A seismic fault detection method based on contrastive semi-supervised learning

Abstract

The application discloses a contrastive semi-supervised learning seismic fault detection method. The method comprises the following steps: obtaining synthetic seismic data and field seismic data, wherein the synthetic seismic data is labeled, and the field seismic data does not contain labels; designing a convolutional neural network with an autoencoder structure; using a gradient descent-based optimization method to train the neural network with labeled synthetic data to obtain a pre-trained model and complete the pre-training process; copying the trained neural network as a teacher network and setting its learning rate to 0, while setting the original neural network as a student network; entering the training step, randomly cutting an n*n*n seismic block X0 from the field seismic data, and performing strong enhancement on the seismic block X0 to obtain an enhanced seismic block X1; performing weak enhancement on the same seismic block to obtain an enhanced seismic block X2; inputting the seismic block X1 into the student network and inputting the seismic block X2 into the teacher network, projecting the outputs of the two through a multilayer perceptron, then calculating the cosine similarity loss, and updating the student network and the teacher network with the loss; and repeating the training step until a specified number of executions is reached.

Description

Technical Field

[0001] This invention belongs to the fields of geophysics and artificial intelligence, specifically relating to a seismic fault detection method based on comparative semi-supervised learning. Background Technology

[0002] In the field of seismic exploration, fault detection is a crucial task, providing vital support for oil and gas exploration and seismic risk assessment. Traditional fault detection methods require manual intervention and specialized knowledge, resulting in low efficiency and susceptibility to subjective factors. In recent years, with the development of deep learning technology, fault detection methods based on convolutional neural networks have received widespread attention. However, due to the lack of labels on field seismic data, they are difficult to directly utilize for supervised learning. Therefore, semi-supervised learning has become an important approach to address this issue. This invention provides a seismic fault detection method based on comparative semi-supervised learning, which overcomes the limitations of traditional methods and improves detection efficiency and accuracy. Summary of the Invention

[0003] This invention relates to a comparative semi-supervised learning method for seismic fault detection, the main contents of which include the following steps:

[0004] 1. Acquire synthetic seismic data and field seismic data. Synthetic seismic data is labeled, while field seismic data is unlabeled.

[0005] 2. Design a convolutional neural network with an autoencoder structure.

[0006] 3. Use labeled synthetic data to train the neural network using a gradient descent-based optimization method to obtain a pre-trained model, thus completing the pre-training process.

[0007] 4. Duplicate the trained neural network to serve as the teacher network, setting its learning rate to 0. The original neural network becomes the student network. The student network updates its parameters using gradient descent, while the teacher network updates its parameters according to the following equation:

[0008]

[0009] in It is the teacher network at this moment. It's the teacher network from the previous moment, V t This is the student network at the current moment.

[0010] 5. Contrastive Semi-Supervised Learning Process. An n×n×n seismic block X0 is randomly selected from the field seismic data and subjected to strong enhancement, denoted as X1. Strong enhancement includes adding Gaussian noise at random scales, random gamma transform, and random flipping. A weak enhancement, denoted as X2, is applied to the same seismic block, consisting only of random flipping. X1 is input into the student network, and X2 into the teacher network. The outputs of both are projected through a multilayer perceptron, and then the cosine similarity loss is calculated to update both the student and teacher networks.

[0011] 6. Repeat step 5 until the prescribed number of training sessions is reached.

[0012] 7. Model Evaluation: After training, the model is evaluated. Qualitative results are obtained using unlabeled field seismic data, and quantitative results are obtained using labeled synthetic data. The optimal parameter combination is selected by combining the test results. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are one embodiment of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of the method of the present invention;

[0015] Figure 2 This is a structural diagram of the method of the present invention. Detailed Implementation

[0016] 1. Data Preparation: In implementing this invention, it is necessary to acquire synthetic seismic data and field seismic data. The synthetic seismic data already has labels, while the field seismic data does not. In this example, we use 200 synthetic data sets with a size of 128×128×128 and 10 field data sets with a size of 1024×1024×1024 for training.

[0017] 2. Network design: Construct a 3D convolutional neural network, which must conform to the structure of an autoencoder. In this embodiment, a 3D-UNet neural network with a basic width of 32 is selected.

[0018] 3. Pre-training: First, the neural network is trained using labeled synthetic data through a gradient descent-based optimization method to obtain a pre-trained model, completing the pre-training process. During training, cross-entropy loss and stochastic gradient descent optimization algorithms can be used. The pre-training process iterates 10,000 times with a batch size of 8.

[0019] 4. Teacher-Student Model Training: A copy of the trained neural network is made and used as the teacher network, with its learning rate set to 0. The original neural network becomes the student network. The student network updates its parameters using gradient descent, while the teacher network updates its parameters according to the following equation:

[0020]

[0021] in It is the teacher network at this moment. It's the teacher network from the previous moment, V t This is the student network at the current moment.

[0022] 5. Contrastive semi-supervised learning process:

[0023] a. Randomly extract a 128×128×128 seismic block X0 from the field seismic data and perform strong enhancement on it, denoted as X1. Strong enhancement includes adding Gaussian noise of random scale, random gamma transform, and random flipping.

[0024] b. Perform weak enhancement on the same seismic block, denoted as X2. Weak enhancement only includes random flipping.

[0025] c. Input X1 into the student network and X2 into the teacher network. Project the outputs of both networks through a multilayer perceptron, then calculate the cosine similarity loss and update the student and teacher networks.

[0026] 6. Repeat step 5 100,000 times, with a batch size of 8.

[0027] 7. Model Evaluation: After training, the model is evaluated. Qualitative results are obtained using unlabeled field seismic data, and quantitative results are obtained using labeled synthetic data. The optimal parameter combination is selected by combining the test results.

[0028] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seismic data fault segmentation method based on contrastive semi-supervised learning, characterized in that, Includes the following steps: Acquire synthetic and field seismic data, where the synthetic data is labeled and the field data is unlabeled; design a convolutional neural network with an autoencoder structure; use a gradient descent-based optimization method to train the neural network using the labeled synthetic data to obtain a pre-trained model and complete the pre-training process; copy the trained neural network as the teacher network and set its learning rate to 0, while using the original neural network as the student network. In the training step, an n×n×n seismic block X0 is randomly selected from the field seismic data and strongly enhanced to obtain the enhanced seismic block X1; the same seismic block is weakly enhanced to obtain the enhanced seismic block X2; seismic block X1 is input into the student network and seismic block X2 is input into the teacher network, and the outputs of the two are projected through a multilayer perceptron, and then the cosine similarity loss is calculated and used to update the student network and the teacher network; the training steps are repeated until the specified number of executions is reached.

Images