Inline direction is represented as p, and the side length of the training sample is represented as l, then the jth training sample can be represented as (X fie,j ,y fie,j ), representing a randomly cropped earthquake cube of size l×l×l from the field data and its corresponding sparse annotations. (X fie,j ,y fie,j ) The coordinates p of the clipping start position t ,p x ,p i Make your selection according to the following formula:

[0028] p t ~U(0,l t -l)

[0029] p x ~U(0,l x -l)

[0030] p i ~U(max(0,p-l+1),min(p,l i -l))

[0031] X fie,j The strong and weak enhancement processes are performed respectively and input into the student model and the teacher model with the same structure as the student model. The output of the student model remains unchanged, and the output of the teacher model is slightly noise-refined. The average of the predicted value of the student model, the predicted value of the teacher model and the original sparse label is used as the new label. Then, the loss of the two outputs is calculated according to the noise-resistant sparse loss function, as follows: Figure 1As shown. This loss function effectively suppresses the overfitting of the model by noisy data by combining the MAE loss function and the Dice loss function, and introduces a specially designed mask matrix M to solve the problem of sparse labels. The positions of unlabeled voxels in this matrix are marked as -1, and the positions of sparsely labeled voxels are marked as 1. When calculating the loss, only the loss value of the voxel corresponding to the position marked as 1 in the mask matrix is ​​considered. The loss function expression is:

[0032]

[0033] Among them, y i ′,y i are the values ​​of the predicted label and the true label at each voxel i, n represents the total number of voxels in the sample, and ∈ is a small constant to avoid the denominator being zero.

[0034] Step 4: Randomly crop the unlabeled field data to form several small cubes. The small cubes are processed with strong and weak enhancement respectively, and are input into the student model and the dual-teacher model with different parameters but the same structure. The output of the student model remains unchanged, and the stability calculation is performed on the two outputs of the dual-teacher model to obtain pseudo labels. The stability calculation uses the intersection over union (IOU) as a stability criterion to determine the stability of the output result. The formula is:

[0035]

[0036] Among them, w represents weak enhancement processing, Represent the probability outputs of the two teacher models t1 and t2 respectively. With the help of this indicator, the degree of overlap of the two prediction results can be judged. The larger the score, the higher the stability. We sort the stability scores of all unlabeled samples, take only the top 50% of the sample outputs as pseudo labels, and then calculate the consistency loss between the pseudo labels and the student model outputs, such as Figure 2 shown.

[0037] Step 5: Repeat steps 2-4 and use their accumulated gradients to update the student network, and then update the teacher network through EMA until the specified number of training times is reached. The whole process is as follows: Figure 3 shown.

[0038] Step 6: Use the trained model to judge the unknown earthquake data and obtain the fault identification result.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention. Any technician familiar with the profession may use the above-described technical solutions to modify or change them into equivalent examples with equivalent changes. Any simple modification, change or modification of the above embodiments based on the technical solutions of the invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the invention.

Claims

1. A semi-supervised 3D fault recognition method based on sparse annotation, characterized in that: The following steps are involved: Step 1: Obtain synthetic seismic data with fault annotations and unlabeled field seismic data as training data, and annotate a small amount of two-dimensional slices of some field data, and divide the training data into three categories: fully annotated data, sparsely annotated data, and unlabeled data; Step 2: Perform strong enhancement on the fully annotated data, use HRNet as the basic architecture of the student model, expand HRNet in three dimensions to adapt to 3D seismic data, and input the enhanced fully annotated data into the student model to calculate the Dice loss; Step 3: Randomly crop the sparse labeled data to form several small cubes, perform strong and weak enhancement processing on the small cubes, and input them into the student model and the teacher model with the same structure as the student model respectively. The output of the student model remains unchanged, and the output of the teacher model is slightly noise-refined. The loss of the two outputs is calculated according to the noise-resistant sparse loss function; Step 4: Randomly crop the unlabeled data to form several small cubes. The small cubes are subjected to strong and weak enhancement respectively and input into the student model and the dual-teacher model respectively. The output of the student model remains unchanged. Stability calculation is performed on the two outputs of the dual-teacher model to obtain pseudo labels, and then the consistency loss is calculated. Step 5: Repeat steps 2-4 and use their accumulated gradients to update the student network, and then update the teacher network through EMA until the specified number of training times is reached; Step 6: Use the trained model to judge the unknown earthquake data and obtain the fault identification result.

2. The semi-supervised 3D fault recognition method based on sparse annotation according to claim 1, characterized in that: The slight noise refinement process takes the average of the prediction value of the student model, the prediction value of the teacher model and the original sparse label as the new label.

3. The semi-supervised 3D fault recognition method based on sparse annotation according to claim 1, characterized in that: The noise-resistant sparse loss function effectively suppresses the overfitting of the model by noisy data by combining the MAE loss function and the Dice loss function, and introduces a specially designed mask matrix M to solve the problem of sparse labels. The positions of unlabeled voxels in this matrix are marked as -1, and the positions of sparsely labeled voxels are marked as 1. When calculating the loss, only the loss values ​​of the voxels corresponding to the positions marked as 1 in the mask matrix are considered.

4. The semi-supervised 3D fault recognition method based on sparse annotation according to claim 1, characterized in that: The stability calculation is based on the stability of the output results of two teacher networks with different parameters and the same structure, and the intersection over union (IOU) is used as the standard to measure the stability. With the help of this indicator, the degree of overlap of two prediction results can be judged. The larger the score, the higher the stability. We sort the stability scores of all unlabeled samples, take only the first 50% of the sample outputs as pseudo labels, and then calculate the consistency loss between the pseudo labels and the student model output.

Images