High-resolution seismic fault detection with adversarial neural networks and regularization

1 . A method for high-resolution seismic fault detection with an adversarial neural network, wherein the method comprises following steps of:

training a target adversarial neural network based on a preset training sample set to obtain a trained target adversarial neural network, wherein the preset training sample set comprises seismic data and fault labels, the target adversarial neural network comprises: a segmentation module, a feature fusion module, and a discriminator module, the segmentation module is a module configured for obtaining a fault feature based on the preset training sample set, and the feature fusion module is a module configured for fusing the fault feature and the seismic data into a global feature map; and

performing seismic fault detection on a target seismic image based on the trained target adversarial neural network;

wherein the step of training a target adversarial neural network based on a preset training set comprises:

a first training step A: training the segmentation module by utilizing the preset training sample set based on a balanced cross entropy loss function, so as to obtain a trained segmentation module;

a predicting step B: substituting the preset training sample set into the trained segmentation module to obtain a predicted fault feature;

a fusing step C: fusing the seismic data and the predicted fault feature into a global feature map based on the feature fusion module;

a second training step D: training the discriminator module by utilizing the global feature map based on a categorical cross entropy loss function, so as to obtain a trained discriminator module;

a discriminating step E: substituting the global feature map into the trained discriminator module to obtain a discriminative difference value; and

an updating step F: updating the balanced cross entropy loss function based on the discriminative difference value and a regularization loss function, and

repeating the steps A through F from the first training step to the updating step, and finishing the training if the discriminative difference value is less than a preset threshold value; and

wherein the step of fusing the seismic data and the predicted fault feature into a global feature map based on the feature fusion module comprises:

performing local feature inversion on the predicted fault feature to obtain a degree of attention of the predicted fault feature; and

calculating the dot product of the degree of attention and the seismic data, and performing normalization processing of local contrast, so as to obtain the global feature map.

2 . The method according to claim 1 , wherein the predicted fault feature comprises a probability of predicted fault and a fault label; the step of performing local feature inversion on the predicted fault feature to obtain the degree of attention of the predicted fault feature comprises:

performing local feature inversion on the predicted fault feature by following equations:

P = 1− P

y = 1− y

where P is the probability of predicted fault, P is the degree of attention corresponding to the probability of predicted fault, y is the fault label, and y is the degree of attention corresponding to the fault label.

3 . The method according to claim 2 , wherein the step of updating the balanced cross entropy loss function by utilizing a regularization loss function comprises:

updating the balanced cross entropy loss function by following equation:

1_ S ( P,y,C )=1_pixel( P,y )+2·1_image( C )

where λ is a hyperparameter, 1 S (P,y,C) is the regularization loss function, 1 pixel (P,y) is the balanced cross entropy loss function, 1 image (C) is the discriminative difference value, and C is the output tensor of the discriminator module.

4 . A system for high-resolution seismic fault detection with an adversarial neural network, wherein the system comprises: a training device and a detection device; wherein

the training device is configured for training a target adversarial neural network based on a preset training sample set to obtain a trained target adversarial neural network, wherein the preset training sample set comprises seismic data and fault labels, the target adversarial neural network comprises: a segmentation module, a feature fusion module, and a discriminator module, the segmentation module is a module configured for obtaining a fault feature based on the preset training sample set, and the feature fusion module is a module configured for fusing the fault feature and the seismic data into a global feature map; and

the detection device is configured for performing seismic fault detection on a target seismic image based on the trained target adversarial neural network;

wherein the training device is further configured for:

a first training step A: training the segmentation module by utilizing the preset training sample set based on a balanced cross entropy loss function, so as to obtain a trained segmentation module;

a predicting step B: substituting the preset training sample set into the trained segmentation module to obtain a predicted fault feature;

a fusing step C: fusing the seismic data and the predicted fault feature into a global feature map based on the feature fusion module;

a second training step D: training the discriminator module by utilizing the global feature map based on a categorical cross entropy loss function, so as to obtain a trained discriminator module;

a discriminating step E: substituting the global feature map into the trained discriminator module to obtain a discriminative difference value; and

an updating step E: updating the balanced cross entropy loss function based on the discriminative difference value and a regularization loss function, and

repeating the steps A through F from the first training step to the updating step, and finishing the training if the discriminative difference value is less than a preset threshold value; and

wherein the feature fusion module is further configured for:

performing local feature inversion on the predicted fault feature to obtain a degree of attention of the predicted fault feature; and

calculating the dot product of the degree of attention and the seismic data, and performing normalization processing of local contrast, so as to obtain the global feature map.

5 . An electronic apparatus, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, wherein the steps of the method according to claim 1 are implemented when the computer program is executed by the processor.